| Original | Translation |
|---|---|
| Arm Holdings plc (ARM) Shareholder/Analyst Call March 24, 2026 1:00 PM EDT Company Participants Rene Haas - CEO & Director Mohamed Awad - Executive Vice President of Cloud AI Business Unit Conference Call Participants Santosh Janardhan - Meta Platforms, Inc. Kevin Weil - OpenAI, L.L.C. Paul Saab Presentation Unknown Attendee [Presentation] Unknown Attendee Please welcome Arm's Chief Executive Officer, Rene Haas. Rene Haas CEO & Director Such a nice warm welcome. Thank you. So welcome to our live stream audience, watching the Arm Everywhere event. | **Arm ํ๋ฉ์ค plc (ARM) ์ฃผ์ฃผ/์ ๋๋ฆฌ์คํธ ์ปจํผ๋ฐ์ค ์ฝ** **2026๋ 3์ 24์ผ ์คํ 1์ (๋ฏธ๊ตญ ๋๋ถ ํ์ค์)** **ํ์ฌ ์ฐธ์์** ๋ฅด๋ค ํ์ค - CEO ๊ฒธ ์ด์ฌ ๋ชจํ๋ฉ๋ ์์๋ - ํด๋ผ์ฐ๋ AI ์ฌ์ ๋ถ ์ด๊ด ๋ถ์ฌ์ฅ **์ปจํผ๋ฐ์ค ์ฝ ์ฐธ์์** ์ฐํ ์ ์๋๋ฅด๋จ - ๋ฉํ ํ๋ซํผ์ค Inc. ์ผ๋น ์จ์ผ - ์คํAI, L.L.C. ํด ์ฌ๋ธ **๋ฐํ** **๋ฐํ์ ๋ถ๋ช ** [๋ฐํ] **๋ฐํ์ ๋ถ๋ช ** ์์ ์ต๊ณ ๊ฒฝ์์(CEO)์ด์ ๋ฅด๋ค ํ์ค๋์ ๋ชจ์๊ฒ ์ต๋๋ค. **๋ฅด๋ค ํ์ค** **CEO ๊ฒธ ์ด์ฌ** ์ด๋ ๊ฒ ๋ฐ๋ปํ๊ฒ ํ์ํด ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. ๊ทธ๋ผ, Arm Everywhere ํ์ฌ๋ฅผ ์์ฒญํ๊ณ ๊ณ์ ๋ผ์ด๋ธ ์คํธ๋ฆผ ์์ฒญ์ ์ฌ๋ฌ๋ถ๊ป๋ ํ์์ ๋ง์์ ์ ํฉ๋๋ค. |
| I don't think we've ever done a live stream event like this and to the folks here in the audience, thank you so much for coming to the historic Fort Mason. And you may not know that Fort Mason here in California, was actually an official defense site for the Civil War. And this is where a very famous battle between Alabama, Georgia and California took place. Now you're thinking to yourself, I don't remember that battle. That's why this area looks so pristine. There actually was not a battle. But it actually was a fort. So I thought that was actually kind of neat. I didn't actually know that. So thank you again for attending a big day for us. We have a lot to share with you. | ์ ํฌ๊ฐ ์ด๋ฐ ๋ผ์ด๋ธ ์คํธ๋ฆฌ๋ฐ ํ์ฌ๋ฅผ ํด๋ณธ ์ ์ ์๋ ๊ฒ ๊ฐ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ฌ๊ธฐ ๊ณ์ ์ฒญ์ค ์ฌ๋ฌ๋ถ๊ป, ์ ์ ๊น์ ํฌํธ ๋ฉ์ด์จ์ ์์ฃผ์
์ ์ง์ฌ์ผ๋ก ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์๋ง ๋ชจ๋ฅด์ค ์๋ ์๊ฒ ์ง๋ง, ์บ๋ฆฌํฌ๋์์ ์๋ ์ด ํฌํธ ๋ฉ์ด์จ์ ์ฌ์ค ๋จ๋ถ ์ ์ ๋น์ ๊ณต์ ๋ฐฉ์ด ๊ธฐ์ง์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ฐ๋ก ์ด๊ณณ์์ ์จ๋ผ๋ฐฐ๋ง, ์กฐ์ง์, ์บ๋ฆฌํฌ๋์ ๊ฐ์ ์์ฃผ ์ ๋ช ํ ์ ํฌ๊ฐ ๋ฒ์ด์ก์ฃ . ์๋ง ์ง๊ธ์ฏค ์ฌ๋ฌ๋ถ์ '๊ทธ ์ ํฌ๋ ๊ธฐ์ต๋์ง ์๋๋ฐ?'๋ผ๊ณ ์๊ฐํ์ค ๊ฒ๋๋ค. ๋ฐ๋ก ๊ทธ ๋๋ฌธ์ ์ด ์ง์ญ์ด ์ด๋ ๊ฒ ๊นจ๋ํ๊ฒ ๋ณด์กด๋์ด ์๋ ๊ฒ๋๋ค. ์ฌ์ค์ ์ ํฌ๊ฐ ์์์ต๋๋ค. ํ์ง๋ง ์ด๊ณณ์ ์ค์ ๋ก ์์์์ต๋๋ค. ๊ทธ๋์ ์ ๋ ๊ทธ ์ฌ์ค์ด ๊ฝค ํฅ๋ฏธ๋กญ๋ค๊ณ ์๊ฐํ์ต๋๋ค. ์ ๋ ์ฌ์ค ๊ทธ ์ฌ์ค์ ๋ชฐ๋์ต๋๋ค. ๋ค์ ํ๋ฒ, ์ ํฌ์๊ฒ ์์ฃผ ์ค์ํ ๋ ์ ์ฐธ์ํด ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. ์ฌ๋ฌ๋ถ๊ณผ ๊ณต์ ํ ๋ด์ฉ์ด ๋ง์ต๋๋ค. |
| So I'm going to jump right into it. When we thought about how to name this event and how to talk about our company, we thought Arm Everywhere was really appropriate because one of the things that we're very proud of that we don't always think about in our daily lives at Arm, but it's really quite impactful. It's just the scale of the company and the scale of the magnitude we have. So when we start looking at numbers, 117 billion, what is the number? That's the total humans ever to live on earth. So that's -- if you count up by all of our calculations, how many people have lived on the planet since inception, about 117 billion. | ๋ฐ๋ก ๋ณธ๋ก ์ผ๋ก ๋ค์ด๊ฐ๊ฒ ์ต๋๋ค. ์ด๋ฒ ํ์ฌ์ ์ด๋ฆ์ ์ง๊ณ ์ ํฌ ํ์ฌ๋ฅผ ์ด๋ป๊ฒ ์๊ฐํ ์ง ๊ณ ๋ฏผํ์ ๋, 'Arm Everywhere'๊ฐ ์ ๋ง ์ ์ ํ๋ค๊ณ ์๊ฐํ์ต๋๋ค. ์ ํฌ๊ฐ ๋งค์ฐ ์๋์ค๋ฝ๊ฒ ์๊ฐํ๋ ๊ฒ ์ค ํ๋๊ฐ, Arm์์์ ์ผ์์ํ์์๋ ๋ ์ธ์งํ์ง ๋ชปํ์ง๋ง, ์ ๋ง ์๋นํ ์ํฅ๋ ฅ์ ๊ฐ์ง๊ณ ์๊ธฐ ๋๋ฌธ์
๋๋ค. ๋ฐ๋ก ์ ํฌ ํ์ฌ์ ๊ท๋ชจ์ ์ ํฌ๊ฐ ๊ฐ์ง ์ํฅ๋ ฅ์ ๊ท๋ชจ์
๋๋ค. ๊ทธ๋์ ์ซ์๋ฅผ ์ดํด๋ณด๋ฉด, 1,170์ต์ด๋ผ๋ ์ซ์๋ ๋ฌด์์ผ๊น์? ๊ทธ๊ฒ์ ์ง๊ตฌ์์ ์กด์ฌํ๋ ์ด ์ธ๋ฅ์ ์์ ๋๋ค. ๋ค์ ๋งํด, ์ ํฌ์ ๋ชจ๋ ๊ณ์ฐ์ ๋ฐ๋ฅด๋ฉด, ์ง๊ตฌ๊ฐ ์ฒ์ ์๊ธด ์ด๋๋ก ์ด ํ์ฑ์์ ์ด์๋ ์ฌ๋๋ค์ ์๊ฐ ์ฝ 1,170์ต ๋ช ์ ๋ ๋ฉ๋๋ค. |
| 350 billion plus are the number of Arm chips to have ever shipped. That is 3x the total number of humans who have ever existed on the planet. So it's not just 1 for every human. It's 3 for every human to have ever lived. 7x the total number of non Arm-based CPUs shipped combined. Just think about that number. And 160 Arm chips for every global household, Mine is probably larger than 160, but 160 is about the average. So that just gives you a sense of the scale of what we've done, and it's really important because it feeds into everything that makes us what we are today and, of course, could not be done without our ecosystem partners. | ์ง๊ธ๊น์ง ์ถํ๋ Arm ์นฉ์ ์ด๋์ 3,500์ต ๊ฐ ์ด์์ ๋ฌํฉ๋๋ค. ์ด๋ ์ธ๋ฅ ์ญ์ฌ์ ์ง๊ตฌ์์ ์กด์ฌํ๋ ์ด ์ธ๊ตฌ์์ 3๋ฐฐ์ ๋ฌํ๋ ์์น์
๋๋ค. ์ฆ, ๋จ์ํ ๋ชจ๋ ์ฌ๋๋น 1๊ฐ๊ฐ ์๋๋ผ, ์ธ๋ฅ ์ญ์ฌ์ ์กด์ฌํ๋ ๋ชจ๋ ์ฌ๋๋น 3๊ฐ์ ๋ฌํ๋ ์์ค์
๋๋ค. ๋น(้) Arm ๊ธฐ๋ฐ CPU์ ์ด ์ถํ๋์ ๋ชจ๋ ํฉ์น ๊ฒ์ 7๋ฐฐ์ ๋ฌํฉ๋๋ค. ์ด ์์น๊ฐ ์๋ฏธํ๋ ๋ฐ๋ฅผ ํ๋ฒ ์๊ฐํด ๋ณด์ญ์์ค. ๊ทธ๋ฆฌ๊ณ ์ ์ธ๊ณ ๋ชจ๋ ๊ฐ๊ตฌ๋น 160๊ฐ์ Arm ์นฉ์ด ๋ณด๊ธ๋์ด ์์ต๋๋ค. ์ ๊ฒฝ์ฐ์๋ ์๋ง 160๊ฐ๋ณด๋ค ๋ ๋ง์ ํ
์ง๋ง, 160๊ฐ๋ ํ๊ท ์น์
๋๋ค. ์ด๋ ์ ํฌ๊ฐ ๋ฌ์ฑํ ๊ท๋ชจ๋ฅผ ๋ณด์ฌ์ฃผ๋ ๊ฒ์ด๋ฉฐ, ์ค๋๋ ์ ์ ํฌ๋ฅผ ์๊ฒ ํ ๋ชจ๋ ๊ฒ์ ๊ธฐ๋ฐ์ด ๋๊ธฐ์ ๋งค์ฐ ์ค์ํฉ๋๋ค. ๋ฌผ๋ก , ์ด๋ ์ ํฌ์ ์์ฝ์์คํ ํํธ๋ ์์ด๋ ๊ฒฐ์ฝ ์ด๋ฃฐ ์ ์์๋ ์ผ์ ๋๋ค. |
| Now the company's DNA was really born to run off batteries. Company started in the early 1990s. It was a spinout of a British computer company named Acorn. And that company had a mandate to build a chip, and that chip had a couple of requirements. One was it had to run in a plastic package, which back then was really important. And number two, it had to be really low power. The first part was important because of heat. The second part was important because battery life met everything because this was going into the world's first PDA. So the company, we nailed that. | ์ ํฌ ํ์ฌ์ DNA๋ ๋ณธ์ง์ ์ผ๋ก ๋ฐฐํฐ๋ฆฌ ๊ตฌ๋์ ์ํด ์ค๊ณ๋์์ต๋๋ค. ํ์ฌ๋ 1990๋
๋ ์ด๋ฐ์ ์ค๋ฆฝ๋์๋๋ฐ, ์๊ตญ์ Acorn์ด๋ผ๋ ์ปดํจํฐ ํ์ฌ์์ ์คํ์์๋ ํ์ฌ์
๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ ํ์ฌ๋ ์นฉ์ ๊ฐ๋ฐํด์ผ ํ๋ ์๋ฌด๋ฅผ ๊ฐ์ง๊ณ ์์๊ณ , ๊ทธ ์นฉ์๋ ๋ช ๊ฐ์ง ์๊ตฌ์ฌํญ์ด ์์์ต๋๋ค. ์ฒซ์งธ, ํ๋ผ์คํฑ ํจํค์ง์์ ์๋ํด์ผ ํ๋๋ฐ, ์ด๋ ๋น์๋ก์๋ ์ ๋ง ์ค์ํ ๋ถ๋ถ์ด์์ต๋๋ค. ๋์งธ, ์ ๋ง ์ ์ ๋ ฅ์ด์ด์ผ ํ์ต๋๋ค. ์ฒซ ๋ฒ์งธ ์๊ตฌ์ฌํญ์ ์ด ๋ฌธ์ ๋๋ฌธ์ ์ค์ํ์ต๋๋ค. ๋ ๋ฒ์งธ ์๊ตฌ์ฌํญ์ ๋ฐฐํฐ๋ฆฌ ์๋ช ์ด ๋ชจ๋ ๊ฒ์ ์ข์ฐํ๊ธฐ ๋๋ฌธ์ธ๋ฐ, ์ด๋ ์ด ์นฉ์ด ์ธ๊ณ ์ต์ด์ PDA์ ํ์ฌ๋ ์์ ์ด์๊ธฐ ๋๋ฌธ์ ๋๋ค. ๊ทธ๋์ ์ ํฌ ํ์ฌ๋ ์ด๋ฅผ ์ฑ๊ณต์ ์ผ๋ก ํด๋์ต๋๋ค. |
| We nailed that objective so solidly that -- and this is a true story, that when the first Arm development board that had the first Arm1 processor was powered up, and these were plugged in now into a back of a wall. So you had a development board, lots of logic chips plugged into an AC outlet. When the AC outlet plug was removed, the chip kept running. And the chip kept running based upon the leakage current that was coming off all the other chips on the board. So the folks came in the next night and they saw the silico was still driving a signal. And that is really what for us launched the revolution of smartphones. | ์ฐ๋ฆฌ๋ ๊ทธ ๋ชฉํ๋ฅผ ๋๋ฌด๋ ํ์คํ๊ฒ ๋ฌ์ฑํ์ต๋๋ค. ์ด๊ฑด ์คํ์ธ๋ฐ์, ์ต์ด์ Arm1 ํ๋ก์ธ์๊ฐ ํ์ฌ๋ ์ฒซ Arm ๊ฐ๋ฐ ๋ณด๋์ ์ ์์ ์ผฐ์ ๋์์ต๋๋ค. ์ด ๋ณด๋๋ค์ ๋ฒฝ๋ฉด ์ฝ์ผํธ์ ์ฐ๊ฒฐ๋์ด ์์์ฃ . ์ฆ, ๊ฐ๋ฐ ๋ณด๋์ ์๋ง์ ๋ก์ง ์นฉ๋ค์ด AC ์ฝ์ผํธ์ ์ฐ๊ฒฐ๋์ด ์์๋ ๊ฒ๋๋ค. ๊ทธ๋ฐ๋ฐ AC ์ฝ์ผํธ ํ๋ฌ๊ทธ๋ฅผ ๋ฝ์๋๋ฐ๋, ์นฉ์ ๊ณ์ ์๋ํ์ต๋๋ค. ๊ทธ ์นฉ์ ๋ณด๋์ ์๋ ๋ค๋ฅธ ๋ชจ๋ ์นฉ๋ค์์ ๋์ค๋ ๋์ค ์ ๋ฅ๋ง์ผ๋ก ๊ณ์ ์๋ํ๋ ๊ฒ๋๋ค. ๊ทธ๋์ ๋ค์๋ ๋ฐค ์ฌ๋๋ค์ด ์์ ๋ณด๋, ๊ทธ ์ค๋ฆฌ์ฝ ์นฉ์ด ์ฌ์ ํ ์ ํธ๋ฅผ ๊ตฌ๋ํ๊ณ ์๋ ๊ฒ์ ๋ณด์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ๊ฒ์ด ๋ฐ๋ก ์ฐ๋ฆฌ์๊ฒ ์ค๋งํธํฐ ํ๋ช ์ ์์ํ๊ฒ ํ ๊ณ๊ธฐ๊ฐ ๋์์ต๋๋ค. |
| We were designed into the very first GSM phone for those who remember that Nokia brick on the far edge. But then the BlackBerry, which many of us who had loved, still love. Wish it came back, all the way to the modern smartphones of Android and iPhones. That is where we started in terms of the battery life. It launched a generation of smartphones. Now one of the breaks we got about 10 years ago was when SoftBank bought Arm. Yes, it was about 10 years ago, it was 2016. And when SoftBank bought Arm, Masa gave us an opportunity now that we're a private company to invest into areas that we were not able to invest in before. | ์ ํฌ ๊ธฐ์ ์ ์ต์ด์ GSM ํฐ์ ํ์ฌ๋์์ต๋๋ค. ๊ทธ ์์ ๋
ธํค์ ๋ฒฝ๋ํฐ์ ๊ธฐ์ตํ์๋ ๋ถ๋ค์ด๋ผ๋ฉด ์์ค ๊ฒ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ ์ค ๋ง์ ์ด๋ค์ด ์ฌ๋ํ๊ณ ์ง๊ธ๋ ์ฌ๋ํ๋ฉฐ, ๋ค์ ๋์์ค๊ธฐ๋ฅผ ๋ฐ๋ผ๋ ๋ธ๋๋ฒ ๋ฆฌ๋ถํฐ ํ์ฌ์ ์๋๋ก์ด๋ ๋ฐ ์์ดํฐ ์ค๋งํธํฐ์ ์ด๋ฅด๊ธฐ๊น์ง ๋ง์
๋๋ค. ๋ฐฐํฐ๋ฆฌ ์๋ช
์ธก๋ฉด์์ ์ ํฌ์ ์์์ ์ด ๋ฐ๋ก ๊ทธ๊ณณ์ด์์ต๋๋ค. ์ด๋ ์ค๋งํธํฐ ์๋๋ฅผ ์ด์์ต๋๋ค. ์ฝ 10๋ ์ , ์ํํธ๋ฑ ํฌ๊ฐ Arm์ ์ธ์ํ์ ๋๊ฐ ์ ํฌ์๊ฒ ์ฐพ์์จ ์ฃผ์ ์ ํ์ ์ค ํ๋์์ต๋๋ค. ๋ค, ์ฝ 10๋ ์ ์ธ 2016๋ ์ด์์ต๋๋ค. ์ํํธ๋ฑ ํฌ๊ฐ Arm์ ์ธ์ํ์ ๋, ์ ํ์ฅ์ ์ ํฌ๊ฐ ๋น์์ฅ ํ์ฌ๊ฐ ๋ ์์ ์์ ์ด์ ์๋ ํฌ์ํ ์ ์์๋ ๋ถ์ผ์ ํฌ์ํ ๊ธฐํ๋ฅผ ์ฃผ์์ต๋๋ค. |
| And that gave us the opportunity to expand the platform to a number of other verticals. We took everything that we knew about smartphones and then expanded that out into the cloud. We launched Neoverse. We got our first design wins in the data center. And then we were also able to invest into autonomous, automotive, physical AI. We could not have done that without that 2016 moment. And this is my thank you to Masa for allowing us to do that. We could not have made that all happen. It's paid significant benefits for the company. | ๊ทธ๋ฆฌ๊ณ ๊ทธ ๋๋ถ์ ์ ํฌ๋ ํ๋ซํผ์ ๋ค๋ฅธ ์ฌ๋ฌ ์ฐ์ ๋ถ์ผ๋ก ํ์ฅํ ๊ธฐํ๋ฅผ ์ป์์ต๋๋ค. ์ ํฌ๋ ์ค๋งํธํฐ์ ๋ํด ์๊ณ ์๋ ๋ชจ๋ ๊ฒ์ ํ์ฉํ์ฌ ํด๋ผ์ฐ๋ ๋ถ์ผ๋ก ํ์ฅํ์ต๋๋ค. ๋ค์ค๋ฒ์ค๋ฅผ ์ถ์ํ๊ณ , ๋ฐ์ดํฐ ์ผํฐ ๋ถ์ผ์์ ์ฒซ ๋ฒ์งธ ๋์์ธ ์์ ํ๋ณดํ์ต๋๋ค. ๋ํ ์์จ์ฃผํ, ์๋์ฐจ, ๋ฌผ๋ฆฌ์ AI ๋ถ์ผ์๋ ํฌ์ํ ์ ์์์ต๋๋ค. 2016๋ ์ ๊ทธ ์๊ฐ์ด ์์๋ค๋ฉด ์ด ๋ชจ๋ ๊ฒ์ ํด๋ผ ์ ์์์ ๊ฒ์ ๋๋ค. ์ ํฌ๊ฐ ๊ทธ๋ ๊ฒ ํ ์ ์๋๋ก ํ๋ฝํด ์ฃผ์ ๋ง์ฌ์๊ฒ ์ด ์๋ฆฌ๋ฅผ ๋น๋ ค ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ด ๋ชจ๋ ๊ฒ์ ๊ฐ๋ฅํ๊ฒ ํ ์ ์์์ ๊ฒ์ ๋๋ค. ์ด๋ ํ์ฌ์ ์๋นํ ์ด์ ์ ๊ฐ์ ธ๋ค์ฃผ์์ต๋๋ค. |
| However, as good as our products are, as competitive as the platform is for physical AI, for autonomous, for the cloud, it is really what I like to call the ecosystem of ecosystems that really differentiates us. And this is where the partnership really comes to life because that mobile platform that we built cannot happen without the software. And the software layer in the case of the mobile area is iOS, it's Windows, it's Android, it's MacOS. And then the litany of applications that not only run on the Arm compute platform, but they're highly optimized, highly tuned and allows the partners in the ecosystem to build great products. | ํ์ง๋ง ์ ํฌ ์ ํ์ด ์๋ฌด๋ฆฌ ํ๋ฅญํ๊ณ ๋ฌผ๋ฆฌ์ AI, ์์จ์ฃผํ, ํด๋ผ์ฐ๋ ๋ถ์ผ์์ ์ ํฌ ํ๋ซํผ์ด ์๋ฌด๋ฆฌ ๊ฒฝ์๋ ฅ์ด ์๋ค ํ๋๋ผ๋, ์ง์ ์ผ๋ก ์ ํฌ๋ฅผ ์ฐจ๋ณํํ๋ ๊ฒ์ ์ ๊ฐ '์ํ๊ณ ์ค์ ์ํ๊ณ(ecosystem of ecosystems)'๋ผ๊ณ ๋ถ๋ฅด๋ ๊ฒ์ ๋๋ค. ๋ฐ๋ก ์ด ์ง์ ์์ ํํธ๋์ญ์ด ์ง์ ์ผ๋ก ๋น์ ๋ฐํฉ๋๋ค. ์ ํฌ๊ฐ ๊ตฌ์ถํ ๋ชจ๋ฐ์ผ ํ๋ซํผ์ ์ํํธ์จ์ด ์์ด๋ ๊ตฌํ๋ ์ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. ๋ชจ๋ฐ์ผ ์์ญ์ ์ํํธ์จ์ด ๊ณ์ธต์ iOS, Windows, Android, MacOS์ด๋ฉฐ, ์ด์ ๋๋ถ์ด Arm ์ปดํจํธ ํ๋ซํผ์์ ์คํ๋ ๋ฟ๋ง ์๋๋ผ ๊ณ ๋๋ก ์ต์ ํ๋๊ณ ํ๋๋์ด ์ํ๊ณ ํํธ๋๋ค์ด ํ๋ฅญํ ์ ํ์ ๋ง๋ค ์ ์๋๋ก ๋๋ ์๋ง์ ์ ํ๋ฆฌ์ผ์ด์ ๋ค์ด ์์ต๋๋ค. |
| That formula applies to every vertical that we participate in. It applies to what takes place in the cloud, whether it's Linux or OpenAI or Anthropic and then the platform that runs with it. And this is why we like to call this the ecosystem of ecosystems because it's not just one vertical. And you can see when we look at the physical AI platform with automotive, same formula. 22 million-plus software developers that are very unique to a vertical, but they leverage a lot across the ecosystem that allows people to get started in other areas. So this is the magic, and this is what is uniquely Arm. It is what's very, very unique about our compute platform. | ์ ํฌ๊ฐ ์ฐธ์ฌํ๋ ๋ชจ๋ ์์ง ์์ฅ์ ์ด ๊ณต์์ด ์ ์ฉ๋ฉ๋๋ค. ํด๋ผ์ฐ๋์์ ๊ตฌํ๋๋ ๊ฒ, ์ฆ ๋ฆฌ๋
์ค๋ OpenAI๋ Anthropic์ด๋ , ๊ทธ๋ฆฌ๊ณ ๊ทธ ์์์ ์คํ๋๋ ํ๋ซํผ์๋ ์ ์ฉ๋์ฃ . ์ด๊ฒ์ด ๋ฐ๋ก ์ ํฌ๊ฐ ์ด๋ฅผ '์์ฝ์์คํ
์ค์ ์์ฝ์์คํ
'์ด๋ผ๊ณ ๋ถ๋ฅด๋ ์ด์ ์
๋๋ค. ๋จ์ผ ์์ง ์์ฅ์๋ง ๊ตญํ๋์ง ์๊ธฐ ๋๋ฌธ์
๋๋ค. ๊ทธ๋ฆฌ๊ณ ์๋์ฐจ ๋ถ์ผ์ ๋ฌผ๋ฆฌ์ AI ํ๋ซํผ์ ๋ณด์๋ฉด, ๋์ผํ ๊ณต์์ด ์ ์ฉ๋ฉ๋๋ค. 2,200๋ง ๋ช ์ด์์ ์ํํธ์จ์ด ๊ฐ๋ฐ์๋ค์ด ํน์ ์์ง ์์ฅ์ ๋งค์ฐ ํนํ๋์ด ์์ง๋ง, ์์ฝ์์คํ ์ ๋ฐ์ ๊ฑธ์ณ ๋ง์ ๊ฒ์ ํ์ฉํ์ฌ ๋ค๋ฅธ ๋ถ์ผ์์๋ ์ฝ๊ฒ ์์ํ ์ ์๋๋ก ํด์ค๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ๋ง๋ฒ์ด๋ฉฐ, Arm๋ง์ ๋ ํนํ ๊ฐ์ ์ ๋๋ค. ์ ํฌ ์ปดํจํ ํ๋ซํผ์ ์ ๋ง ๋ ๋ณด์ ์ธ ํน์ง์ด์ฃ . |
| There's no one on the planet who can serve the edge to the cloud in the way our ecosystem does. Now over the past few years, we've been evolving our strategies largely because we see the demands in the marketplace are around the chips are more complex. The cycle times to build these chips are getting longer, 5-nanometer to 3-nanometer to 2-nanometer means longer fab times, longer peg times. There's a need to do more and to do it faster. We've traditionally provided IP, IP in a stand-alone form, the CPU, the GPU, system IP. And that has served us well for the first 30-plus years of the company. | ์ ์ธ๊ณ์ ์ผ๋ก ์ฐ๋ฆฌ ์ํ๊ณ์ฒ๋ผ ์ฃ์ง์์ ํด๋ผ์ฐ๋๊น์ง ์๋น์ค๋ฅผ ์ ๊ณตํ ์ ์๋ ๊ณณ์ ์์ต๋๋ค. ์ต๊ทผ ๋ช ๋ ๊ฐ ์ ํฌ๋ ์ ๋ต์ ์งํ์์ผ ์์ต๋๋ค. ์ด๋ ์ฃผ๋ก ์์ฅ์ ์์๊ฐ ์นฉ์ ๋ณต์ก์ฑ ์ฆ๊ฐ์ ๊ด๋ จ๋์ด ์๊ธฐ ๋๋ฌธ์ ๋๋ค. 5๋๋ ธ๋ฏธํฐ์์ 3๋๋ ธ๋ฏธํฐ, 2๋๋ ธ๋ฏธํฐ๋ก ๋ฐ์ ํ๋ฉด์ ์นฉ ๊ฐ๋ฐ ์ฃผ๊ธฐ๊ฐ ๊ธธ์ด์ง๊ณ ์์ผ๋ฉฐ, ์ด๋ ๋ ๊ธด ์ ์กฐ(fab) ์๊ฐ๊ณผ ๋ ๊ธด ์กฐ๋ฆฝ(peg) ์๊ฐ์ ์๋ฏธํฉ๋๋ค. ๋ ๋ง์ ๊ฒ์ ๋ ๋น ๋ฅด๊ฒ ํด์ผ ํ ํ์๊ฐ ์์ต๋๋ค. ์ ํฌ๋ ์ ํต์ ์ผ๋ก ๋ ๋ฆฝํ IP, ์ฆ CPU, GPU, ์์คํ IP๋ฅผ ์ ๊ณตํด ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ๊ฒ์ ํ์ฌ ์ฐฝ๋ฆฝ ํ 30์ฌ ๋ ๊ฐ ์ ํฌ์๊ฒ ํฐ ๋์์ด ๋์์ต๋๋ค. |
| But as I said, we were starting to see huge demand for the need to go faster, make products better and get time to market sooner. And we introduced something called compute subsystems. We did this about 3, 4 years ago. We invested very heavily in terms of the engineering requirements to do this. And what this does is it takes all the blocks of IP and puts them together in a finished way, verified, performant, tested that the end customer can then take to market. And in some cases, it shaves a year, in some cases, 18 months off the time of starting design to get into production. It was a very significant investment for us. We put a lot of effort and engineering into it. | ํ์ง๋ง ๋ง์๋๋ ธ๋ฏ์ด, ์ ํฌ๋ ๋ ๋น ๋ฅด๊ฒ ๋์๊ฐ๊ณ , ์ ํ์ ๋ ์ข๊ฒ ๋ง๋ค๊ณ , ์์ฅ ์ถ์ ์๊ฐ์ ๋จ์ถํด์ผ ํ ํ์์ฑ์ ๋ํ ์์ฒญ๋ ์์๋ฅผ ๋ณด๊ธฐ ์์ํ์ต๋๋ค. ๊ทธ๋์ ์ ํฌ๋ '์ปดํจํธ ์๋ธ์์คํ
(compute subsystems)'์ด๋ผ๋ ๊ฒ์ ๋์
ํ์ต๋๋ค. ์ฝ 3~4๋
์ ์ ์ด๋ฅผ ์์ํ์ผ๋ฉฐ, ์ด๋ฅผ ์ํด ํ์ํ ์์ง๋์ด๋ง์ ๋ง๋ํ ํฌ์๋ฅผ ํ์ต๋๋ค. ์ด ์์คํ ์ ๋ชจ๋ IP ๋ธ๋ก์ ๊ฐ์ ธ์์, ๊ฒ์ฆ๋๊ณ , ์ฑ๋ฅ์ด ๋ฐ์ด๋๋ฉฐ, ํ ์คํธ๊น์ง ์๋ฃ๋ ํํ๋ก ๊ฒฐํฉํ์ฌ ์ต์ข ๊ณ ๊ฐ์ด ๋ฐ๋ก ์์ฅ์ ์ถ์ํ ์ ์๋๋ก ํฉ๋๋ค. ๊ฒฝ์ฐ์ ๋ฐ๋ผ์๋ ์ค๊ณ ์์๋ถํฐ ์์ฐ๊น์ง ๊ฑธ๋ฆฌ๋ ์๊ฐ์ 1๋ , ๊ธธ๊ฒ๋ 18๊ฐ์๊น์ง ๋จ์ถ์์ผ ์ค๋๋ค. ์ ํฌ์๊ฒ๋ ๋งค์ฐ ์๋ฏธ ์๋ ํฌ์์์ผ๋ฉฐ, ์ฌ๊ธฐ์ ๋ง์ ๋ ธ๋ ฅ๊ณผ ์์ง๋์ด๋ง ์ญ๋์ ์์๋ถ์์ต๋๋ค. |
| But we've already seen massive benefits in terms of the customer base. We introduced this 3 or 4 years ago. Our business model is a license plus royalty. Royalty is the laggard. So royalties start to show up 2, 3 years after we license a product. Already, CSS represents almost 20% of our royalties and growing. Now that's our evolution. But of course, we're now in an era where everything is different than we knew it before. And when I think about artificial intelligence, and I get a lot of questions when I talk to analysts or media about did AI just come up on us by surprise. | ํ์ง๋ง ์ ํฌ๋ ์ด๋ฏธ ๊ณ ๊ฐ ๊ธฐ๋ฐ ์ธก๋ฉด์์ ๋ง๋ํ ํํ์ ํ์ธํ์ต๋๋ค. ์ ํฌ๋ ์ด๋ฅผ 3, 4๋
์ ์ ๋์
ํ์ต๋๋ค. ์ ํฌ์ ์ฌ์
๋ชจ๋ธ์ ๋ผ์ด์ ์ค ๋ฐ ๋ก์ดํฐ์
๋๋ค. ๋ก์ดํฐ๋ ํํ ์งํ์
๋๋ค. ๋ฐ๋ผ์ ์ ํ ๋ผ์ด์ ์ค๋ฅผ ์ ๊ณตํ ํ 2, 3๋
์ด ์ง๋์ผ ๋ก์ดํฐ๊ฐ ๋ฐ์ํ๊ธฐ ์์ํฉ๋๋ค. ์ด๋ฏธ CSS๋ ์ ํฌ ๋ก์ดํฐ์ ๊ฑฐ์ 20%๋ฅผ ์ฐจ์งํ๊ณ ์์ผ๋ฉฐ, ๊ณ์ ์ฑ์ฅํ๊ณ ์์ต๋๋ค. ์ด๊ฒ์ด ์ ํฌ์ ๋ฐ์ ๊ณผ์ ์
๋๋ค. ํ์ง๋ง ๋ฌผ๋ก , ์ ํฌ๋ ์ด์ ์ด์ ์ ์๋ ๊ฒ๊ณผ๋ ๋ชจ๋ ๊ฒ์ด ๋ค๋ฅธ ์๋์ ์ด๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ธ๊ณต์ง๋ฅ์ ๋ํด ์๊ฐํด๋ณด๋ฉด, ์ ๋๋ฆฌ์คํธ๋ ์ธ๋ก ๊ณผ ๋ํํ ๋ AI๊ฐ ์ ํฌ์๊ฒ ๊ฐ์๊ธฐ ๋ฅ์ณ์จ ๊ฒ์ธ์ง์ ๋ํ ์ง๋ฌธ์ ๋ง์ด ๋ฐ์ต๋๋ค. |
| And I think back to a time I was in Bletchley Park about 1.5 years ago, and the Bletchley Park is where the original crypto work was done by Alan Turing to help the West against the Germans and World War II. There is an area there where you can go in the museum and you see papers from Alan Turing about can machines think. I think those papers were written in the 1940s. So the idea of AI is obviously not new. And if you're a sci-fi, aficionado or fan, I certainly was growing up. Arthur C. Clark was one of my favorite authors. 2001 A Space Odyssey, now we have people who weren't even born in 2001, who are here. I always looked at this and said, of course, this is going to happen. | ๊ทธ๋ฆฌ๊ณ ์ฝ 1๋
๋ฐ ์ ๋ธ๋ ์ธจ๋ฆฌ ํํฌ์ ์์๋ ๋๋ฅผ ๋ ์ฌ๋ ค ๋ด
๋๋ค. ๋ธ๋ ์ธจ๋ฆฌ ํํฌ๋ 2์ฐจ ์ธ๊ณ๋์ ๋น์ ์จ๋ฐ ํ๋ง์ด ์๋ฐฉ ์ฐํฉ๊ตฐ์ด ๋
์ผ์ ๋ง์ ์ธ์ฐ๋ ๊ฒ์ ๋๊ธฐ ์ํด ์ต์ด์ ์ํธ ํด๋
์์
์ ์ํํ๋ ๊ณณ์
๋๋ค. ๊ทธ๊ณณ ๋ฐ๋ฌผ๊ด์ ๊ฐ๋ฉด ์จ๋ฐ ํ๋ง์ด '๊ธฐ๊ณ๋ ์๊ฐํ ์ ์๋๊ฐ'์ ๋ํด ์ด ๋
ผ๋ฌธ๋ค์ ๋ณผ ์ ์๋ ๊ตฌ์ญ์ด ์์ต๋๋ค. ์ ์๊ฐ์ ๊ทธ ๋
ผ๋ฌธ๋ค์ 1940๋
๋์ ์ฐ์ฌ์ก์ ๊ฒ๋๋ค. ๋ฐ๋ผ์ ์ธ๊ณต์ง๋ฅ์ด๋ผ๋ ๊ฐ๋
์ ๋ถ๋ช
ํ ์๋ก์ด ๊ฒ์ด ์๋๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ง์ฝ ์ฌ๋ฌ๋ถ์ด ๊ณต์ ๊ณผํ ์์ค์ ์ ํธ๊ฐ๋ ํฌ์ด๋ผ๋ฉด, ์ ์ญ์ ์ด๋ฆด ์ ์๋ ๋ถ๋ช ํ ๊ทธ๋ฌ์ต๋๋ค. ์์ C. ํด๋ผํฌ๋ ์ ๊ฐ ๊ฐ์ฅ ์ข์ํ๋ ์๊ฐ ์ค ํ ๋ช ์ด์์ต๋๋ค. '2001 ์คํ์ด์ค ์ค๋์ธ์ด'๋ฅผ ๋ณด๋ฉด, ์ง๊ธ ์ด ์๋ฆฌ์๋ 2001๋ ์ ํ์ด๋์ง๋ ์์ ๋ถ๋ค๋ ๊ณ์์ฃ . ์ ๋ ํญ์ ์ด๊ฒ์ ๋ณด๋ฉด์ '๋ฌผ๋ก , ์ด๋ฐ ์ผ์ ์ผ์ด๋ ์๋ฐ์ ์๋ค'๊ณ ์๊ฐํ์ต๋๋ค. |
| I just didn't think in my lifetime, I would see it at the pace that we've seen it. And for anyone who says this is a bubble and it's going to pass, it may be a financial bubble in the case of investment may slow down and it may be an investment bubble in the sense of the valuations may not be what they are today tomorrow. But if anyone thinks that this is something that is going to go away, it's a little bit of an ostrich syndrome. This is here with us. And it's really changed how people think about computing. However, somewhere along the way, people kind of thought CPUs were dead. | ๋ด ํ์์ ์ฐ๋ฆฌ๊ฐ ๋ณธ ์๋๋ก ์ด๋ฐ ๋ณํ๋ฅผ ๋ณด๊ฒ ๋ ๊ฑฐ๋ผ๊ณ ๋ ์ ๋ง ์๊ฐ ๋ชป ํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๊ฒ์ด ๊ฑฐํ์ด๊ณ ๊ณง ์ฌ๋ผ์ง ๊ฒ์ด๋ผ๊ณ ๋งํ๋ ์ฌ๋๋ค์๊ฒ๋, ํฌ์ ๋ํ๋ก ์ธํ ๊ธ์ต ๊ฑฐํ์ผ ์๋ ์๊ณ , ๋ด์ผ์ ๊ฐ์น ํ๊ฐ๊ฐ ์ค๋๊ณผ ๊ฐ์ง ์์ ์ ์๋ค๋ ์๋ฏธ์์ ํฌ์ ๊ฑฐํ์ผ ์๋ ์์ต๋๋ค. ํ์ง๋ง ์ด๊ฒ์ด ์ฌ๋ผ์ง ๊ฒ์ด๋ผ๊ณ ์๊ฐํ๋ ์ฌ๋์ด ์๋ค๋ฉด, ๊ทธ๊ฒ์ ์ฝ๊ฐ์ ํ์กฐ ์ฆํ๊ตฐ์ ๋๋ค. ์ด๊ฒ์ ์ฐ๋ฆฌ์ ํจ๊ปํ ๊ฒ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๊ฒ์ ์ฌ๋๋ค์ด ์ปดํจํ ์ ๋ํด ์๊ฐํ๋ ๋ฐฉ์์ ์ ๋ง๋ก ๋ฐ๊ฟ๋์์ต๋๋ค. ํ์ง๋ง ์ด๋ ์์ ๋ถํฐ์ธ๊ฐ, ์ฌ๋๋ค์ CPU๊ฐ ์ฃฝ์๋ค๊ณ ์๊ฐํ๋ ๊ฒฝํฅ์ด ์์์ต๋๋ค. |
| And there was a thought that the only way you handle AI is through accelerated computing, that the CPU's role in the AI world is no longer relevant. Now if we think about the role of the CPU and what happens in the cloud, now this is the cloud before AI. So I'm going to say it's before that last slide that I showed. Huge growth in compute cloud. We saw growth from AWS, Microsoft, GCP. And the conventional use of the cloud was you type in an answer, you do a search, -- any seats left for the Warrior's game? I think there are a lot of seats left for tomorrow's game, by the way, I have seen or tonight's game. You got the prompt back. This is the cloud. | AI๋ฅผ ๋ค๋ฃจ๋ ์ ์ผํ ๋ฐฉ๋ฒ์ ๊ฐ์ ์ปดํจํ ๋ฟ์ด๋ฉฐ, AI ์ธ์์์ CPU์ ์ญํ ์ ๋ ์ด์ ์ค์ํ์ง ์๋ค๋ ์๊ฐ์ด ์์์ต๋๋ค. ์ด์ CPU์ ์ญํ ๊ณผ ํด๋ผ์ฐ๋์์ ์ผ์ด๋๋ ์ผ๋ค์ ์๊ฐํด ๋ณด๋ฉด, ์ด๊ฒ์ AI ์ด์ ์ ํด๋ผ์ฐ๋์ ๋๋ค. ์ ๊ฐ ๋ณด์ฌ๋๋ฆฐ ๋ง์ง๋ง ์ฌ๋ผ์ด๋ ์ด์ ์ ์ํฉ์ด๋ผ๊ณ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ปดํจํ ํด๋ผ์ฐ๋๋ ์์ฒญ๋๊ฒ ์ฑ์ฅํ์ต๋๋ค. AWS, ๋ง์ดํฌ๋ก์ํํธ, GCP์์ ์ฑ์ฅ์ธ๋ฅผ ๋ณด์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ํด๋ผ์ฐ๋์ ๊ธฐ์กด ์ฉ๋๋ ๋ต๋ณ์ ์ ๋ ฅํ๊ณ ๊ฒ์์ ํ๋ ๊ฒ์ด์์ต๋๋ค. ์๋ฅผ ๋ค์ด, "์๋ฆฌ์ด์ค ๊ฒฝ๊ธฐ ์ข์์ด ๋จ์์๋์?" ๊ฐ์ ์ง๋ฌธ์ ํ๋ฉด (๊ทธ๋์ ๋, ์ ๊ฐ ๋ดค์ ๋๋ ๋ด์ผ ๊ฒฝ๊ธฐ๋ ์ค๋ ๊ฒฝ๊ธฐ ์ข์์ด ๋ง์ด ๋จ์์๋ ๊ฒ ๊ฐ๋๊ตฐ์), ๋ต๋ณ์ ๋๋ ค๋ฐ๋ ์์ด์์ฃ . ์ด๊ฒ์ด ๋ฐ๋ก ํด๋ผ์ฐ๋์์ต๋๋ค. |
| Very simple, you do search, but CPU is very heavy. So when we look at the growth of SaaS 10-plus years ago, 10, 15 years ago and all the growth around cloud, the CPUs were doing literally all the work. Now when you add the AI cloud, if you will, and now you are a human and you're putting in a prompt into your device, whether it's your phone or your PC, Well, of course, there are still CPUs involved. The cloud is servicing that request, and that request gets sent for a token, which the accelerator generates and a CPU in that data center orchestrates and sends a token back, the token being a word or an answer that provides the request for the query. | ๊ฐ๋จํ ๋งํด์, ๊ฒ์์ ํ ๋ CPU ๋ถํ๊ฐ ๋งค์ฐ ํฝ๋๋ค. ๊ทธ๋์ 10์ฌ ๋
์ , 10๋
์์ 15๋
์ SaaS์ ์ฑ์ฅ๊ณผ ํด๋ผ์ฐ๋ ๊ด๋ จ ๋ชจ๋ ์ฑ์ฅ์ ๋ณด๋ฉด, CPU๊ฐ ๋ง ๊ทธ๋๋ก ๋ชจ๋ ์์
์ ์ฒ๋ฆฌํ์ต๋๋ค. ์ด์ AI ํด๋ผ์ฐ๋๋ผ๊ณ ํ ๊น์, ์ด๊ฒ์ ์ถ๊ฐํ๊ณ , ์ฌ๋์ด ํด๋ํฐ์ด๋ PC๋ ์์ ์ ๊ธฐ๊ธฐ์ ํ๋กฌํํธ๋ฅผ ์ ๋ ฅํ๋ค๊ณ ๊ฐ์ ํด ๋ด ์๋ค. ๋ฌผ๋ก , ์ฌ์ ํ CPU๊ฐ ๊ด์ฌํฉ๋๋ค. ํด๋ผ์ฐ๋๊ฐ ํด๋น ์์ฒญ์ ์ฒ๋ฆฌํ๊ณ , ๊ทธ ์์ฒญ์ ๊ฐ์๊ธฐ๊ฐ ์์ฑํ๋ ํ ํฐ์ ์ํด ์ ์ก๋๋ฉฐ, ํด๋น ๋ฐ์ดํฐ์ผํฐ์ CPU๊ฐ ์ด๋ฅผ ์กฐ์จํ์ฌ ํ ํฐ์ ๋ค์ ๋ณด๋ ๋๋ค. ์ด ํ ํฐ์ ์ฟผ๋ฆฌ ์์ฒญ์ ๋ํ ์๋ต์ ์ ๊ณตํ๋ ๋จ์ด ๋๋ ๋ต๋ณ์ ๋๋ค. |
| So this is all the work that's being done by the AI data center. So CPUs are involved both in the cloud and obviously, they're involved in the AI data center. And we estimate that in this data center, there's probably 30 million CPU cores per gigawatt. So there's a lot. And data center here is a combination of what sits right in the AI cluster, whether it's your head node to your accelerator or what sits next to a dedicated rack. But the math is basically about 30 million CPU cores per gigawatt, okay? And that is the world that we've seen coming up to about the last year or so or maybe even less. And what has changed in the last number of months has been this explosion of agents. | ์ด๊ฒ์ด ๋ฐ๋ก AI ๋ฐ์ดํฐ ์ผํฐ๊ฐ ์ํํ๋ ๋ชจ๋ ์์
์
๋๋ค. ๋ฐ๋ผ์ CPU๋ ํด๋ผ์ฐ๋์ ๋น์ฐํ AI ๋ฐ์ดํฐ ์ผํฐ ๋ชจ๋์ ์ฌ์ฉ๋ฉ๋๋ค. ์ ํฌ๋ ์ด ๋ฐ์ดํฐ ์ผํฐ์ ๊ธฐ๊ฐ์ํธ๋น ์ฝ 3์ฒ๋ง ๊ฐ์ CPU ์ฝ์ด๊ฐ ์์ ๊ฒ์ผ๋ก ์ถ์ ํฉ๋๋ค. ๊ทธ ์์ด ์๋นํฉ๋๋ค. ์ฌ๊ธฐ์ ๋ฐ์ดํฐ ์ผํฐ๋ AI ํด๋ฌ์คํฐ ๋ด๋ถ์ ์ง์ ์์นํ๋ ๊ตฌ์ฑ ์์๋ค, ์๋ฅผ ๋ค์ด ํค๋ ๋ ธ๋๋ถํฐ ๊ฐ์๊ธฐ๊น์ง, ๋๋ ์ ์ฉ ๋ ์์ ์๋ ๊ตฌ์ฑ ์์๋ค์ ๋ชจ๋ ํฌํจํฉ๋๋ค. ํ์ง๋ง ํต์ฌ์ ์ธ ์์น๋ ๊ธฐ๋ณธ์ ์ผ๋ก ๊ธฐ๊ฐ์ํธ๋น ์ฝ 3์ฒ๋ง ๊ฐ์ CPU ์ฝ์ด์ ๋๋ค. ์์๊ฒ ์ฃ ? ๊ทธ๋ฆฌ๊ณ ์ด๊ฒ์ด ๋ฐ๋ก ์ ํฌ๊ฐ ์ง๋ 1๋ ์ ๋, ์ด์ฉ๋ฉด ๊ทธ๋ณด๋ค ๋ ์งง์ ๊ธฐ๊ฐ ๋์ ๋ชฉ๊ฒฉํด ์จ ์ํฉ์ ๋๋ค. ์ง๋ ๋ช ๋ฌ ๋์ ๋ณํํ ๊ฒ์ ๋ฐ๋ก ์์ด์ ํธ์ ํญ๋ฐ์ ์ธ ์ฆ๊ฐ์ ๋๋ค. |
| Agents are essentially tools that act on a request and come back with a full flow of answers. So it's not just a query for an answer, but it's actually work. It's run a payroll task, do a scheduler, go off and write a number of analyses relative to a tool flow and provide me an answer. And we heard so much about OpenClaw here in the last few weeks as an example, and it's not the only example. Now why is this important? Why am I talking about this? Because as we move to agentic query, the number of tokens per human go up by 15x, if not greater. And if you think about the why of that, it's pretty straightforward. | ์์ด์ ํธ๋ ๋ณธ์ง์ ์ผ๋ก ์์ฒญ์ ๋ฐ์ ์ฒ๋ฆฌํ๊ณ , ๊ทธ์ ๋ํ ํฌ๊ด์ ์ธ ๋ต๋ณ ํ๋ฆ์ ์ ๊ณตํ๋ ๋๊ตฌ์
๋๋ค. ๋ฐ๋ผ์ ์ด๋ ๋จ์ํ ๋ต๋ณ์ ์ป๊ธฐ ์ํ ์ง์๊ฐ ์๋๋ผ, ์ค์ ์
๋ฌด๋ฅผ ์ํํ๋ ๊ฒ์
๋๋ค. ๊ธ์ฌ ์
๋ฌด๋ฅผ ์คํํ๊ณ , ์ค์ผ์ค๋ฌ๋ฅผ ์๋์ํค๋ฉฐ, ๋๊ตฌ ์ฐ๋ ํ๋ฆ์ ๋ง์ถฐ ์ฌ๋ฌ ๋ถ์์ ์์ฑํ ํ ์ ์๊ฒ ๋ต๋ณ์ ์ ๊ณตํ๋ ์์
๋๋ค. ์ง๋ ๋ช ์ฃผ๊ฐ ์ฌ๊ธฐ์ ์คํํด๋ก(OpenClaw)์ ๋ํด ๋ง์ด ์ธ๊ธ๋์๋๋ฐ, ์ด๊ฒ์ด ์ ์ผํ ์์๋ ์๋๋๋ค. ์, ์ด๊ฒ์ด ์ ์ค์ํ ๊น์? ์ ๊ฐ ์ ์ด ๋ง์์ ๋๋ฆฌ๋ ๊ฑธ๊น์? ์์ด์ ํธํ ์ง์๋ก ์ ํ๋ ์๋ก, ์ฌ์ฉ์๋น ํ ํฐ(token) ์๊ฐ 15๋ฐฐ, ํน์ ๊ทธ ์ด์ ์ฆ๊ฐํ๊ธฐ ๋๋ฌธ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ ์ด์ ๋ฅผ ์๊ฐํด ๋ณด๋ฉด, ๋งค์ฐ ๋ช ํํฉ๋๋ค. |
| Agents can generate requests, a, far faster than humans; and b, they don't sleep. They're at a 24/7. So the agents are now pushing these requests into the cloud into the data center and what's happening? The data center is choking. These accelerators, which are very expensive, that generate the tokens now need to send those tokens back through the cloud. Now if we think about what an agent is, an agent is a workflow. As I said, it's a payroll task, it's a scheduler task. It's asynchronous. It is a lot of work relative to scheduling. That's what CPUs do. That is what CPUs do. That is not a work that can be done by an accelerator. | ์์ด์ ํธ๋ ์ฒซ์งธ, ์ธ๊ฐ๋ณด๋ค ํจ์ฌ ๋น ๋ฅด๊ฒ ์์ฒญ์ ์์ฑํ ์ ์์ต๋๋ค. ๋์งธ, ์ ์ ์์ง ์์ต๋๋ค. 24์๊ฐ ๋ด๋ด ์๋ํ์ฃ . ๊ทธ๋์ ์์ด์ ํธ๋ค์ ์ด์ ์ด ์์ฒญ๋ค์ ํด๋ผ์ฐ๋, ์ฆ ๋ฐ์ดํฐ์ผํฐ๋ก ๋ฐ์ด ๋ฃ๊ณ ์๋๋ฐ, ๋ฌด์จ ์ผ์ด ๋ฒ์ด์ง๊ณ ์์๊น์? ๋ฐ์ดํฐ์ผํฐ๊ฐ ๋ง๋น๋๊ณ ์์ต๋๋ค. ํ ํฐ์ ์์ฑํ๋ ์ด ๊ฐ์๊ธฐ๋ค์ ๋งค์ฐ ๋น์ผ๋ฐ๋ ๋ถ๊ตฌํ๊ณ , ์ด์ ๊ทธ ํ ํฐ๋ค์ ๋ค์ ํด๋ผ์ฐ๋๋ฅผ ํตํด ๋ณด๋ด์ผ ํฉ๋๋ค. ์ด์ ์์ด์ ํธ๊ฐ ๋ฌด์์ธ์ง ์๊ฐํด ๋ณด๋ฉด, ์์ด์ ํธ๋ ์ํฌํ๋ก์ฐ์ ๋๋ค. ์ ๊ฐ ๋ง์๋๋ ธ๋ฏ์ด, ๊ธ์ฌ ์ฒ๋ฆฌ ์์ , ์ค์ผ์ค๋ฌ ์์ ๊ณผ ๊ฐ์ ๊ฒ์ ๋๋ค. ๋น๋๊ธฐ์ ์ด์ฃ . ์ค์ผ์ค๋ง๊ณผ ๊ด๋ จํ์ฌ ๋ง์ ์์ ์ ์ฒ๋ฆฌํฉ๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก CPU๊ฐ ํ๋ ์ผ์ ๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก CPU๊ฐ ํ๋ ์ผ์ ๋๋ค. ์ด๊ฒ์ ๊ฐ์๊ธฐ๊ฐ ์ฒ๋ฆฌํ ์ ์๋ ์์ ์ด ์๋๋๋ค. |
| The way to think about this is the accelerator generates the tokens, but it's almost like pushing a dump truck up and someone's got to move all that dirt. The CPUs are the pieces of equipment that move that dirt and Agentic AI only increases that. So what you see is a huge bottleneck now in terms of flow. So what does that mean? You need more and more CPUs, lots of them. CPUs near the head node, CPUs next to the accelerator rack, more CPU racks inside the data center, you just need more. And by our calculations, and we think this may be a little bit light, goes up about 4x, 120 million CPU cores for that same gigawatt, okay? So in that same profile, we now need 120 million CPU cores. | ์ด๋ฅผ ์ด๋ ๊ฒ ์๊ฐํ์๋ฉด ๋ฉ๋๋ค. ๊ฐ์๊ธฐ(accelerator)๊ฐ ํ ํฐ์ ์์ฑํ์ง๋ง, ์ด๋ ๋ง์น ๋คํํธ๋ญ์ด ํ์ ์์๋ด๋ ๊ฒ๊ณผ ๊ฐ๊ณ , ๋๊ตฐ๊ฐ๋ ๊ทธ ํ์ ๋ชจ๋ ์ฎ๊ฒจ์ผ ํฉ๋๋ค. CPU๋ ๊ทธ ํ์ ์ฎ๊ธฐ๋ ์ฅ๋น์ด๋ฉฐ, ์์ด์ ํธ AI(Agentic AI)๋ ์ด๋ฌํ ์๊ตฌ๋ฅผ ๋์ฑ ์ฆ๊ฐ์ํต๋๋ค. ๋ฐ๋ผ์ ํ์ฌ ๋ฐ์ดํฐ ํ๋ฆ ์ธก๋ฉด์์ ์์ฒญ๋ ๋ณ๋ชฉ ํ์์ด ๋ฐ์ํ๊ณ ์์ต๋๋ค. ๊ทธ๋ผ ์ด๊ฒ ๋ฌด์์ ์๋ฏธํ ๊น์? ์ ์ ๋ ๋ง์ CPU๊ฐ, ๊ทธ๊ฒ๋ ์์ฃผ ๋ง์ด ํ์ํ๋ค๋ ๋ป์ ๋๋ค. ํค๋ ๋ ธ๋ ๊ทผ์ฒ์ CPU, ๊ฐ์๊ธฐ ๋ ์์ CPU, ๋ฐ์ดํฐ ์ผํฐ ๋ด์ ๋ ๋ง์ CPU ๋ ๋ฑ, ๊ทธ์ ๋ ๋ง์ด ํ์ํฉ๋๋ค. ์ ํฌ ๊ณ์ฐ์ ๋ฐ๋ฅด๋ฉด, ์ด ์์น๊ฐ ๋ค์ ๋ณด์์ ์ผ ์ ์๋ค๊ณ ์๊ฐํ์ง๋ง, ๋์ผํ ๊ธฐ๊ฐ์ํธ(gigawatt) ์ ๋ ฅ๋์์ ์ฝ 4๋ฐฐ ์ฆ๊ฐํ 1์ต 2์ฒ๋ง ๊ฐ์ CPU ์ฝ์ด๊ฐ ํ์ํฉ๋๋ค. ์์๊ฒ ์ฃ ? ๋ฐ๋ผ์ ๋์ผํ ์กฐ๊ฑด์์, ์ด์ ์ฐ๋ฆฌ๋ 1์ต 2์ฒ๋ง ๊ฐ์ CPU ์ฝ์ด๊ฐ ํ์ํฉ๋๋ค. |
| Now we're trying to put 4x the amount of CPU cores in that same power envelope. Power is precious, obviously. The capital required for it is precious. So trying to put all those extra CPUs into a data center that is already stuffed to the brim with accelerators and CPUs doing the core work, that is a problem. [Presentation] Rene Haas CEO & Director Now every tough problem needs a good solution. And we're announcing our first silicon chip that we are selling to customers for revenue. The Arm AGI CPU. Now this is a big, big deal. And I would love to tell you every feed and speed about the product right now, but Mohammad will kill me if I do that. | ํ์ฌ ์ฐ๋ฆฌ๋ ๋์ผํ ์ ๋ ฅ ๋ฒ์ ๋ด์์ 4๋ฐฐ ๋ ๋ง์ CPU ์ฝ์ด๋ฅผ ํ์ฌํ๋ ค๊ณ ๋
ธ๋ ฅํ๊ณ ์์ต๋๋ค. ์ ๋ ฅ์ ๋น์ฐํ ๊ทํฉ๋๋ค. ๊ทธ๊ฒ์ ํ์ํ ์๋ณธ ๋ํ ๊ทํฉ๋๋ค. ๋ฐ๋ผ์, ํต์ฌ ์์
์ ์ํํ๋ ๊ฐ์๊ธฐ์ CPU๋ค๋ก ์ด๋ฏธ ํฌํ ์ํ์ธ ๋ฐ์ดํฐ ์ผํฐ์ ์ด ๋ชจ๋ ์ถ๊ฐ CPU๋ค์ ๋ฃ์ผ๋ ค๋ ์๋๋ ๋ฌธ์ ๊ฐ ๋ฉ๋๋ค. **๋ฅด๋ค ํ์ค (CEO ๊ฒธ ์ด์ฌ)** ์ด์ ๋ชจ๋ ์ด๋ ค์ด ๋ฌธ์ ์๋ ์ข์ ํด๊ฒฐ์ฑ ์ด ํ์ํฉ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ค๋, ์ ํฌ๋ ๊ณ ๊ฐ์๊ฒ ํ๋งคํ์ฌ ์์ต์ ์ฐฝ์ถํ ์ฒซ ๋ฒ์งธ ์ค๋ฆฌ์ฝ ์นฉ์ ๋ฐํํฉ๋๋ค. ๋ฐ๋ก Arm AGI CPU์ ๋๋ค. ์ด๊ฒ์ ์ ๋ง ์์ฒญ๋ ์ผ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ๋ ์ง๊ธ ๋น์ฅ ์ด ์ ํ์ ๋ชจ๋ ์ธ๋ถ ์ฌ์๊ณผ ์ฑ๋ฅ์ ๋ํด ๋ง์๋๋ฆฌ๊ณ ์ถ์ง๋ง, ๊ทธ๋ ๊ฒ ํ๋ฉด ๋ชจํ๋ง๋๊ฐ ์ ๋ฅผ ๊ฐ๋ง๋์ง ์์ ๊ฒ๋๋ค. |
| So we'll go into a lot of detail about the product and how we conceived it and the why. But let me be clear, we are now in a new business for Arm, and we are supplying CPUs as chips. The biggest reason we're doing this is that our partners have asked for it. But we're also really doing this to solve the problem I just described. As Agentic AI becomes mainstream, all of the work required to make that happen is CPU bound and you need a CPU that has the DNA of being born to run off a battery. So as I said, -- reason 0 is our partners have asked us for it. And one of the partners we work the closest with on this is Meta. | ์, ๊ทธ๋ผ ์ ํ๊ณผ ์ ํฌ๊ฐ ์ด๋ป๊ฒ ์ด๋ฅผ ๊ตฌ์ํ๋์ง, ๊ทธ๋ฆฌ๊ณ ๊ทธ ์ด์ ์ ๋ํด ์์ธํ ์ค๋ช ๋๋ฆฌ๊ฒ ์ต๋๋ค. ํ์ง๋ง ๋ถ๋ช ํ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ ํฌ Arm์ ์ด์ ์๋ก์ด ์ฌ์ ์์ญ์ ์ง์ถํ์ผ๋ฉฐ, ์นฉ ํํ๋ก CPU๋ฅผ ๊ณต๊ธํ๊ณ ์์ต๋๋ค. ์ ํฌ๊ฐ ์ด๋ ๊ฒ ํ๋ ๊ฐ์ฅ ํฐ ์ด์ ๋ ํํธ๋๋ค์ด ์์ฒญํ๊ธฐ ๋๋ฌธ์ ๋๋ค. ํ์ง๋ง ์ ํฌ๋ ๋ํ ์ ๊ฐ ๋ฐฉ๊ธ ์ค๋ช ๋๋ฆฐ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด ์ด ์ผ์ ์ถ์งํ๊ณ ์์ต๋๋ค. ์์ด์ ํธํ AI๊ฐ ์ฃผ๋ฅ๊ฐ ๋จ์ ๋ฐ๋ผ, ์ด๋ฅผ ๊ตฌํํ๋ ๋ฐ ํ์ํ ๋ชจ๋ ์์ ์ CPU์ ์ข ์๋๋ฉฐ, ๋ฐฐํฐ๋ฆฌ ๊ตฌ๋์ ์ํด ํ์ด๋ DNA๋ฅผ ๊ฐ์ง CPU๊ฐ ํ์ํฉ๋๋ค. ๊ทธ๋์ ์์ ๋ง์๋๋ ธ๋ฏ์ด, 0์์ ์ด์ ๋ ์ ํฌ ํํธ๋๋ค์ด ์์ฒญํ๊ธฐ ๋๋ฌธ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด์ ๊ด๋ จํ์ฌ ์ ํฌ๊ฐ ๊ฐ์ฅ ๊ธด๋ฐํ๊ฒ ํ๋ ฅํ๋ ํํธ๋ ์ค ํ๋๋ Meta์ ๋๋ค. |
| And I'm super pleased to have Santosh Janardhan with me today, who's going to do a better job than I can to tell you why Meta made that choice. Santosh? Santosh Janardhan Meta Platforms, Inc. Hey folks, welcome. Every year, I try to run the San Francisco half Marathon and they distribute the bits the day before you run right here. I can tell you, it looks very, very different compared to what I guess you're seeing now. So hi, my name is Santosh Janardhan. I lead infrastructure at Meta. So what does that mean? Well, it means that we traditionally go and custom build and design our data centers, we run it. | ์ค๋ Santosh Janardhan ์จ๋ฅผ ๋ชจ์๊ฒ ๋์ด ์ ๋ง ๊ธฐ์ฉ๋๋ค. ๋ฉํ๊ฐ ์ ๊ทธ๋ฐ ์ ํ์ ํ๋์ง ์ ๋ณด๋ค ๋ ํ๋ฅญํ๊ฒ ์ค๋ช
ํด ์ฃผ์ค ๊ฒ๋๋ค. Santosh? **Santosh Janardhan (๋ฉํ ํ๋ซํผ์ค)** ์๋ ํ์ธ์, ํ์ํฉ๋๋ค. ๋งค๋ ์ ๋ ์ํ๋์์ค์ฝ ํํ ๋ง๋ผํค์ ์ฐธ๊ฐํ๋ ค๊ณ ๋ ธ๋ ฅํ๋๋ฐ์, ๋ฐ๋ก ์ด๊ณณ์์ ๊ฒฝ๊ธฐ ์ ๋ ๋ฐฐ๋ฒ(bibs)์ ๋ฐฐ๋ถํฉ๋๋ค. ์์งํ ๋ง์๋๋ฆฌ์๋ฉด, ์ง๊ธ ์ฌ๋ฌ๋ถ์ด ๋ณด์๋ ๊ฒ๊ณผ๋ ๋งค์ฐ, ๋งค์ฐ ๋ค๋ฅธ ๋ชจ์ต์ ๋๋ค. ๋ค, ์ ๋ Santosh Janardhan์ ๋๋ค. ๋ฉํ์์ ์ธํ๋ผ๋ฅผ ์ด๊ดํ๊ณ ์์ต๋๋ค. ๊ทธ๋ผ ๊ทธ๊ฒ ๋ฌด์์ ์๋ฏธํ ๊น์? ์ผ๋ฐ์ ์ผ๋ก ์ ํฌ๋ ๋ฐ์ดํฐ์ผํฐ๋ฅผ ๋ง์ถคํ์ผ๋ก ๊ตฌ์ถํ๊ณ ์ค๊ณํ๋ฉฐ, ์ง์ ์ด์ํ๋ค๋ ๋ป์ ๋๋ค. |
| We custom build our hardware, our GPUs, our CPUs, and we'll get into that quite a bit, the network that connects them. And obviously, the software that sort of binds it all together. It's a fancy way to say that if your Instagram is not working, if your WhatsApp is not working, your message is not arriving, I am the person to blame. Now if you think through our family of apps, that marks about 3 billion, 3.5 billion users that use our products daily. Every single day, about half of humanity logs into one of our sort of apps and hammers away at it. And as you can imagine, that creates a decent amount of scale. We run a decent amount of the Internet. | ์ ํฌ๋ ํ๋์จ์ด, GPU, CPU๋ฅผ ์ง์ ๋ง์ถค ์ ์ํ๋ฉฐ, ์ด๋ค์ ์ฐ๊ฒฐํ๋ ๋คํธ์ํฌ์ ๋ํด์๋ ๋์ค์ ์์ธํ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ๋ฌผ๋ก , ์ด ๋ชจ๋ ๊ฒ์ ํ๋๋ก ๋ฌถ๋ ์ํํธ์จ์ด๊น์ง ํฌํจํด์ ๋ง์ด์ฃ . ์ฝ๊ฒ ๋งํด, ๋ง์ฝ ์ธ์คํ๊ทธ๋จ์ด ์๋ํ์ง ์๊ฑฐ๋, ์์ธ ์ฑ์ด ์ ๋๊ฑฐ๋, ๋ฉ์์ง๊ฐ ๋์ฐฉํ์ง ์๋๋ค๋ฉด, ๊ทธ๊ฑด ๋ฐ๋ก ์ ์ฑ
์์ด๋ผ๋ ๋ป์
๋๋ค. ์ ํฌ ์ฑ ์ ํ๊ตฐ์ ์๊ฐํด ๋ณด์๋ฉด, ๋งค์ผ ์ ํฌ ์ ํ์ ์ฌ์ฉํ๋ ์ฌ์ฉ์๊ฐ ์ฝ 30์ต์์ 35์ต ๋ช ์ ๋ฌํฉ๋๋ค. ๋งค์ผ, ์ธ๋ฅ์ ์ ๋ฐ ๊ฐ๋์ด ์ ํฌ ์ฑ ์ค ํ๋์ ๋ก๊ทธ์ธํ์ฌ ํ๋ฐํ๊ฒ ์ด์ฉํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ง์ํ์๊ฒ ์ง๋ง, ์ด๋ ์๋นํ ๊ท๋ชจ๋ฅผ ๋ง๋ค์ด๋ ๋๋ค. ์ ํฌ๊ฐ ์ธํฐ๋ท์ ์๋น ๋ถ๋ถ์ ์ด์ํ๊ณ ์๋ ์ ์ ๋๋ค. |
| And we're probably the only hyperscaler that's not a cloud, right? So if you think about gigawatts of capacity, tens of millions of servers and increasingly, more and more, you're seeing bigger and bigger CPU and GPU, AI clusters. Rene sort of went through that quite a bit. I think it's interesting to go and look at how this has grown over the last years. AI clusters are a fairly new thing, really started sort of post-COVID 2022, '23, just after sort of a ChatGPT came along. And our initial clusters were pretty small. In fact, when I look back for this, in '23, our initial clusters are about 128 GPUs. That's it. But as you can see, even in '23, we started scaling quite a bit. | ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ ์๋ง๋ ํด๋ผ์ฐ๋๊ฐ ์๋ ์ ์ผํ ํ์ดํผ์ค์ผ์ผ๋ฌ์ผ ๊ฒ๋๋ค, ๊ทธ๋ ์ฃ ? ๊ธฐ๊ฐ์ํธ๊ธ ์ฉ๋, ์์ฒ๋ง ๋์ ์๋ฒ๋ฅผ ์๊ฐํด ๋ณด์๋ฉด, ์ ์ ๋ ์ปค์ง๋ CPU์ GPU ๊ธฐ๋ฐ AI ํด๋ฌ์คํฐ๋ค์ ๋ณด์ค ์ ์์ ๊ฒ๋๋ค. ๋ฅด๋ค๊ฐ ๊ทธ ๋ถ๋ถ์ ๊ฝค ์์ธํ ๋ค๋ค์ต๋๋ค. ์ง๋ ๋ช ๋ ๊ฐ ์ด๊ฒ์ด ์ด๋ป๊ฒ ์ฑ์ฅํด์๋์ง ์ดํด๋ณด๋ ๊ฒ์ ํฅ๋ฏธ๋กญ๋ค๊ณ ์๊ฐํฉ๋๋ค. AI ํด๋ฌ์คํฐ๋ ๋น๊ต์ ์๋ก์ด ํ์์ผ๋ก, ์ฝ๋ก๋19 ์ดํ์ธ 2022๋ , 23๋ , ์ฆ ChatGPT๊ฐ ๋ฑ์ฅํ ์งํ์ ๋ณธ๊ฒฉ์ ์ผ๋ก ์์๋์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ์ ์ด๊ธฐ ํด๋ฌ์คํฐ๋ ๊ฝค ์์์ต๋๋ค. ์ฌ์ค, ์ด ์๋ฃ๋ฅผ ์ฐพ์๋ณด๋ 23๋ ์๋ ์ ํฌ ์ด๊ธฐ ํด๋ฌ์คํฐ๊ฐ ์ฝ 128๊ฐ์ GPU๋ก ๊ตฌ์ฑ๋์ด ์์์ต๋๋ค. ๊ทธ๊ฒ ์ ๋ถ์์ฃ . ํ์ง๋ง ๋ณด์๋ค์ํผ, 23๋ ์๋ ์ ํฌ๋ ๊ฝค ๋ง์ด ํ์ฅํ๊ธฐ ์์ํ์ต๋๋ค. |
| And as you sort of fast forward, it really started growing. The demand for this has far surpassed sort of what any one of us could imagine it was. We are in the tens of thousands of GPUs stitched together in a single cluster now. And if I project it forward, and this is the thing I really want to set context, there is absolutely no sign of this slowing down. In fact, it's almost exponential. I only see it accelerating. right? So the demand is exponential. And as Rene was saying, power is constrained. I want to talk a little bit about some of our clusters. That is Prometheus. Prometheus is one of our bigger clusters. It will surpass well over 1 gigawatt by the end of this year. | ์๊ฐ์ด ํ๋ฅด๋ฉด์, ์ ๋ง๋ก ์ฑ์ฅํ๊ธฐ ์์ํ์ต๋๋ค. ์ด๊ฒ์ ๋ํ ์์๋ ์ฐ๋ฆฌ ์ค ๊ทธ ๋๊ตฌ๋ ์์ํ๋ ์์ค์ ํจ์ฌ ๋ฐ์ด๋์์ต๋๋ค. ํ์ฌ ์ ํฌ๋ ๋จ์ผ ํด๋ฌ์คํฐ ๋ด์ ์๋ง ๋์ GPU๋ฅผ ํ๋๋ก ์ฎ์ด ์ฌ์ฉํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์์ผ๋ก ์ ๋งํด๋ณด๋ฉด, ์ ๊ฐ ํนํ ๋ง์๋๋ฆฌ๊ณ ์ถ์ ์ ์, ์ด๋ฌํ ์ถ์ธ๊ฐ ์ ํ ๋ํ๋ ์กฐ์ง์ด ๋ณด์ด์ง ์๋๋ค๋ ๊ฒ์
๋๋ค. ์ฌ์ค์ ๊ฑฐ์ ๊ธฐํ๊ธ์์ ์
๋๋ค. ์ ๋ ์คํ๋ ค ๊ฐ์ํ๋ ๊ฒ์ผ๋ก๋ง ๋ณด์
๋๋ค. ๊ทธ๋ ์ฃ ? ๋ฐ๋ผ์ ์์๋ ๊ธฐํ๊ธ์์ ์ผ๋ก ์ฆ๊ฐํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ฅด๋ค๊ฐ ๋งํ๋ฏ์ด, ์ ๋ ฅ ๊ณต๊ธ์ ์ ํ์ ์
๋๋ค. ์ ํฌ ํด๋ฌ์คํฐ ์ค ์ผ๋ถ์ ๋ํด ์ ์ ๋ง์๋๋ฆฌ๊ณ ์ ํฉ๋๋ค. ์ ๊ฒ์ด ํ๋ก๋ฉํ ์ฐ์ค์ ๋๋ค. ํ๋ก๋ฉํ ์ฐ์ค๋ ์ ํฌ์ ๋๊ท๋ชจ ํด๋ฌ์คํฐ ์ค ํ๋์ ๋๋ค. ์ฌํด ๋ง๊น์ง 1๊ธฐ๊ฐ์ํธ๋ฅผ ํจ์ฌ ์ํํ ๊ฒ์ ๋๋ค. |
| There's a lot of GPUs, I can tell you. And we stitched together a bunch of data centers, a bunch of tent. That thing you see, the blue colored thing is actually a tent. It's a fancy tent, but still a tent, right? It's weatherproof. It can survive about a category 2 hurricane. But we're putting together all of this, stitching it together with a network. And so to our developers, to our researchers, what they end up getting is about 1 gigawatt worth of an AI cluster in a single combined entity, which is pretty powerful, as you can imagine. But like I was saying, the demand is exponential to put it mildly. That is Hyperion. It is going to go up to 5 gigawatts in a few years. | ๋ถ๋ช
ํ ๋ง์๋๋ฆฌ์ง๋ง, GPU๊ฐ ์์ฒญ ๋ง์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ ์ฌ๋ฌ ๋ฐ์ดํฐ ์ผํฐ์ ์ฌ๋ฌ ํ
ํธ ์์ค์ ์ด์ด ๋ถ์์ต๋๋ค. ์ฌ๋ฌ๋ถ์ด ๋ณด์๋ ์ ํ๋์ ๋ฌผ์ฒด๋ ์ฌ์ค ํ
ํธ์
๋๋ค. ๊ณ ๊ธ ํ
ํธ์ด๊ธด ํ์ง๋ง, ์ฌ์ ํ ํ
ํธ์
๋๋ค, ๊ทธ๋ ์ฃ ? ๋ฐฉ์ ๋ฐฉํ์ด ๋๋ฉฐ, ๋๋ต ์นดํ
๊ณ ๋ฆฌ 2๊ธ ํ๋ฆฌ์ผ์ธ์๋ ๊ฒฌ๋ ์ ์์ต๋๋ค. ํ์ง๋ง ์ ํฌ๋ ์ด ๋ชจ๋ ๊ฒ์ ๋คํธ์ํฌ๋ก ์ฐ๊ฒฐํ์ฌ ํ๋๋ก ์ด์ด ๋ถ์ด๊ณ ์์ต๋๋ค. ๊ทธ๋์ ์ ํฌ ๊ฐ๋ฐ์๋ค๊ณผ ์ฐ๊ตฌ์๋ค์ ๊ฒฐ๊ตญ ๋จ์ผ ํตํฉ๋ ํํ๋ก ์ฝ 1๊ธฐ๊ฐ์ํธ ๊ท๋ชจ์ AI ํด๋ฌ์คํฐ๋ฅผ ์ป๊ฒ ๋ฉ๋๋ค. ์์ํ์๊ฒ ์ง๋ง, ์ด๋ ์๋นํ ๊ฐ๋ ฅํฉ๋๋ค. ํ์ง๋ง ์์ ๋ง์๋๋ ธ๋ฏ์ด, ์์๋ ๊ธฐํ๊ธ์์ ์ด๋ผ๊ณ ํด๋ ๊ณผ์ธ์ด ์๋ ์ ๋์ ๋๋ค. ๊ทธ๊ฒ์ด ๋ฐ๋ก ํ์ดํ๋ฆฌ์จ์ ๋๋ค. ๋ช ๋ ์์ 5๊ธฐ๊ฐ์ํธ๊น์ง ๋์ด๋ ๊ฒ์ ๋๋ค. |
| Most people can't fathom what a gigawatt is. A gigawatt is about 10 Palo Altos, the town of Palo Alto, 10x what it consumes is 1 gigawatt. This will be 5. That's 50 Palo Altos, right? That's what we are building out. So it's going to go and go really, really big. So why do we do this? At Meta, we have this vision of delivering personal super intelligence for every single one of our users. This means creating models that can go and figure out the most relevant experience, the most engaging experience for every one of you on our platforms. It means creating a personal assistant for every one of you, right? | ๋๋ถ๋ถ์ ์ฌ๋๋ค์ ๊ธฐ๊ฐ์ํธ๊ฐ ์ด๋ ์ ๋ ๊ท๋ชจ์ธ์ง ์ ๋ชจ๋ฅด์ค ๊ฒ๋๋ค. 1 ๊ธฐ๊ฐ์ํธ๋ ํ๋ก์ํ ์ 10๊ฐ ์ ๋์ ๊ท๋ชจ์
๋๋ค. ํ๋ก์ํ ์๊ฐ ์๋นํ๋ ์ ๋ ฅ๋์ 10๋ฐฐ๊ฐ ๋ฐ๋ก 1 ๊ธฐ๊ฐ์ํธ์ธ ์
์ด์ฃ . ์ ํฌ๊ฐ ๊ตฌ์ถํ๋ ๊ท๋ชจ๋ 5 ๊ธฐ๊ฐ์ํธ์
๋๋ค. ์ด๋ ํ๋ก์ํ 50๊ฐ ๊ท๋ชจ์ ํด๋นํ์ฃ ? ์ ํฌ๊ฐ ์ง๊ธ ๊ตฌ์ถํ๊ณ ์๋ ๊ฒ์ด ๋ฐ๋ก ์ด ์ ๋ ๊ท๋ชจ์
๋๋ค. ๋ฐ๋ผ์ ์ ํฌ์ ๊ตฌ์ถ ๊ท๋ชจ๋ ์ ๋ง ์์ฒญ๋๊ฒ ์ปค์ง ๊ฒ์
๋๋ค. ๊ทธ๋ ๋ค๋ฉด ์ ํฌ๋ ์ ์ด๋ฐ ์ผ์ ํ ๊น์? ๋ฉํ๋ ๋ชจ๋ ์ฌ์ฉ์ ํ ๋ถ ํ ๋ถ๊ป ๊ฐ์ธ ๋ง์ถคํ ์ด์ง๋ฅ์ ์ ๊ณตํ๊ฒ ๋ค๋ ๋น์ ์ ๊ฐ์ง๊ณ ์์ต๋๋ค. ์ด๋ ์ ํฌ ํ๋ซํผ์์ ์ฌ๋ฌ๋ถ ๊ฐ์์๊ฒ ๊ฐ์ฅ ๊ด๋ จ์ฑ ๋๊ณ , ๊ฐ์ฅ ๋ชฐ์ ๋ ๋์ ๊ฒฝํ์ ์ฐพ์๋ด๊ณ ํ์ ํ ์ ์๋ ๋ชจ๋ธ์ ๋ง๋๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค. ์ด๋ ์ฌ๋ฌ๋ถ ๊ฐ์๋ฅผ ์ํ ๊ฐ์ธ ๋น์๋ฅผ ๋ง๋๋ ๊ฒ์ ์๋ฏธํ๊ธฐ๋ ํฉ๋๋ค, ๊ทธ๋ ์ฃ ? |
| Now if you have to go and deliver sort of personal super intelligence to billions of people, what kind of systems would that take? We're talking about billions of people each using sort of exact amount of compute over and over. And like I said, over 3 billion users a day, right? This -- if this advances, there you go. So what does it take? Well, it takes power, it takes land. It takes a decent amount of hardware, software, obviously. And most of all, it takes silicon, a lot of silicon, right? And this is why I think Arm is such a natural partner for us. What we want is a partner who can match our ambition who can match our cadence of velocity of innovation. | ๋ง์ฝ ์์ญ์ต ๋ช
์ ์ฌ๋๋ค์๊ฒ ๊ฐ์ธ ์ด์ง๋ฅ์ ์ ๊ณตํด์ผ ํ๋ค๋ฉด, ์ด๋ค ์ข
๋ฅ์ ์์คํ
์ด ํ์ํ ๊น์? ์ฐ๋ฆฌ๋ ์์ญ์ต ๋ช
์ ์ฌ๋๋ค์ด ๊ฐ์ ์๋นํ ์์ ์ปดํจํ
์์์ ๋ฐ๋ณต์ ์ผ๋ก ์ฌ์ฉํ๋ ์ํฉ์ ์ด์ผ๊ธฐํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ง์๋๋ ธ๋ฏ์ด, ํ๋ฃจ์ 30์ต ๋ช
์ด์์ ์ฌ์ฉ์๊ฐ ์์ต๋๋ค, ๊ทธ๋ ์ฃ ? ์ด๊ฒ์ด ๋ฐ์ ํ๋ค๋ฉด, ๊ทธ์ ๋ฐ๋ฅธ ์๊ตฌ์ฌํญ์ ๋ช
ํํด์ง๋๋ค. ๊ทธ๋ ๋ค๋ฉด ๋ฌด์์ด ํ์ํ ๊น์? ์, ์ ๋ ฅ์ด ํ์ํ๊ณ , ํ ์ง๊ฐ ํ์ํฉ๋๋ค. ๋น์ฐํ ์๋นํ ์์ ํ๋์จ์ด์ ์ํํธ์จ์ด๋ ํ์ํ์ฃ . ๊ทธ๋ฆฌ๊ณ ๋ฌด์๋ณด๋ค๋, ์ค๋ฆฌ์ฝ์ด ํ์ํฉ๋๋ค. ์์ฃผ ๋ง์ ์ค๋ฆฌ์ฝ์ด์, ๊ทธ๋ ์ฃ ? ์ด๊ฒ์ด ๋ฐ๋ก Arm์ด ์ฐ๋ฆฌ์๊ฒ ๋งค์ฐ ์ ํฉํ ํํธ๋๋ผ๊ณ ์๊ฐํ๋ ์ด์ ์ ๋๋ค. ์ฐ๋ฆฌ๊ฐ ์ํ๋ ๊ฒ์ ์ฐ๋ฆฌ์ ์ผ๋ง์ ๋ถํฉํ๊ณ , ์ฐ๋ฆฌ์ ํ์ ์๋(์ผ์ด๋์ค)์ ๋ฐ๋ง์ถ ์ ์๋ ํํธ๋์ ๋๋ค. |
| And what we realized when we're sitting down with Arm is that they codevelop it. They were as hungry as we were and most importantly for us, we were as power conscious and as efficient as we wanted sort of them to be. This is why while Arm is now the primary co-collaborator and the primary sort of partner, the CPU that we are ending up developing is pretty foundational. It can be -- it's not just a meta CPU. It's not just an Arm CPU. This is something that I think will end up being a foundational CPU for the whole ecosystem. I think we are at the threshold of something pretty sweet here because you're going to hear more and more about sort of the constraints that data centers are facing. | Arm๊ณผ ๋ ผ์ํ๋ฉด์ ์ฐ๋ฆฌ๊ฐ ๊นจ๋ฌ์ ๊ฒ์ ๊ทธ๋ค์ด ๊ณต๋ ๊ฐ๋ฐ์ ์ฐธ์ฌํ๋ค๋ ์ ์ด์์ต๋๋ค. ๊ทธ๋ค๋ ์ฐ๋ฆฌ๋งํผ ์ด์ ์ ์ด์๊ณ , ์ฐ๋ฆฌ์๊ฒ ๊ฐ์ฅ ์ค์ํ๋ ๊ฒ์ ์ฐ๋ฆฌ๊ฐ ๊ทธ๋ค์๊ฒ ๊ธฐ๋ํ๋ ๋งํผ ์ ๋ ฅ ํจ์จ์ฑ๊ณผ ํจ์จ์ฑ์ ์ค์ํ๊ฒ ์๊ฐํ๋ค๋ ์ ์ ๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก Arm์ด ์ด์ ์ฃผ์ ๊ณต๋ ํ๋ ฅ์์ด์ ์ฃผ์ ํํธ๋์ด์ง๋ง, ์ฐ๋ฆฌ๊ฐ ๊ฒฐ๊ตญ ๊ฐ๋ฐํ๊ฒ ๋ CPU๊ฐ ์๋นํ ๊ธฐ์ด์ ์ธ ์ด์ ์ ๋๋ค. ์ด๊ฒ์ ๋จ์ํ ๋ฉํ(Meta) CPU๊ฐ ์๋๋๋ค. ๋จ์ํ Arm CPU๋ ์๋๋๋ค. ์ ๊ฐ ์๊ฐํ๊ธฐ์ ์ด๊ฒ์ ๊ฒฐ๊ตญ ์ ์ฒด ์ํ๊ณ๋ฅผ ์ํ ๊ธฐ์ด์ ์ธ CPU๊ฐ ๋ ๊ฒ์ ๋๋ค. ์ ๋ ์ฐ๋ฆฌ๊ฐ ์ฌ๊ธฐ์ ์๋นํ ํ๊ธฐ์ ์ธ ์ผ์ ๋ฌธํฑ์ ์ ์๋ค๊ณ ์๊ฐํฉ๋๋ค. ์๋ํ๋ฉด ์ฌ๋ฌ๋ถ์ ๋ฐ์ดํฐ ์ผํฐ๋ค์ด ์ง๋ฉดํ๊ณ ์๋ ์ ์ฝ ์ฌํญ๋ค์ ๋ํด ์ ์ ๋ ๋ง์ด ๋ฃ๊ฒ ๋ ๊ฒ์ด๊ธฐ ๋๋ฌธ์ ๋๋ค. |
| You're going to hear more and more about while the demand for compute is growing, the power is not growing at the exact same curve. So this marriage is -- I think about it personally as a win-win situation, right? So it's extremely sort of heartening to see Arm moving on from not just being an IP license provider, but actually getting into the game of sort of building something that is production scale and production ready. Exciting times. 2 years, 3 years in the making, but I think about this as the sweetest of things take some time, but we're getting there. Now like I said, we are obsessed with efficiency. | ์ปดํจํ ์์๋ ๊ณ์ ์ฆ๊ฐํ๋๋ฐ, ์ ๋ ฅ์ ๊ทธ์ ์ ํํ ๊ฐ์ ๊ณก์ ์ผ๋ก ์ฆ๊ฐํ์ง ๋ชปํ๊ณ ์๋ค๋ ์ด์ผ๊ธฐ๊ฐ ์ ์ ๋ ๋ง์ด ๋ค๋ฆด ๊ฒ์ ๋๋ค. ๊ทธ๋์ ์ด๋ฌํ ๊ฒฐํฉ์ โ ์ ๋ ๊ฐ์ธ์ ์ผ๋ก ์์(win-win) ์ํฉ์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค, ๋ง์ฃ ? Arm์ด ๋จ์ํ IP ๋ผ์ด์ ์ค ์ ๊ณต์ ์ฒด์ ๋จธ๋ฌด๋ฅด์ง ์๊ณ , ์ค์ ๋ก ์์ฐ ๊ท๋ชจ(production scale)์ ๋ง๊ณ ์ฆ์ ์ ์ฉ ๊ฐ๋ฅํ(production ready) ๋ฌด์ธ๊ฐ๋ฅผ ๊ตฌ์ถํ๋ ์ผ์ ๋ฐ์ด๋ค๊ณ ์๋ค๋ ์ ์ ๋ณด๊ฒ ๋์ด ๋งค์ฐ ๊ณ ๋ฌด์ ์ ๋๋ค. ํฅ๋ฏธ๋ก์ด ์๊ธฐ์ ๋๋ค. 2๋ , 3๋ ์ด๋ผ๋ ์๊ฐ์ด ๊ฑธ๋ ธ์ง๋ง, ์ ๋ ์ข์ ์ผ์๋ ์๊ฐ์ด ๊ฑธ๋ฆฌ๋ ๋ฒ์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค. ํ์ง๋ง ์ฐ๋ฆฌ๋ ๋ชฉํ๋ฅผ ํฅํด ๋์๊ฐ๊ณ ์์ต๋๋ค. ๋ค์ ๋ง์๋๋ฆฌ์ง๋ง, ์ ํฌ๋ ํจ์จ์ฑ์ ์ง์ฐฉํ๊ณ ์์ต๋๋ค. |
| And if you think about one of the biggest appeal that Arm has had over the years, it's power profile. Arm can go -- Rene had this fascinating experience that he was talking about taking 30 million cores instead of 30 million, now making it 120 million and fitting in the same power envelope. But that's one thing. You don't want to compromise on performance, right? This is the thing that I really want to make sure we drive here. The biggest reason why we sat down with Arm and had this conversation was we want to put in a lot more cores per watt, but we do not want to compromise on the performance piece. That marriage is why I really think it's a win-win situation here. | Arm์ด ์๋ ๊ฐ ๊ฐ์ ธ์จ ๊ฐ์ฅ ํฐ ๋งค๋ ฅ ์ค ํ๋๋ ๋ฐ๋ก ์ ๋ ฅ ํจ์จ์ฑ์ ๋๋ค. Arm์ ๊ฐ๋ฅํฉ๋๋ค. ๋ฅด๋ค๊ฐ ์ด์ผ๊ธฐํ๋ ํฅ๋ฏธ๋ก์ด ๊ฒฝํ์ด ์๋๋ฐ, 3์ฒ๋ง ๊ฐ์ ์ฝ์ด๋ฅผ 1์ต 2์ฒ๋ง ๊ฐ๋ก ๋๋ฆฌ๋ฉด์๋ ๋์ผํ ์ ๋ ฅ ๋ฒ์ ๋ด์ ๋ง์ถ๋ ๊ฒ์ด์ฃ . ํ์ง๋ง ๊ทธ๊ฑด ํ ๊ฐ์ง ์ธก๋ฉด์ผ ๋ฟ์ ๋๋ค. ์ฑ๋ฅ์ ํํํ๊ณ ์ถ์ง๋ ์์ ๊ฒ๋๋ค, ๊ทธ๋ ์ฃ ? ์ด๊ฒ์ด ๋ฐ๋ก ์ ๊ฐ ์ฌ๊ธฐ์ ํ์คํ ๊ฐ์กฐํ๊ณ ์ถ์ ๋ถ๋ถ์ ๋๋ค. ์ฐ๋ฆฌ๊ฐ Arm๊ณผ ๋ง๋ ์ด ๋ํ๋ฅผ ๋๋ ๊ฐ์ฅ ํฐ ์ด์ ๋ ์ํธ๋น ํจ์ฌ ๋ ๋ง์ ์ฝ์ด๋ฅผ ๋ฃ๊ณ ์ถ์์ง๋ง, ์ฑ๋ฅ ๋ถ๋ถ์์๋ ํํํ๊ณ ์ถ์ง ์์๊ธฐ ๋๋ฌธ์ ๋๋ค. ๊ทธ๋ฌํ ๊ฒฐํฉ์ด ๋ฐ๋ก ์ ๊ฐ ์ด๊ฒ์ ์-์ ์ํฉ์ด๋ผ๊ณ ์๊ฐํ๋ ์ด์ ์ ๋๋ค. |
| In fact, about 2, 2.5 years ago, we sat down with Arm, we actually first surveyed the market to see was there a CPU that could meet the specs that we wanted. If we met the performance, we couldn't get the power. If we got the power, we couldn't get the performance. And this is why Arm ended up being such a partner. The ability to scale that Arm gives us when you push in a lot more cores. And if you think about personal super intelligence, if you think about the orchestration that Rene showed, you don't want to starve your CPUs nor do you want to starve your GPUs. That marriage that you end up doing is, I think, that most people are going to realize pretty soon. | ์ฌ์ค ์ฝ 2๋ ์์ 2๋ ๋ฐ ์ , ์ ํฌ๋ Arm๊ณผ ๋ ผ์๋ฅผ ์์ํ์ต๋๋ค. ๋จผ์ ์์ฅ ์กฐ์ฌ๋ฅผ ํตํด ์ ํฌ๊ฐ ์ํ๋ ์ฌ์์ ์ถฉ์กฑํ ์ ์๋ CPU๊ฐ ์๋์ง ํ์ธํ์ฃ . ์ฑ๋ฅ์ ์ถฉ์กฑ์ํค๋ฉด ์ ๋ ฅ์ ๋ง์ถ ์ ์์๊ณ , ์ ๋ ฅ์ ๋ง์ถ๋ฉด ์ฑ๋ฅ์ ์ถฉ์กฑ์ํฌ ์ ์์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก Arm์ด ์ด๋ ๊ฒ ์ค์ํ ํํธ๋๊ฐ ๋ ์ด์ ์ ๋๋ค. Arm์ด ์ ๊ณตํ๋ ํ์ฅ์ฑ์ ๋ ๋ง์ ์ฝ์ด๋ฅผ ํ์ฌํ ๋ ๋งค์ฐ ์ค์ํฉ๋๋ค. ๊ฐ์ธ ์ํผ ์ธํ ๋ฆฌ์ ์ค๋ฅผ ์๊ฐํด๋ณด๊ฑฐ๋ Rene๊ฐ ๋ณด์ฌ์ค ์ค์ผ์คํธ๋ ์ด์ ์ ์๊ฐํด๋ณด๋ฉด, CPU์ GPU ๋ชจ๋ ์์ ๋ถ์กฑ์ผ๋ก ํ๋์ด๊ฒ ํ๊ณ ์ถ์ง ์์ ๊ฒ์ ๋๋ค. ๊ฒฐ๊ตญ ์ด๋ฃจ๊ฒ ๋๋ ๊ทธ ๊ฒฐํฉ์ ๋๋ถ๋ถ์ ์ฌ๋๋ค์ด ๊ณง ๊นจ๋ซ๊ฒ ๋ ๊ฒ์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค. |
| Now the design point that we chose for this was something to minimize risk for this iteration. We wanted to make sure we get our first CPU right, get it working out of the box. But this is a multigenerational partnership. I just want to emphasize this. When we look at subsequent iterations of things that are already in the hopper of what we're going to build out, I truly believe that this chip is going to expand sort of the performance on multiple axis. In fact, this ecosystem is actually going to be awesome. When you challenge the incumbents, you see innovation across the board. That, I think, is what all of us will end up achieving. Now I want to talk about why. | ์ด๋ฒ ๊ฐ๋ฐ ๋จ๊ณ์์ ์ ํฌ๊ฐ ์ ํํ ์ค๊ณ ์ฃผ์์ ์ ์ํ์ ์ต์ํํ๋ ๊ฒ์ด์์ต๋๋ค. ์ฒซ CPU๋ฅผ ์ฑ๊ณต์ ์ผ๋ก ๊ฐ๋ฐํ๊ณ , ์ฆ์ ๋ฌธ์ ์์ด ์๋ํ๋๋ก ํ๋ ๊ฒ์ด ๋ชฉํ์์ต๋๋ค. ํ์ง๋ง ์ด๊ฒ์ ๋ค์ธ๋์ ํ๋ ฅ ๊ด๊ณ์ ๋๋ค. ์ด ์ ์ ๊ฐ์กฐํ๊ณ ์ถ์ต๋๋ค. ์ด๋ฏธ ๊ณํ๋์ด ๊ฐ๋ฐ๋ ํ์ ์ดํฐ๋ ์ด์ ๋ค์ ์ดํด๋ณด๋ฉด, ์ ๋ ์ด ์นฉ์ด ๋ค๋ฐฉ๋ฉด์ผ๋ก ์ฑ๋ฅ์ ํ์ฅํ ๊ฒ์ด๋ผ๊ณ ํ์ ํฉ๋๋ค. ์ฌ์ค, ์ด ์ํ๊ณ๋ ์ ๋ง ๋๋จํ ๊ฒ์ ๋๋ค. ๊ธฐ์กด ์์ฅ ์ง๋ฐฐ์๋ค์๊ฒ ๋์ ํ๋ฉด, ์ ๋ฐ์ ์ธ ํ์ ์ ๋ณผ ์ ์์ต๋๋ค. ์ ๋ ๊ทธ๊ฒ์ด ์ฐ๋ฆฌ ๋ชจ๋๊ฐ ๊ฒฐ๊ตญ ๋ฌ์ฑํ๊ฒ ๋ ๊ฒ์ด๋ผ๊ณ ์๊ฐํฉ๋๋ค. ์ด์ ๊ทธ ์ด์ ์ ๋ํด ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. |
| I want to take this back, I guess, to why we do this work. But like I say, 3 billion, 3.5 billion people use our products every single day. This means there's your friends messaging each other on WhatsApp. It could be a small or medium business messaging the users on a platform. It could be somebody going and doing an AI interaction with Meta AI. None of this is possible without infrastructure. Infrastructure has now become -- has gone from being on the backside of sort of technology innovation to being the enabler of technology innovation, right? AI is built on the backbone of infrastructure. | ๋ค์ ํ๋ฒ ์ฐ๋ฆฌ๊ฐ ์ด ์ผ์ ํ๋ ์ด์ ์ ๋ํด ๋ง์๋๋ฆฌ๊ณ ์ ํฉ๋๋ค. ํ์ง๋ง ์์ ๋ง์๋๋ ธ๋ฏ์ด, ๋งค์ผ 30์ต์์ 35์ต ๋ช
์ ์ฌ๋๋ค์ด ์ ํฌ ์ ํ์ ์ฌ์ฉํ๊ณ ์์ต๋๋ค. ์ด๋ ์น๊ตฌ๋ค์ด ์์ธ ์ฑ(WhatsApp)์ผ๋ก ์๋ก ๋ฉ์์ง๋ฅผ ์ฃผ๊ณ ๋ฐ๋ ๊ฒฝ์ฐ์ผ ์๋ ์๊ณ , ์ค์๊ธฐ์ ์ด ํ๋ซํผ์ ํตํด ์ฌ์ฉ์๋ค์๊ฒ ๋ฉ์์ง๋ฅผ ๋ณด๋ด๋ ๊ฒฝ์ฐ์ผ ์๋ ์์ต๋๋ค. ๋๋ ๋๊ตฐ๊ฐ ๋ฉํ AI(Meta AI)์ AI ์ํธ์์ฉ์ ํ๋ ๊ฒฝ์ฐ์ผ ์๋ ์์ต๋๋ค. ์ด ๋ชจ๋ ๊ฒ์ ์ธํ๋ผ ์์ด๋ ๋ถ๊ฐ๋ฅํฉ๋๋ค. ์ธํ๋ผ๋ ์ด์ ๊ธฐ์ ํ์ ์ '๋ท๋จ'์ ๋จธ๋ฌผ๋ ์ญํ ์์ ๋ฒ์ด๋, ๊ธฐ์ ํ์ ์ ๊ฐ๋ฅํ๊ฒ ํ๋ 'ํต์ฌ ๋๋ ฅ'์ผ๋ก ์๋ฆฌ๋งค๊นํ์ต๋๋ค. ๊ทธ๋ ์ง ์์ต๋๊น? AI๋ ์ธํ๋ผ๋ผ๋ ๊ทผ๊ฐ ์์ ๊ตฌ์ถ๋๋ ๊ฒ์ ๋๋ค. |
| So every interaction, every post, every feed, every call is done on the basis of what we build out on the back end. And at least for us, we're custom building data centers, we're custom building hardware and custom building silicon. That's why Arm, I think, is such a big partner for us because for us, we want to squeeze every bit of performance out of what we build out. We think about optimizing things like performance per watt, performance per gigawatt, and Arm allows us to do that. It allows us to go and increase the efficacy of everything we build out. Why? So that we can go and serve more users so that we can hopefully improve every one of your lives in some way, shape or form. | ๋ฐ๋ผ์ ๋ชจ๋ ์ํธ์์ฉ, ๊ฒ์๋ฌผ, ํผ๋, ํตํ๋ ์ ํฌ๊ฐ ๋ฐฑ์๋์์ ๊ตฌ์ถํ๋ ์์คํ ์ ๊ธฐ๋ฐ์ผ๋ก ์ด๋ฃจ์ด์ง๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ์ด๋ ์ ํฌ๋ ๋ฐ์ดํฐ ์ผํฐ, ํ๋์จ์ด, ์ค๋ฆฌ์ฝ์ ์ง์ ์ค๊ณํ์ฌ ๊ตฌ์ถํ๊ณ ์์ต๋๋ค. ๋ฐ๋ก ๊ทธ ์ด์ ๋ก Arm์ ์ ํฌ์๊ฒ ๋งค์ฐ ์ค์ํ ํํธ๋๋ผ๊ณ ์๊ฐํฉ๋๋ค. ์ ํฌ๋ ์ ํฌ๊ฐ ๊ตฌ์ถํ๋ ์์คํ ์์ ํ ์น์ ์ฑ๋ฅ๋ ๋์น์ง ์๊ณ ์ต๋ํ ๋์ด๋ด๊ณ ์ ํ๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ ํฌ๋ ์ํธ๋น ์ฑ๋ฅ(performance per watt), ๊ธฐ๊ฐ์ํธ๋น ์ฑ๋ฅ(performance per gigawatt)๊ณผ ๊ฐ์ ์งํ๋ค์ ์ต์ ํํ๋ ๋ฐ ์ค์ ์ ๋๋ฉฐ, Arm์ ์ ํฌ๊ฐ ๊ทธ๋ ๊ฒ ํ ์ ์๋๋ก ์ง์ํฉ๋๋ค. ์ด๋ฅผ ํตํด ์ ํฌ๊ฐ ๊ตฌ์ถํ๋ ๋ชจ๋ ๊ฒ์ ํจ์จ์ฑ์ ๋์ผ ์ ์์ต๋๋ค. ๊ทธ ์ด์ ๋ ๋ฌด์์ผ๊น์? ๋ ๋ง์ ์ฌ์ฉ์์๊ฒ ์๋น์ค๋ฅผ ์ ๊ณตํ๊ณ , ๋์๊ฐ ์ฌ๋ฌ๋ถ ๊ฐ์์ ์ถ์ ์ด๋ค ํํ๋ก๋ ๊ฐ์ ํ๋ ๋ฐ ๊ธฐ์ฌํ ์ ์๊ธฐ๋ฅผ ํฌ๋งํ๊ธฐ ๋๋ฌธ์ ๋๋ค. |
| And that's why I think Arm has been an awesome partner. So thank you, Rene and team. It's been absolutely a pleasure to work with you, and hopefully, we'll do this for years together. Thank you. | ๊ทธ๋์ ์ ๋ Arm์ด ์ ๋ง ๋ ๋ ํ ํํธ๋์๋ค๊ณ ์๊ฐํฉ๋๋ค. ๋ฅด๋ค์ ํ์ ์ฌ๋ฌ๋ถ๊ป ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ํจ๊ป ์ผํ๋ ๊ฒ์ด ๋ํ ๋์ ์์ด ์ฆ๊ฑฐ์ ๊ณ , ์์ผ๋ก๋ ์ค๋ซ๋์ ํจ๊ป ํ ์ ์๊ธฐ๋ฅผ ๋ฐ๋๋๋ค. ๊ฐ์ฌํฉ๋๋ค. |
๋ค์์ Arm Holdings plc์ ์ค์ ๋ฐํ ์์ฝ์
๋๋ค.
* **์ฌ์
๋ชจ๋ธ ํ์ฅ:** Arm์ AI ๋ฐ์ดํฐ์ผํฐ์ ์์ด์ ํธ AI ์ํฌ๋ก๋์ ํ์ํ ์ ๋ ฅ ํจ์จ์ ์ธ CPU ์์ ๊ธ์ฆ์ ๋์ํ๊ธฐ ์ํด, IP ๋ผ์ด์ ์ฑ์ ๋์ด ์์ฒด ์ค๋ฆฌ์ฝ์ธ "Arm AGI CPU"๋ฅผ ์ง์ ํ๋งคํ๋ ์ฌ์
๋ชจ๋ธ ํ์ฅ์ ๋ฐํํ์ต๋๋ค. ์ด๋ ์๋ก์ด ๋งค์ถ์ ์ฐฝ์ถ์ ์๋ฏธํฉ๋๋ค.
* **CSS ์ ๋ต ์ฑ๊ณผ:** 3~4๋
์ ๋์
๋ Compute Subsystems (CSS) ์ ๋ต์ ์ด๋ฏธ ๋ก์ดํฐ ๋งค์ถ์ ์ฝ 20%๋ฅผ ์ฐจ์งํ๋ฉฐ ์ฑ์ฅ ์ค์ด๋ฉฐ, ๊ณ ๊ฐ์ฌ์ ์ ํ ์ถ์ ๊ธฐ๊ฐ์ 12~18๊ฐ์ ๋จ์ถ์ํค๋ ๋ฑ ๊ธ์ ์ ์ธ ์ฑ๊ณผ๋ฅผ ๋ณด์ด๊ณ ์์ต๋๋ค.
* **Meta์์ ํต์ฌ ํํธ๋์ญ:** Meta Platforms์์ ํต์ฌ ํํธ๋์ญ์ ํตํด Arm AGI CPU์ ์์ฅ์ฑ์ ์
์ฆํ์ผ๋ฉฐ, Meta๋ Hyperion ํด๋ฌ์คํฐ(ํฅํ 5๊ธฐ๊ฐ์ํธ ๊ท๋ชจ) ๋ฑ์์ Arm์ ๊ณ ์ฑ๋ฅ-์ ์ ๋ ฅ ์๋ฃจ์
์ด ๋ฐ์ดํฐ์ผํฐ์ ์ ๋ ฅ ์ ์ฝ์ ํด๊ฒฐํ๋ ๋ฐ ํ์์ ์์ ๊ฐ์กฐํ์ต๋๋ค.
* **๊ฒฝ์์ง์ ์์ ๊ฐ:** ๊ฒฝ์์ง์ Arm์ ๋
๋ณด์ ์ธ ์ ์ ๋ ฅ DNA์ ๊ด๋ฒ์ํ ์ํ๊ณ๋ฅผ ํ์ฉํ์ฌ ๊ธ์ฑ์ฅํ๋ AI ์์ฅ์ ํต์ฌ ์ธํ๋ผ ๋ณ๋ชฉ ํ์์ ํด๊ฒฐํ ์ ์๋ค๋ ๊ฐ๋ ฅํ ์์ ๊ฐ์ ํ๋ช
ํ๋ฉฐ, AI ์๋์ CPU์ ์ค์์ฑ์ ์ฌ๊ฐ์กฐํ์ต๋๋ค.
| Original | Translation |
|---|---|
| CEO & Director: Amazing. Santosh, thank you. That was terrific. I have someone else I'd like to ask to join us to also talk about how they plan to use our Arm AGI CPU. And that's Kevin Weil from OpenAI. Kevin? | **CEO & Director:** ์ ๋ง ๋๋จํฉ๋๋ค. ์ฐํ ์ค๋, ๊ฐ์ฌํฉ๋๋ค. ์ ๋ง ํ๋ฅญํ์ต๋๋ค. ์ ํฌ Arm AGI CPU๋ฅผ ์ด๋ป๊ฒ ํ์ฉํ ๊ณํ์ด์ ์ง์ ๋ํด ์ด์ผ๊ธฐํด์ฃผ์ค ๋ค๋ฅธ ๋ถ์ ๋ชจ์๊ฒ ์ต๋๋ค. ๋ฐ๋ก OpenAI์ ์ผ๋น ์์ผ(Kevin Weil)๋์ ๋๋ค. ์ผ๋น๋? |
| OpenAI, L.L.C.: Thank you, sir. | **OpenAI, L.L.C.:** ๋ค, ๊ฐ์ฌํฉ๋๋ค. |
| CEO & Director: Kevin, thanks for joining us. | **CEO & Director:** ์ผ๋น๋, ์ฐธ์ฌํด์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. |
| OpenAI, L.L.C.: Thank you for having me. | **OpenAI, L.L.C.:** ์ด๋ํด ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. |
| CEO & Director: Welcome to Fort Mason. Have you been here before? | **CEO & Director:** ํฌํธ ๋ฉ์ด์จ์ ์ค์ ๊ฒ์ ํ์ํฉ๋๋ค. ์ฌ๊ธฐ๋ ์ด์ ์ ์๋ณด์ ์ ์์ผ์ ๊ฐ์? |
| OpenAI, L.L.C.: I have. There are a few conferences in the past. | **OpenAI, L.L.C.:** ๋ค, ์์ต๋๋ค. ๊ณผ๊ฑฐ์ ๋ช ์ฐจ๋ก ์ปจํผ๋ฐ์ค(conference)์ ์ฐธ์ํ ๊ฒฝํ์ด ์์ต๋๋ค. |
| CEO & Director: Well, welcome. So first off, just tell us and tell me, why does -- why does this launch today matter to OpenAI? | **CEO & Director:** ๋ค, ๋ฐ๊ฐ์ต๋๋ค. ๋จผ์ ์ ํฌ์๊ฒ ์ค๋ช ํด ์ฃผ์ญ์์ค. ์ค๋ ์ด ์ถ์๊ฐ OpenAI์ ์ ์ค์ํ๊ฐ์? |
| OpenAI, L.L.C.: Well, I thought you did a good job painting this. AI performance these days is system performance. And GPUs kind of get top billing wherever they go, but really the CPU is playing an incredibly important role as an orchestrator. But also, I think as AI becomes more agentic, when you look at a rollout that an agent is doing, it's using it's using tools inside containers, that's CPUs, it's running Python scripts as it does -- as it performs skills. Those are CPUs. So the CPU plays an incredibly important role. And it's really the whole system together that makes it all possible. | **OpenAI, L.L.C.:** ์, ์ด ์ํฉ์ ์ ์ง์ด์ฃผ์
จ๋ค๊ณ ์๊ฐํฉ๋๋ค. ์์ฆ AI ์ฑ๋ฅ(AI performance)์ ๊ณง ์์คํ
์ฑ๋ฅ(system performance)์
๋๋ค. GPU(๊ทธ๋ํฝ ์ฒ๋ฆฌ ์ฅ์น)๊ฐ ์ด๋์๋ ๊ฐ์ฅ ์ฃผ๋ชฉ์ ๋ฐ์ง๋ง, ์ฌ์ค CPU(์ค์ ์ฒ๋ฆฌ ์ฅ์น)๋ ์ค์ผ์คํธ๋ ์ดํฐ(orchestrator)๋ก์ ์์ฒญ๋๊ฒ ์ค์ํ ์ญํ ์ ํ๊ณ ์์ต๋๋ค. ๋ํ, AI๊ฐ ์ ์ ๋ ์์ด์ ํธํ(agentic)๋ ์๋ก, ์์ด์ ํธ๊ฐ ์ํํ๋ ๋ฐฐํฌ(rollout)๋ฅผ ๋ณด๋ฉด ์ปจํ ์ด๋(container) ๋ด๋ถ์ ๋๊ตฌ๋ค์ ์ฌ์ฉํ๋๋ฐ, ์ด๊ฒ์ด ๋ฐ๋ก CPU์ ๋๋ค. ๊ธฐ๋ฅ์ ์ํํ ๋ ํ์ด์ฌ ์คํฌ๋ฆฝํธ(Python script)๋ฅผ ์คํํ๋ ๊ฒ๋ CPU์ ์ญํ ์ ๋๋ค. ๋ฐ๋ผ์ CPU๋ ์ ๋ง ์์ฒญ๋๊ฒ ์ค์ํ ์ญํ ์ ํฉ๋๋ค. ๊ฒฐ๊ตญ ์ด ๋ชจ๋ ๊ฒ์ ๊ฐ๋ฅํ๊ฒ ํ๋ ๊ฒ์ ์ ์ฒด ์์คํ (system)์ด ํจ๊ป ์๋ํ๋ ๊ฒ์ ๋๋ค. |
| CEO & Director: Now your role at OpenAI is a pretty cool one, right? You're doing math and science and the stuff that's super compute heavy. And when you think about compute constraints, and I know when I talk to you or Sam or Mark or anyone at your company, it's I need more compute. Yes. Tell us about that. | **CEO & Director:** OpenAI์์ ๋งก๊ณ ๊ณ์ ์ญํ ์ ๊ต์ฅํ ํฅ๋ฏธ๋ก์ด ์ญํ ์ด์์ฃ . ์ํ, ๊ณผํ ๋ถ์ผ์์ ๋ง๋ํ ์ฐ์ฐ๋(compute)์ ์๊ตฌํ๋ ์์ ๋ค์ ์ฃผ๋ก ํ์์์์. ๊ทธ๋ฆฌ๊ณ ์ปดํจํ ์์ ์ ์ฝ(compute constraint)์ ๋ํด ๋ง์๋๋ฆฌ์๋ฉด, ์ ๊ฐ ๋ํ๋์ด๋ ์, ๋งํฌ ๋ฑ ํ์ฌ ๊ด๊ณ์๋ถ๋ค๊ณผ ๋ํํ ๋๋ง๋ค ๋ '์ปดํจํ ์์์ด ๋ ํ์ํ๋ค'๋ ๋ง์์ ๋ฃ๊ณค ํฉ๋๋ค. ์ด์ ๋ํด ์์ธํ ๋ง์ํด์ฃผ์๊ฒ ์ด์? |
| CEO & Director: Tell us about that. | **CEO & Director:** ๊ทธ๊ฒ์ ๋ํด ๋ง์ํด ์ฃผ์๊ฒ ์ด์? |
| OpenAI, L.L.C.: That is one of the most common things I hear inside OpenAI. I need more compute. It's kind of the coin of the realm. I mean the root of it is we have more demand from customers. We have more ideas internally that we want to experiment with. We have more things that we want to do than frankly, the industry can keep up with. And when you get to the bottom of all this, it's certainly it's about silicon, but it's also about power. And so if you have a CPU that can draw less power, it could be just as performant, but use less power, it means you have more leftover for everything else that you want to do. That means more inference and more compute. That means more intelligence. And if there's one thing that I've learned in my couple of years now at OpenAI, it's that more intelligence leads us to be able to build better products for all of you. The thing that I keep coming back to that I try and remind myself of all times is as amazing as the models are today, -- and every year, I'm blown away by the amount of progress we make. As amazing as the models are, the model that you use today is the worst AI model that you will ever use for the rest of your life. It's the worst AI model you're going to use for the rest of your life. And a year from now, you're going to be -- you couldn't imagine coming back to the AI models of today because they're getting better at such a rapid pace, which just means there's basically infinite demand for intelligence. So we are not stopping from here. | **OpenAI, L.L.C.:** OpenAI ๋ด๋ถ์์ ์ ๊ฐ ๊ฐ์ฅ ํํ๊ฒ ๋ฃ๋ ๋ง ์ค ํ๋๋ '์ปดํจํ
์์(compute)์ด ๋ ํ์ํ๋ค'๋ ๊ฒ์
๋๋ค. ์ด๋ ๊ฐ์ฅ ์ค์ํ ์์์ด๋ผ๊ณ ํ ์ ์์ต๋๋ค. ๊ทธ ๊ทผ๋ณธ์ ์ธ ๋ฐฐ๊ฒฝ์๋ ๊ณ ๊ฐ ์์๊ฐ ์ฆ๊ฐํ๊ณ ์๊ณ , ๋ด๋ถ์ ์ผ๋ก ์คํํ๊ณ ์ถ์ ์์ด๋์ด๊ฐ ๋ ๋ง์ผ๋ฉฐ, ์์งํ ๋งํด ์ ๊ณ๊ฐ ๋ฐ๋ผ์ก๊ธฐ ํ๋ค ์ ๋๋ก ์ ํฌ๊ฐ ํ๊ณ ์ถ์ ์ผ์ด ๋ง๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ด ๋ชจ๋ ๊ฒ์ ๋ณธ์ง์ ํ๊ณ ๋ค๋ฉด, ๋ฌผ๋ก ๋ฐ๋์ฒด(silicon) ๋ฌธ์ ์ด๊ธฐ๋ ํ์ง๋ง, ๋์์ ์ ๋ ฅ(power) ๋ฌธ์ ์ด๊ธฐ๋ ํฉ๋๋ค. ๋ฐ๋ผ์ ์ ๋ ฅ ์๋ชจ๊ฐ ์ ์ CPU๊ฐ ์๋ค๋ฉด, ์ฑ๋ฅ์ ๋์ผํ๋ฉด์๋ ์ ๋ ฅ์ ๋ ์ฌ์ฉํ๊ฒ ๋๋ฏ๋ก, ๋ค๋ฅธ ๋ชจ๋ ์์ ์ ํ์ฉํ ์ ์๋ ์ฌ์ ์์์ด ๋ ๋ง์์ง๋ค๋ ์๋ฏธ์ ๋๋ค. ์ด๋ ๋ ๋ง์ ์ถ๋ก (inference)๊ณผ ๋ ๋ง์ ์ปดํจํ ์์์ ๊ฐ๋ฅํ๊ฒ ํ๋ฉฐ, ๊ถ๊ทน์ ์ผ๋ก๋ ๋ ๋์ ์ง๋ฅ์ผ๋ก ์ด์ด์ง๋๋ค. OpenAI์์ ์ง๋ ๋ช ๋ ๊ฐ ์ ๊ฐ ๋ฐฐ์ด ํ ๊ฐ์ง๊ฐ ์๋ค๋ฉด, ๋ ๋ง์ ์ง๋ฅ(intelligence)์ด ์ฌ๋ฌ๋ถ ๋ชจ๋๋ฅผ ์ํ ๋ ๋์ ์ ํ์ ๋ง๋ค ์ ์๊ฒ ํด์ค๋ค๋ ๊ฒ์ ๋๋ค. ์ ๊ฐ ํญ์ ๋์๊ธฐ๊ณ ์ค์ค๋ก์๊ฒ ์๊ธฐ์ํค๋ ค๊ณ ๋ ธ๋ ฅํ๋ ๊ฒ์, ์ค๋๋ ์ ๋ชจ๋ธ๋ค์ด ์๋ฌด๋ฆฌ ๋๋๋ค๊ณ ํด๋ โ ๋งค๋ ์ ํฌ๊ฐ ์ด๋ฃจ๋ ๋ฐ์ ์ ์ ๋ง ๊น์ง ๋๋ผ์ง๋ง โ ์ฌ๋ฌ๋ถ์ด ์ค๋ ์ฌ์ฉํ์๋ ๋ชจ๋ธ์ ์ฌ๋ฌ๋ถ์ด ํ์ ์ฌ์ฉํ๊ฒ ๋ AI ๋ชจ๋ธ ์ค ๊ฐ์ฅ ์ ์ข์ ๋ชจ๋ธ์ผ ๊ฒ๋๋ค. ๋ค์ ๋งํด, ์ค๋ ์ฌ์ฉํ์๋ AI ๋ชจ๋ธ์ด ์ฌ๋ฌ๋ถ์ด ์์ผ๋ก ์ฌ์ฉํ์ค ๋ชจ๋ธ ์ค ๊ฐ์ฅ ์ ์ข์ ๋ชจ๋ธ์ด๋ผ๋ ๊ฑฐ์ฃ . ๊ทธ๋ฆฌ๊ณ 1๋ ํ์๋, ์ค๋๋ ์ AI ๋ชจ๋ธ๋ก ๋ค์ ๋์๊ฐ๋ ๊ฒ์ ์์์กฐ์ฐจ ํ ์ ์์ ๊ฒ๋๋ค. ์๋ํ๋ฉด AI ๋ชจ๋ธ๋ค์ ์ ๋ง ๋น ๋ฅธ ์๋๋ก ๋ฐ์ ํ๊ณ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ด๋ ๊ณง ์ง๋ฅ์ ๋ํ ์์๊ฐ ์ฌ์ค์ ๋ฌดํํ๋ค๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค. ์ ํฌ๋ ์ฌ๊ธฐ์ ๋ฉ์ถ์ง ์์ ๊ฒ๋๋ค. |
| CEO & Director: And in your world, in your new role, where you're looking at verticals that are somewhat untapped today, math and science and things of that nature. When you think about the Arm AGI CPU or more broadly, what does more compute do for you in that space? | **CEO & Director:** ์๋ก์ด ์ญํ ์์ ํ์ฌ ๋ค์ ๋ฏธ๊ฐ์ฒ๋ ์์ง ์์ฅ(verticals), ์๋ฅผ ๋ค์ด ์ํ ๋ฐ ๊ณผํ ๋ถ์ผ ๋ฑ์ ๋ณด๊ณ ๊ณ์ ๋ฐ์. Arm AGI CPU๋ ๋ ๋์ ๊ด์ ์์ ๋ณผ ๋, ๊ทธ๋ฌํ ๋ถ์ผ์์ ๋ ๋ง์ ์ปดํจํ ํ์(compute)๊ฐ ์ด๋ค ์๋ฏธ๋ฅผ ๊ฐ์ง๋์? |
| OpenAI, L.L.C.: Well, I mean, the more compute you have, the more inference you're able to do, the longer the rollouts you're able to do. AI, as we go -- as we're sort of progressing from this world of AI as chat to AI solving harder and harder problems. And just like you or me, when you solve harder and harder problems, you're going to need to think a little bit longer. So the more important problems we solve as we start to think about things like enterprise AGI, science, you're going to need more compute, which means if you can draw the power that you have, which will always be finite, you can draw that more efficiently. You can do more and we can solve more problems. | **OpenAI, L.L.C.:** ๋ค, ๊ทธ๋ฌ๋๊น ์ปดํจํ
ํ์(compute)๊ฐ ๋ง์์๋ก ๋ ๋ง์ ์ถ๋ก (inference)์ ํ ์ ์๊ณ , ๋ ๊ด๋ฒ์ํ๊ฒ ์ ์ฉํ ์ ์์ต๋๋ค. AI๋ ์ฑ๋ด(chat)์ผ๋ก์์ ์ญํ ์ ๋์ด ์ ์ ๋ ์ด๋ ค์ด ๋ฌธ์ ๋ค์ ํด๊ฒฐํ๋ ๋ฐฉํฅ์ผ๋ก ๋ฐ์ ํ๊ณ ์์ต๋๋ค. ์ฐ๋ฆฌ ๋ชจ๋๊ฐ ๊ทธ๋ ๋ฏ์ด, ๋ ์ด๋ ค์ด ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๋ ค๋ฉด ๋ ๋ง์ ์๊ฐ๊ณผ ๊น์ ์ฌ๊ณ ๊ฐ ํ์ํ์ฃ . ๋ฐ๋ผ์ ๊ธฐ์ ์ฉ AGI(Artificial General Intelligence, ์ธ๊ณต ์ผ๋ฐ ์ง๋ฅ)๋ ๊ณผํ ๋ถ์ผ์ ๊ฐ์ ๋ ์ค์ํ ๋ฌธ์ ๋ค์ ํด๊ฒฐํ๊ธฐ ์์ํ๋ฉด, ๋ ๋ง์ ์ปดํจํ ํ์๊ฐ ํ์ํ๊ฒ ๋ฉ๋๋ค. ์ด๋ ์ฐ๋ฆฌ๊ฐ ๊ฐ์ง ์ ํํ ์ ๋ ฅ์ ๋ ํจ์จ์ ์ผ๋ก ํ์ฉํ ์ ์๋ค๋ฉด, ๋ ๋ง์ ์ผ์ ํด๋ผ ์ ์๊ณ , ๋ ๋ง์ ๋ฌธ์ ๋ค์ ํด๊ฒฐํ ์ ์๋ค๋ ์๋ฏธ์ ๋๋ค. |
| CEO & Director: And for you personally, what are you most excited about broadly in terms of everything we see going on with AI? | **CEO & Director:** ๊ทธ๋ฆฌ๊ณ ๊ฐ์ธ์ ์ผ๋ก๋, AI ๋ถ์ผ์์ ํ์ฌ ์ผ์ด๋๊ณ ์๋ ์ ๋ฐ์ ์ธ ์ํฉ๋ค์ ๋ณด์ค ๋, ์ด๋ค ์ ์ด ๊ฐ์ฅ ๊ธฐ๋๋์๋์ง์? |
| OpenAI, L.L.C.: Well, I mean, I kind of -- I think I have the coolest job in the world. I get to work on accelerating science with AI. And you've seen sort of a revolution in the past, even just 3 months with GPT 5.2, 5.4, codecs. I mean it used to be that people said, oh, well, these are just stochastic parsers. They're sampling from a distribution of data that they were trained on, but they can't do novel things. Now we're seeing every day AI solve open problems in science, in mathematics and physics and biology. We're seeing AI help us understand the nature of the universe. We're seeing AI work for weeks on end using a robotic lab to run 36,000 different experiments to optimize the synthesis of a new protein faster and better than any human could. So it's an exciting world. I think science is going to move faster than ever, and it's all built on the kind of infrastructure that you're providing. | **OpenAI, L.L.C.:** ์์งํ ์ ๋ ์ ๊ฐ ์ธ์์์ ๊ฐ์ฅ ๋ฉ์ง ์ง์
์ ๊ฐ์ง๊ณ ์๋ค๊ณ ์๊ฐํฉ๋๋ค. AI๋ฅผ ํ์ฉํด ๊ณผํ ๋ฐ์ ์ ๊ฐ์ํํ๋ ์ผ์ด์ฃ . ์ฌ๋ฌ๋ถ๋ ์์๋ค์ํผ, ์ง๋ ๋จ 3๊ฐ์ ๋์์๋ GPT 5.2, 5.4, ์ฝ๋ฑ์ค(codecs) ๋ฑ์์ ์ผ์ข ์ ํ๋ช ์ด ์ผ์ด๋ฌ์ต๋๋ค. ์์ ์๋ ์ฌ๋๋ค์ด '์, ์ด๊ฑด ๊ทธ์ ํ๋ฅ ์ ํ์(stochastic parsers)์ผ ๋ฟ์ด์ผ. ํ์ต๋ ๋ฐ์ดํฐ ๋ถํฌ์์ ์ํ๋งํ๋ ๊ฒ์ผ ๋ฟ, ์๋ก์ด ๊ฒ์ ๋ง๋ค์ด๋ผ ์๋ ์์ด'๋ผ๊ณ ๋งํ๊ณค ํ์ต๋๋ค. ํ์ง๋ง ์ด์ ์ฐ๋ฆฌ๋ AI๊ฐ ๊ณผํ, ์ํ, ๋ฌผ๋ฆฌ, ์๋ฌผํ ๋ถ์ผ์ ๋ฏธํด๊ฒฐ ๋ฌธ์ ๋ค์ ๋งค์ผ ํด๊ฒฐํ๋ ๊ฒ์ ๋ชฉ๊ฒฉํ๊ณ ์์ต๋๋ค. AI๊ฐ ์ฐ์ฃผ์ ๋ณธ์ง์ ์ดํดํ๋ ๋ฐ ๊ธฐ์ฌํ๋ ๋ชจ์ต๋ ๋ณด๊ณ ์์ต๋๋ค. ๋ํ, AI๊ฐ ๋ก๋ด ์คํ์ค(robotic lab)์ ํ์ฉํด ๋ช ์ฃผ๊ฐ ์ฌ์ง ์๊ณ 36,000๊ฐ์ง์ ๋ค์ํ ์คํ์ ์ํํ๋ฉฐ, ์๋ก์ด ๋จ๋ฐฑ์ง ํฉ์ฑ์ ๊ทธ ์ด๋ค ์ธ๊ฐ๋ณด๋ค๋ ๋น ๋ฅด๊ณ ํจ์จ์ ์ผ๋ก ์ต์ ํํ๋ ๋ชจ์ต๋ ๋ณด๊ณ ์์ต๋๋ค. ์ ๋ง ํฅ๋ฏธ์ง์งํ ์ธ์์ ๋๋ค. ๊ณผํ์ ๊ทธ ์ด๋ ๋๋ณด๋ค ๋น ๋ฅด๊ฒ ์ง๋ณดํ ๊ฒ์ด๋ผ๊ณ ๋ด ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๋ฌํ ๋ชจ๋ ์ง๋ณด๋ ๋ฐ๋ก ์ฌ๋ฌ๋ถ์ด ์ ๊ณตํ๊ณ ๊ณ์ ์ธํ๋ผ(infrastructure)๋ฅผ ๊ธฐ๋ฐ์ผ๋ก ์ด๋ฃจ์ด์ง ๊ฒ์ ๋๋ค. |
| CEO & Director: We are grateful for your support. Kevin, thanks. | **CEO & Director:** ์ฑ์์ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ง๋ฌธ ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. |
| OpenAI, L.L.C.: Thank you so much. | **OpenAI, L.L.C.:** ์ ๋ง ๊ฐ์ฌํฉ๋๋ค. |
| CEO & Director: Thank you. I love the idea that the model that we're using today is about as bad as it's going to get. That's crazy. I want to repeat, in case I wasn't crystal clear on the first go around. We are now delivering IP, CSS and chips. IP, CSS and chips. Contact your local sales representative. Will is here. He can be reached afterwards. Now seriously, I talked earlier about the ecosystem of ecosystems and none of this could be done without the ecosystem that we have, particularly around Neoverse. We have many partners that we work with on the supply side, whether it's around memory or connectivity. But we've also got great customers who use our IP today. And they are so supportive of what we're doing. Santosh talked about the demand. The market is so large. The demand is so significant that no one company can serve it. So what I'd like to do is rather than me going on and on and talking about it is have you hear from some of our partners and friends who I think you'll probably recognize a few. [Presentation] | **CEO & Director:** ๊ฐ์ฌํฉ๋๋ค. ์ค๋ ์ ํฌ๊ฐ ํ์ฉํ๋ ๋ชจ๋ธ์ด ์ง๊ธ์ด ๊ฐ์ฅ ์ ์ข์ ์ํ์ด๊ณ ์์ผ๋ก๋ ๋ ์ข์์ง ์ผ๋ง ๋จ์๋ค๋ ์ ์ด ์ ๋ง ์ข์ต๋๋ค. ๋๋์ต๋๋ค. ํน์ ์ ๊ฐ ์ฒซ ์ค๋ช ์์ ๋ช ํํ ์ ๋ฌํ์ง ๋ชปํ์๊น ๋ด ๋ค์ ํ๋ฒ ๋ง์๋๋ฆฝ๋๋ค. ์ ํฌ๋ ์ด์ IP (์ง์ ์ฌ์ฐ, Intellectual Property), CSS (์นฉ๋ ์์คํ ์๋ฃจ์ , Chiplet System Solution), ๊ทธ๋ฆฌ๊ณ ์นฉ (Chip)์ ์ ๊ณตํ๊ณ ์์ต๋๋ค. IP, CSS, ์นฉ์ ๋๋ค. ์ง์ญ ์์ ๋ด๋น์์๊ฒ ๋ฌธ์ํ์ญ์์ค. Will์ด ์ฌ๊ธฐ ์ ์์ต๋๋ค. ๋์ค์ ๊ทธ์๊ฒ ์ฐ๋ฝํ์๋ฉด ๋ฉ๋๋ค. ์ด์ ์ง์งํ๊ฒ ๋ง์๋๋ฆฌ์๋ฉด, ์ ๊ฐ ์์ '์ํ๊ณ์ ์ํ๊ณ (ecosystem of ecosystems)'์ ๋ํด ๋ง์๋๋ ธ๋๋ฐ, ์ด ๋ชจ๋ ๊ฒ์ ํนํ Neoverse (๋ค์ค๋ฒ์ค)๋ฅผ ์ค์ฌ์ผ๋ก ์ ํฌ๊ฐ ๊ตฌ์ถํ ์ํ๊ณ ์์ด๋ ๋ถ๊ฐ๋ฅํ์ ๊ฒ์ ๋๋ค. ์ ํฌ๋ ๊ณต๊ธ ์ธก๋ฉด์์ ๋ฉ๋ชจ๋ฆฌ (Memory)๋ ์ฐ๊ฒฐ์ฑ (Connectivity)์ด๋ ๋ง์ ํํธ๋๋ค๊ณผ ํ๋ ฅํ๊ณ ์์ต๋๋ค. ํ์ง๋ง ์ค๋๋ ์ ํฌ IP๋ฅผ ์ฌ์ฉํ๋ ํ๋ฅญํ ๊ณ ๊ฐ์ฌ๋ค๋ ์์ต๋๋ค. ๊ทธ๋ค์ ์ ํฌ๊ฐ ํ๋ ์ผ์ ๋ํด ๋งค์ฐ ์ง์ง์ ์ ๋๋ค. Santosh๊ฐ ์์ (Demand)์ ๋ํด ์ด์ผ๊ธฐํ๋ฏ์ด, ์์ฅ์ ์ ๋ง ๊ฑฐ๋ํฉ๋๋ค. ์์๊ฐ ์๋ ๋ง๋ํด์ ์ด๋ค ํ ๊ธฐ์ ๋ ์ด ๋ชจ๋ ์์๋ฅผ ๋ค ๊ฐ๋นํ ์ ์์ ์ ๋์ ๋๋ค. ๊ทธ๋์ ์ ๊ฐ ๊ณ์ํด์ ๊ธธ๊ฒ ๋ง์๋๋ฆฌ๋ ๊ฒ๋ณด๋ค๋, ์ ํฌ ํํธ๋์ฌ๋ค๊ณผ ๊ด๊ณ์๋ถ๋ค์ ์ด์ผ๊ธฐ๋ฅผ ์ง์ ๋ค์ด๋ณด์๋ ๊ฒ ์ข๊ฒ ์ต๋๋ค. ์๋ง ๋ช ๋ถ์ ์ ์์๋ ๋ถ๋ค์ผ ๊ฒ๋๋ค. |
| CEO & Director: Charlie and Matt and Sanjay and even my old boss did better than I could in terms of talking about this. But this has not happened without a fantastic partnership and support from the ecosystem. Now I know you are dying to hear about this product as am I. And I'm now going to turn over to Mohamed Awad, who is going to tell you all about the Arm's AGI CPU, and why it is absolutely amazing. Mohamed? | **CEO & Director:** ์ฐฐ๋ฆฌ, ๋งท, ์ฐ์ ์ด, ๊ทธ๋ฆฌ๊ณ ์ฌ์ง์ด ์ ์ ์์ฌ๋ถ๊น์ง๋ ์ด ์ ์ ๋ํด์๋ ์ ๋ณด๋ค ํจ์ฌ ๋ ์ ์ค๋ช
ํด ์ฃผ์
จ์ต๋๋ค. ํ์ง๋ง ์ด ๋ชจ๋ ๊ฒ์ ํ๋ฅญํ ํํธ๋์ญ๊ณผ ์ํ๊ณ(ecosystem)์ ์ง์ ์์ด๋ ๋ถ๊ฐ๋ฅํ์ ๊ฒ๋๋ค. ์ด์ ์ฌ๋ฌ๋ถ๊ป์ ์ด ์ ํ์ ๋ํด ์ ๋ง ๊ถ๊ธํดํ๊ณ ๊ณ์๋ค๋ ๊ฒ์ ์ ๋ ์ ์๊ณ ์์ต๋๋ค. ์ ์ญ์ ๋ง์ฐฌ๊ฐ์ง๊ณ ์. ์ด์ ๋ชจํ๋ฉ๋ ์์๋(Mohamed Awad)์๊ฒ ๋ง์ดํฌ๋ฅผ ๋๊ธฐ๊ฒ ์ต๋๋ค. ๊ทธ๊ฐ Arm์ AGI CPU์ ๋ํด ์์ธํ ์ค๋ช ํด ๋๋ฆด ๊ฒ์ด๋ฉฐ, ์ด ์ ํ์ด ์ ์ ๋ง ๋๋ผ์ด์ง ๋ง์ํด ๋๋ฆด ๊ฒ์ ๋๋ค. ๋ชจํ๋ฉ๋? |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit Thank you, Rene. Thank you, Santosh. Thank you, Kevin. Thanks to all of you. Thanks to the entire Arm team that made today possible. We have been looking forward to this, and it is so exciting to be here. It's so exciting to talk to you guys. Thank you. Thank you. Thank you. Rene talked about how the world is transitioning from sort of legacy data centers to AGI data centers to agentic data centers heading down this path and how the CPU is at the heart of it. We've designed our AGI CPU around 3 simple principles. We believe that's the heart of what we're doing. It's the heart of what we focused on. It's the heart of how we think about it. First, performance, performance, performance. With this many threads going on, with this much work to do, with this much orchestration to happen, you can't slow down 24 hours a day, as Rene said, these agents are going to be running. And if they're not performing fast enough, then the rest of that infrastructure that's relying on it grinds to a halt. So we focused on performance. Second, we focused on scale. The scale of what we're talking about here is just incredible. You heard Santosh talk about gigawatts, gigawatts, scale at the CPU level, scale at the Board level, scale at the rack level, scale at the warehouse level, all the way up. We focused on that. And finally, we focused on efficiency, maybe most importantly, because at the end of the day, with this much at stake, with this much compute we're trying to deploy, we're not going to get there unless we provide that performance, we provide that scale and we do it in an efficient package. Those are the principles that have guided us. Wait for it. Those are the principles that have guided us, and we refuse to compromise. We've designed on all 3. Play the video now. [Presentation] | **Mohamed Awad:** Rene, Santosh, Kevin๊ป ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์๋ฆฌ์ ๊ณ์ ๋ชจ๋ ๋ถ๋ค๊ป๋ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ค๋ ์ด ์๋ฆฌ๋ฅผ ๋ง๋ จํด ์ฃผ์ Arm ํ ์ ์ฒด์๋ ๊น์ด ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ ํฌ๋ ์ด ์๊ฐ์ ์ค๋ซ๋์ ๊ณ ๋ํด ์์ต๋๋ค. ์ด ์๋ฆฌ์ ํจ๊ปํ๊ฒ ๋์ด ์ ๋ง ๊ธฐ์๊ณ , ์ฌ๋ฌ๋ถ๊ณผ ์ด์ผ๊ธฐ ๋๋ ์ ์๊ฒ ๋์ด ๋งค์ฐ ์ค๋ ๋๋ค. ์ ๋ง ๊ฐ์ฌํฉ๋๋ค. Rene๊ฐ ์ธ๊ธํ๋ฏ์ด, ํ์ฌ ์ธ์์ ๊ธฐ์กด ๋ฐ์ดํฐ์ผํฐ (legacy data centers)์์ AGI ๋ฐ์ดํฐ์ผํฐ (AGI data centers)๋ก, ๊ทธ๋ฆฌ๊ณ ์์ด์ ํธ ๊ธฐ๋ฐ ๋ฐ์ดํฐ์ผํฐ (agentic data centers)๋ก ์ ํ๋๋ ๊ธธ์ ๊ฑท๊ณ ์์ผ๋ฉฐ, ๊ทธ ์ค์ฌ์๋ CPU๊ฐ ์์ต๋๋ค. ์ ํฌ๋ ์ด๋ฌํ AGI CPU๋ฅผ ์ธ ๊ฐ์ง ๊ฐ๋จํ ์์น์ ์ค์ฌ์ผ๋ก ์ค๊ณํ์ต๋๋ค. ์ ํฌ๋ ์ด ์์น๋ค์ด ์ ํฌ๊ฐ ํ๋ ์ผ์ ํต์ฌ์ด์, ์ ํฌ๊ฐ ์ง์คํด ์จ ๋ถ๋ถ์ ํต์ฌ์ด๋ฉฐ, ์ ํฌ๊ฐ ์ด ๋ฌธ์ ๋ฅผ ๋ฐ๋ผ๋ณด๋ ๋ฐฉ์์ ํต์ฌ์ด๋ผ๊ณ ๋ฏฟ์ต๋๋ค. ์ฒซ์งธ, ์ฑ๋ฅ์ ๋๋ค. ์ฑ๋ฅ, ๊ทธ๋ฆฌ๊ณ ๋ฌด์๋ณด๋ค ์ฑ๋ฅ์ ๋๋ค. ์ด๋ ๊ฒ ๋ง์ ์์ ์ค๋ ๋(threads)๊ฐ ๋์์ ์งํ๋๊ณ , ์ฒ๋ฆฌํด์ผ ํ ์ผ์ด ๋ง์ผ๋ฉฐ, ๋ง์ ์กฐ์จ(orchestration)์ด ํ์ํ ์ํฉ์์๋ 24์๊ฐ ๋ด๋ด ์๋๋ฅผ ๋ฆ์ถ ์ ์์ต๋๋ค. ๋ฅด๋ค๊ฐ ๋งํ๋ฏ์ด, ์ด ์์ด์ ํธ๋ค์ ๊ณ์ ๊ฐ๋๋์ด์ผ ํฉ๋๋ค. ๋ง์ฝ ์ด๋ค์ด ์ถฉ๋ถํ ๋น ๋ฅด๊ฒ ์๋ํ์ง ์์ผ๋ฉด, ์ด์ ์์กดํ๋ ๋๋จธ์ง ์ธํ๋ผ ์ ์ฒด๊ฐ ๋ฉ์ถฐ ์๊ฒ ๋ฉ๋๋ค. ๊ทธ๋์ ์ ํฌ๋ ์ฑ๋ฅ(performance)์ ์ง์คํ์ต๋๋ค. ๋ ๋ฒ์งธ๋ก, ์ ํฌ๋ ์ค์ผ์ผ(scale)์ ์ง์คํ์ต๋๋ค. ์ ํฌ๊ฐ ์ด์ผ๊ธฐํ๋ ์ค์ผ์ผ์ ์ ๋ง ์์ฒญ๋ฉ๋๋ค. ์ฐํ ์๊ฐ ๊ธฐ๊ฐ์ํธ(gigawatts), ๊ธฐ๊ฐ์ํธ์ ๋ํด ์ด์ผ๊ธฐํ๋ ๊ฒ์ ๋ค์ผ์ จ์ ๊ฒ๋๋ค. CPU ์์ค, ๋ณด๋(Board) ์์ค, ๋(rack) ์์ค, ์ฐฝ๊ณ (warehouse) ์์ค, ๊ทธ๋ฆฌ๊ณ ๊ทธ ์ด์์ผ๋ก ํ์ฅ๋๋ ์ค์ผ์ผ ๋ง์ ๋๋ค. ์ ํฌ๋ ๊ทธ ์ ์ ์ง์คํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ง์ง๋ง์ผ๋ก, ์๋ง๋ ๊ฐ์ฅ ์ค์ํ๊ฒ๋ ์ ํฌ๋ ํจ์จ์ฑ์ ์ง์คํ์ต๋๋ค. ๊ถ๊ทน์ ์ผ๋ก ์ด๋ ๊ฒ ์ค์ํ ์ฌ์์ด ๊ฑธ๋ ค ์๊ณ , ๋ฐฐํฌํ๋ ค๋ ์ปดํจํ ์์์ ๊ท๋ชจ๊ฐ ์๋นํ ๋งํผ, ์ฑ๋ฅ๊ณผ ๊ท๋ชจ๋ฅผ ํ๋ณดํ๊ณ ์ด๋ฅผ ํจ์จ์ ์ธ ๋ฐฉ์์ผ๋ก ์ ๊ณตํ์ง ์๋๋ค๋ฉด ๋ชฉํ๋ฅผ ๋ฌ์ฑํ ์ ์์ ๊ฒ์ด๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ด๊ฒ์ด ์ ํฌ๋ฅผ ์ด๋์ด์จ ์์น๋ค์ ๋๋ค. ์ด๊ฒ์ด ์ ํฌ๋ฅผ ์ด๋์ด์จ ์์น๋ค์ด๋ฉฐ, ์ ํฌ๋ ํํํ์ง ์์ ๊ฒ์ ๋๋ค. ์ ํฌ๋ ์ด ์ธ ๊ฐ์ง ์์น ๋ชจ๋๋ฅผ ๊ธฐ๋ฐ์ผ๋ก ์ค๊ณํ์ต๋๋ค. ์ด์ ์์์ ์ฌ์ํด ์ฃผ์ญ์์ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit I got to tell you, we are so, so proud. Our team has done a fantastic job on this, and it's really been designed for the ground up from this. Let me tell you a little bit more about what you just saw because I know there was a lot packed into that video. Arm AI CPU starts off with our standard Neoverse V3 compute subsystem. That's the same compute subsystem we make available to the entire ecosystem, and we have other partners building on it. We're incredibly proud of that. We pack in 136 of those cores, which are very high-performance cores designed to be high performance. Our V-Series is our most performant line, and you've seen it set records across lots of different hyperscale implementations and those of other system providers. We add to that a dedicated 2-megabyte L2 cache, and we support up to 3.7 gigahertz in frequency. But it's not just the CPU core. We thought about the entire system. As part of the design, we went with 96 lanes of PCIe Gen 6, which supports CXL 3, which means you can attach it to any accelerator you like. It also means that you can support things like memory expansion. On the memory side, DDR5, with up to 6 gigabytes per second of memory per core, which can be sustained to each core. That is unique. That level of performance to every single core on both the I/O and the memory is unique to us in this type of a package, in this type of a performance point at this efficiency level. And it's not just about the bandwidth. It's not just about the I/O. It's about the overall design. You see we designed the whole thing to be low latency so that you could get to less than 100 nanoseconds of latency from the memory. We did so by sticking with a dual chiplet design, each chiplet having all of the memory in the I/O directly on it rather than having to worry about complicated pneuma domains and multiple hops across the silicon. The result, it wasn't a typo in the slide, 300-watt TDP, 300 watts. That is amazing. It's built on a 3-nanometer TSMC process and allows for that maximum compute density. This is what purpose-built design looks like. This is what we're so proud of. The AGI CPU is breaking records all over the place for performance, for scale and for efficiency. You saw some of that in the video. This is a standard OCP air-cooled rack. Nothing unique about it, nothing especially exotic about it, just OCP rack, standard, right? That's our Head of OCP right there clapping, just so everyone's aware. 36 kilowatts, we pack in over 8,000 of these performance CPU cores. We do so by going to a 2-node 1-use server, 30 of them. You can't do that in other systems because the power consumption is just too high. This is setting records for air cooled. But you know what, if you want liquid cooled, we can do that, too. Over 45,000 CPU cores and a 200-kilowatt, again, a standard rack from OCP. Over a petabyte of memory in this thing. And oh, by the way, fun fact on this one. It's a 200-kilowatt rack. We actually will consume about half that much power. We ran out of space. That's why we couldn't put more cores in there. Yes, it's pretty wild. The scale of this stuff is crazy. It's just really inspiring. These are standard racks, but there's nothing else like them. To get to this level of efficiency, we really had to design the Arm AGI CPU from the ground up. And that's what I'm so proud about, and I'll tell you about in a minute. But before I get that, I want to just talk about the fact that these are standard racks because it's not only about the fact that we are using -- there we go. It's not only that we're taking from OCP and leveraging some of their platforms, we're also giving back. We're in the process of making a bunch of contributions to OCP, things like Arm server ready, authenticated access control and diagnostic tools. And those contributions won't just be for the Arm AGI CPU, they will apply to the entire ecosystem. So it will make it available so that -- and they will be beneficial for all Arm-based platforms because it really is an ecosystem that we're building here. Arm has always been about nurturing and partnering with the ecosystem. That's always been core to our identity. And those relationships are paying great dividends now. You saw the video that Rene played, and we're so grateful about all those partnerships. It's those partnerships actually, which have allowed us to build the Arm AGI CPU. Some of them are very long-standing. Partners like TSMC and Samsung and Micron and SK hynix, these are partners that we've been working with for well over -- for decades, literally for decades. And we've also got some new partnerships, which is why we're so proud to say that the Arm AGI CPU -- went a little bit far there. Can you go back, please? -- is available now. Said it there. It doesn't say it there. Yes. So Arm available. Arm AGI CPU is available now, and we're so proud of that. It's actually in customers' hands. Customers are actually evaluating it as we speak. We are ready to go. And we're so grateful for our partners, both on the ODM side, on the memory side, on the CPU side, on the manufacturing side who have helped us get to this point. We'll be in production by the end of the year, and we are excited to share that with you. We've got -- today, we've got firmware ready to go. We've got specifications ready to go. I talked to you about platforms. I talked to you about supply. The one thing I haven't talked to you about yet is software. Let's talk about software. Now the next slide. Okay. So the reality is that Arm has been investing in data center software ecosystem for well over 15 years. I don't know if everyone understands how long we've been investing in the software ecosystem. For the beginning of that time, in the early days, it was just Arm investing in the software ecosystem. And then something happened in 2019. We launched Arm Neoverse. And what Arm Neoverse did, that compute platform when we launched it, it allowed our customers to begin to launch products with a much lower barrier to entry. It allowed them to build their own silicon and start to coalesce around a common platform. And that started that software flywheel turning. You see when tech leaders started adopting Neoverse, they started to optimize software around it. And the more of those tech leaders that adopted Neoverse, the faster that flywheel started to spin. Today, we've got AWS and Google and Meta and Microsoft and Oracle and NVIDIA, all investing alongside us in the software ecosystem. And that really was what allowed us to kind of really make some great traction in software. Together, we've made Arm a first-class citizen on most modern software packages. And for our AI software ecosystem specifically, not only are we a first-class citizen, not only the software run well on Arm, software actually runs best on Arm. And the reason for that is very simple. For AI, the Arm software ecosystem, the Arm architecture is the primary CPU architecture in support of AI today. In fact, the work we've done together with technology leaders means that tens of thousands of companies today run their software on Arm in the cloud on over 1.25 billion Arm Neoverse cores, which we've already shipped into data centers around the world. And that growth is only accelerating. That's actually the curve. You see Arm in the data center just works. This is a key point. And I don't know if I'm making it well enough. So I'm going to bring somebody on stage who's got a little experience with software. Paul Saab has worked on Meta's infrastructure for over 18 years. He's one of the longest tenured employees at the company. There's a laundry list of things that he's been responsible for, including the adoption of flash storage all the way through the implementation of IPV6. Today, he's specifically focused on making AI more efficient in their infrastructure. And that's how we got to know each other. Please welcome Paul Saab. | **Mohamed Awad:** ์ ๋ง ์๋์ค๋ฝ์ต๋๋ค. ์ ํฌ ํ์ด ์ด ํ๋ก์ ํธ๋ฅผ ์ ๋ง ํ๋ฅญํ๊ฒ ์ํํ์ผ๋ฉฐ, ์ด๋ ์ฒ์๋ถํฐ ์์ ํ ์๋กญ๊ฒ ์ค๊ณ๋์์ต๋๋ค. ๋ฐฉ๊ธ ๋ณด์ ๋ด์ฉ์ ๋ํด ์ข ๋ ์์ธํ ์ค๋ช ํด ๋๋ฆฌ๊ฒ ์ต๋๋ค. ๊ทธ ์์์ ๋ง์ ๋ด์ฉ์ด ๋ด๊ฒจ ์์๋ค๋ ๊ฒ์ ์๊ณ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. Arm AI CPU๋ ์ ํฌ์ ํ์ค Neoverse V3 ์ปดํจํธ ์๋ธ์์คํ (compute subsystem)์ผ๋ก ์์ํฉ๋๋ค. ์ด๋ ์ ํฌ๊ฐ ์ ์ฒด ์ํ๊ณ(ecosystem)์ ์ ๊ณตํ๋ ๋์ผํ ์ปดํจํธ ์๋ธ์์คํ ์ด๋ฉฐ, ๋ค๋ฅธ ํํธ๋์ฌ๋ค๋ ์ด๋ฅผ ๊ธฐ๋ฐ์ผ๋ก ๊ฐ๋ฐํ๊ณ ์์ต๋๋ค. ์ ํฌ๋ ์ด์ ๋ํด ์์ฒญ๋ ์๋ถ์ฌ์ ๋๋๋๋ค. ์ ํฌ๋ 136๊ฐ์ ์ฝ์ด(core)๋ฅผ ํ์ฌํ์ผ๋ฉฐ, ์ด ์ฝ์ด๋ค์ ๋งค์ฐ ๋์ ์ฑ๋ฅ์ ๋ฐํํ๋๋ก ์ค๊ณ๋ ๊ณ ์ฑ๋ฅ ์ฝ์ด๋ค์ ๋๋ค. ์ ํฌ์ V-์๋ฆฌ์ฆ๋ ๊ฐ์ฅ ๋ฐ์ด๋ ์ฑ๋ฅ์ ์๋ํ๋ ๋ผ์ธ์ ์ด๋ฉฐ, ์ฌ๋ฌ๋ถ๋ ๋ณด์ จ๋ฏ์ด, ๋ค์ํ ํ์ดํผ์ค์ผ์ผ(hyperscale) ๊ตฌํ ์ฌ๋ก์ ๋ค๋ฅธ ์์คํ ์ ๊ณต์ ์ฒด๋ค์ ๊ฒฝ์ฐ์์๋ ๊ธฐ๋ก์ ์ธ์ ์ต๋๋ค. ์ฌ๊ธฐ์ ์ ์ฉ 2๋ฉ๊ฐ๋ฐ์ดํธ(MB) L2 ์บ์(L2 cache)๋ฅผ ์ถ๊ฐํ์ผ๋ฉฐ, ์ต๋ 3.7๊ธฐ๊ฐํค๋ฅด์ธ (GHz)์ ์ฃผํ์(frequency)๋ฅผ ์ง์ํฉ๋๋ค. ํ์ง๋ง ๋จ์ํ CPU ์ฝ์ด(CPU core)์๋ง ๊ตญํ๋ ๊ฒ์ด ์๋๋๋ค. ์ ํฌ๋ ์ ์ฒด ์์คํ ์ ๊ณ ๋ คํ์ต๋๋ค. ์ค๊ณ์ ์ผํ์ผ๋ก, ์ ํฌ๋ 96๊ฐ ๋ ์ธ์ PCIe Gen 6๋ฅผ ์ฑํํ์ผ๋ฉฐ, ์ด๋ CXL 3์ ์ง์ํฉ๋๋ค. ๋ค์ ๋งํด, ๊ณ ๊ฐ์ด ์ํ๋ ์ด๋ค ๊ฐ์๊ธฐ(accelerator)์๋ ์ฐ๊ฒฐํ ์ ์๋ค๋ ์๋ฏธ์ ๋๋ค. ์ด๋ ๋ํ ๋ฉ๋ชจ๋ฆฌ ํ์ฅ(memory expansion)๊ณผ ๊ฐ์ ๊ธฐ๋ฅ๋ ์ง์ํ ์ ์๋ค๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค. ๋ฉ๋ชจ๋ฆฌ ์ธก๋ฉด์์๋ DDR5๋ฅผ ์ฌ์ฉํ๋ฉฐ, ์ฝ์ด๋น ์ต๋ ์ด๋น 6๊ธฐ๊ฐ๋ฐ์ดํธ(GB/s)์ ๋ฉ๋ชจ๋ฆฌ ๋์ญํญ์ ์ ๊ณตํ๋๋ฐ, ์ด ๋์ญํญ์ ๊ฐ ์ฝ์ด์ ์ง์์ ์ผ๋ก ์ ์ง๋ ์ ์์ต๋๋ค. ์ด๋ ๋ ๋ณด์ ์ธ ์์ค์ ๋๋ค. I/O์ ๋ฉ๋ชจ๋ฆฌ ์์ชฝ ๋ชจ๋์์ ๋ชจ๋ ๋จ์ผ ์ฝ์ด์ ์ ๊ณต๋๋ ์ด๋ฌํ ์ฑ๋ฅ ์์ค์, ์ด๋ฌํ ์ข ๋ฅ์ ํจํค์ง(package)์์, ์ด ์ ๋์ ์ฑ๋ฅ ์ง์ (performance point)๊ณผ ํจ์จ์ฑ ์์ค(efficiency level)์ ๊ณ ๋ คํ ๋, ์ ํฌ์๊ฒ๋ง ์๋ ๋ ํนํ ๊ฐ์ ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๋ ๋จ์ํ ๋์ญํญ(bandwidth)์๋ง ๊ตญํ๋ ๊ฒ์ด ์๋๋๋ค. I/O์๋ง ๊ตญํ๋ ๊ฒ๋ ์๋๊ณ ์. ํต์ฌ์ ์ ์ฒด์ ์ธ ์ค๊ณ(design)์ ์์ต๋๋ค. ์ ํฌ๋ ์ ์ฒด ์์คํ ์ ์ ์ง์ฐ(low latency)์ผ๋ก ์ค๊ณํ์ฌ ๋ฉ๋ชจ๋ฆฌ์์ 100๋๋ ธ์ด ๋ฏธ๋ง์ ์ง์ฐ ์๊ฐ(latency)์ ๋ฌ์ฑํ ์ ์๋๋ก ํ์ต๋๋ค. ์ด๋ฅผ ์ํด ๋์ผ ์นฉ๋ (chiplet) ์ค๊ณ๋ฅผ ๊ณ ์ํ์ผ๋ฉฐ, ๊ฐ ์นฉ๋ ์ ๋ชจ๋ ๋ฉ๋ชจ๋ฆฌ์ I/O๊ฐ ์ง์ ํตํฉ๋์ด ๋ณต์กํ ๋ด๋ง ๋๋ฉ์ธ(pneuma domain)์ด๋ ์ค๋ฆฌ์ฝ์ ๊ฐ๋ก์ง๋ฅด๋ ์ฌ๋ฌ ๋ฒ์ ํ(hop)์ ๊ฑฑ์ ํ ํ์๊ฐ ์๋๋ก ํ์ต๋๋ค. ๊ทธ ๊ฒฐ๊ณผ, ์ฌ๋ผ์ด๋์ ์คํ๊ฐ ์๋์์ต๋๋ค. 300์ํธ์ TDP (์ด ์ค๊ณ ์ ๋ ฅ, Thermal Design Power), 300์ํธ์ ๋๋ค. ์ ๋ง ๋๋ผ์ด ์์น์ ๋๋ค. ์ด๋ TSMC์ 3๋๋ ธ๋ฏธํฐ ๊ณต์ (process)์ผ๋ก ์ ์๋์์ผ๋ฉฐ, ์ต๋ ์ปดํจํ ๋ฐ๋(compute density)๋ฅผ ๊ฐ๋ฅํ๊ฒ ํฉ๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ๋ชฉ์ ์ ๋ง๊ฒ ์ค๊ณ๋ ๋์์ธ์ ๋ชจ์ต์ ๋๋ค. ์ ํฌ๊ฐ ๋งค์ฐ ์๋์ค๋ฌ์ํ๋ ๋ถ๋ถ์ด์ฃ . AGI CPU๋ ์ฑ๋ฅ, ํ์ฅ์ฑ(scale), ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ(efficiency) ๋ฉด์์ ๋ชจ๋ ๊ธฐ๋ก์ ๊ฒฝ์ ํ๊ณ ์์ต๋๋ค. ๋น๋์ค์์ ๊ทธ ์ผ๋ถ๋ฅผ ํ์ธํ์ จ์ ๊ฒ๋๋ค. ์ด๊ฒ์ ํ์ค OCP (Open Compute Project) ๊ณต๋ญ์ ๋(air-cooled rack)์ ๋๋ค. ํน๋ณํ๊ฑฐ๋ ์ด์์ ์ธ ๊ฑด ์์ต๋๋ค. ๊ทธ์ ํ์ค OCP ๋(rack)์ด์ฃ , ์์์ฃ ? ์ ๊ธฐ ๋ฐ์ ์น๊ณ ๊ณ์ ๋ถ์ด ์ ํฌ OCP ์ด๊ด์ด์ญ๋๋ค. ์ฐธ๊ณ ๋ก ๋ง์๋๋ฆฝ๋๋ค. ์ ํฌ๋ 36ํฌ๋ก์ํธ(kW)๋ก 8,000๊ฐ ์ด์์ ๊ณ ์ฑ๋ฅ CPU ์ฝ์ด(CPU core)๋ฅผ ํ์ฌํ์ต๋๋ค. ์ด๋ฅผ ์ํด 2๋ ธ๋(node) 1U ์๋ฒ(server) 30๊ฐ๋ฅผ ์ฌ์ฉํ์ฃ . ๋ค๋ฅธ ์์คํ ์์๋ ์ ๋ ฅ ์๋น๋(power consumption)์ด ๋๋ฌด ๋์์ ์ด๋ ๊ฒ ๊ตฌํํ๊ธฐ๊ฐ ์ด๋ ต์ต๋๋ค. ๊ณต๋ญ์(air-cooled)์ผ๋ก๋ ๊ธฐ๋ก์ ์ธ ์์ค์ ๋๋ค. ํ์ง๋ง ์๋ญ์(liquid-cooled)์ ์ํ์๋ฉด, ๊ทธ๊ฒ๋ ๋ฌผ๋ก ๊ฐ๋ฅํฉ๋๋ค. OCP์ ํ์ค ๋์ 45,000๊ฐ ์ด์์ CPU ์ฝ์ด์ 200ํฌ๋ก์ํธ(kW)๋ฅผ ํ์ฌํ์ต๋๋ค. ๋ค์ ํ๋ฒ ๊ฐ์กฐํ์ง๋ง, ์ด๊ฒ๋ ํ์ค ๋์ ๋๋ค. ์ฌ๊ธฐ์ 1ํํ๋ฐ์ดํธ(PB) ์ด์์ ๋ฉ๋ชจ๋ฆฌ(memory)๊ฐ ๋ค์ด๊ฐ๋๋ค. ๊ทธ๋ฆฌ๊ณ ํ ๊ฐ์ง ์ฌ๋ฏธ์๋ ์ฌ์ค์ ๋ง์๋๋ฆฌ์๋ฉด์. 200ํฌ๋ก์ํธ(kW) ๋์ธ๋ฐ, ์ค์ ๋ก๋ ๊ทธ ์ ๋ฐ ์ ๋์ ์ ๋ ฅ๋ง ์๋นํฉ๋๋ค. ๊ณต๊ฐ์ด ๋ถ์กฑํด์ ๋ ๋ง์ ์ฝ์ด๋ฅผ ๋ฃ์ ์ ์์์ต๋๋ค. ๋ค, ์ ๋ง ๋๋จํฉ๋๋ค. ์ด ๊ท๋ชจ๋ ์์์ ์ด์ํ๊ณ , ์ ๋ง ๊ณ ๋ฌด์ ์ ๋๋ค. ์ด๊ฒ๋ค์ ํ์ค ๋(standard rack)์ด์ง๋ง, ๊ทธ ์ด๋ค ๊ฒ๊ณผ๋ ๋น๊ตํ ์ ์์ ๋งํผ ํน๋ณํฉ๋๋ค. ์ด๋ฌํ ์์ค์ ํจ์จ์ฑ(efficiency)์ ๋ฌ์ฑํ๊ธฐ ์ํด, ์ ํฌ๋ Arm AGI CPU๋ฅผ ์ฒ์๋ถํฐ ์์ ํ ์๋กญ๊ฒ ์ค๊ณํด์ผ ํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ฐ๋ก ๊ทธ ์ ์ด ์ ๊ฐ ์ ๋ง ์๋์ค๋ฝ๊ฒ ์๊ฐํ๋ ๋ถ๋ถ์ด๋ฉฐ, ์ ์ ํ์ ๋ ์์ธํ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ํ์ง๋ง ๊ทธ ์ ์, ์ด๊ฒ๋ค์ด ํ์ค ๋์ด๋ผ๋ ์ ์ ๋ํด ๋ง์๋๋ฆฌ๊ณ ์ถ์ต๋๋ค. ์ ํฌ๋ ๋จ์ํ OCP์ ํ๋ซํผ์ ํ์ฉํ๋ ๊ฒ์ ๋์ด, ์ ํฌ์ ๊ธฐ์ฌ๋ฅผ ํตํด ๋ค์ ํ์ํ๊ณ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ ํฌ๋ ํ์ฌ OCP์ Arm ์๋ฒ ๋ ๋(Arm server ready), ์ธ์ฆ๋ ์ ๊ทผ ์ ์ด(authenticated access control), ๊ทธ๋ฆฌ๊ณ ์ง๋จ ๋๊ตฌ(diagnostic tool)์ ๊ฐ์ ๋ค์ํ ๊ธฐ์ฌ๋ฅผ ํ๊ณ ์์ต๋๋ค. ์ด๋ฌํ ๊ธฐ์ฌ๋ Arm AGI CPU์๋ง ๊ตญํ๋์ง ์๊ณ ์ ์ฒด ์ํ๊ณ(ecosystem)์ ์ ์ฉ๋ ๊ฒ์ ๋๋ค. ๊ทธ ๊ฒฐ๊ณผ, ๋ชจ๋ Arm ๊ธฐ๋ฐ ํ๋ซํผ(platform)์ ์ ์ตํ๊ฒ ํ์ฉ๋ ๊ฒ์ ๋๋ค. ์ ํฌ๊ฐ ๊ตฌ์ถํ๊ณ ์๋ ๊ฒ์ ์ง์ ํ ์ํ๊ณ์ด๊ธฐ ๋๋ฌธ์ ๋๋ค. Arm์ ํญ์ ์ํ๊ณ๋ฅผ ์ก์ฑํ๊ณ ํํธ๋์ญ(partnership)์ ๋งบ๋ ๋ฐ ์ฃผ๋ ฅํด์์ต๋๋ค. ์ด๋ ์ ํฌ์ ์ ์ฒด์ฑ(identity)์ ๋ ํต์ฌ์ ์ธ ๋ถ๋ถ์ด์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ๋ฌํ ๊ด๊ณ๋ค์ด ์ด์ ํฐ ๊ฒฐ์ค์ ๋งบ๊ณ ์์ต๋๋ค. Rene๊ฐ ๋ณด์ฌ๋๋ฆฐ ์์์ ๋ณด์ จ๊ฒ ์ง๋ง, ์ ํฌ๋ ์ด ๋ชจ๋ ํํธ๋์ญ์ ๋ํด ์ง์ฌ์ผ๋ก ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ฌ์ค, Arm AGI CPU๋ฅผ ๊ฐ๋ฐํ ์ ์์๋ ๊ฒ๋ ๋ฐ๋ก ์ด๋ฌํ ํํธ๋์ญ ๋๋ถ์ ๋๋ค. ์ด ์ค ์ผ๋ถ๋ ๋งค์ฐ ์ค๋ ๊ธฐ๊ฐ ์ง์๋์ด ์์ต๋๋ค. TSMC, ์ผ์ฑ(Samsung), ๋ง์ดํฌ๋ก (Micron), SKํ์ด๋์ค(SK Hynix)์ ๊ฐ์ ํํธ๋์ฌ๋ค์ ์ ํฌ๊ฐ ์์ญ ๋ , ๋ง ๊ทธ๋๋ก ์์ญ ๋ ๋์ ํจ๊ป ํ๋ ฅํด ์จ ์์คํ ๋๋ฐ์๋ค์ ๋๋ค. ์ ํฌ๋ ๋ํ ์๋ก์ด ํํธ๋์ญ์ ๊ตฌ์ถํ์ผ๋ฉฐ, ์ด๋ฅผ ํตํด Arm AGI CPU๊ฐ... (์, ์ฌ๋ผ์ด๋๊ฐ ๋๋ฌด ๋์ด๊ฐ๋ค์. ์ด์ ์ผ๋ก ๋์๊ฐ ์ฃผ์๊ฒ ์ด์?) ...์ด์ ์ถ์๋์ด ์ฌ์ฉ ๊ฐ๋ฅํ๋ค๊ณ ๋ง์๋๋ฆฌ๊ฒ ๋์ด ๋งค์ฐ ์๋์ค๋ฝ์ต๋๋ค. (์, ์ฌ๊ธฐ๋ ์๋ค์.) ๋ค, Arm AGI CPU๋ ์ด์ ์ฌ์ฉ ๊ฐ๋ฅํ๋ฉฐ, ์ ํฌ๋ ์ด์ ๋ํด ํฐ ์๋ถ์ฌ์ ๋๋๋๋ค. ์ค์ ๋ก ์ด๋ฏธ ๊ณ ๊ฐ๋ค์๊ฒ ์ ๋ฌ๋์ด ์์ต๋๋ค. ์ง๊ธ ์ด ์๊ฐ์๋ ๊ณ ๊ฐ๋ค์ด ์ง์ ํ๊ฐํ๊ณ ์์ผ๋ฉฐ, ์ ํฌ๋ ๋ชจ๋ ์ค๋น๋ฅผ ๋ง์ณค์ต๋๋ค. ๋ํ, ODM (Original Design Manufacturer) ํ๋ ฅ์ฌ, ๋ฉ๋ชจ๋ฆฌ, CPU, ๊ทธ๋ฆฌ๊ณ ์ ์กฐ ๋ถ์ผ์ ๋ชจ๋ ํํธ๋๋ถ๋ค๊ป ๊น์ด ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ด ๋จ๊ณ๊น์ง ์ฌ ์ ์๋๋ก ํฐ ๋์์ ์ฃผ์ จ์ต๋๋ค. ์ฐ๋ง๊น์ง๋ ์์ฐ (mass production)์ ๋์ ํ ์์ ์ด๋ฉฐ, ์ด ์์์ ์ฌ๋ฌ๋ถ๊ณผ ๊ณต์ ํ๊ฒ ๋์ด ๋งค์ฐ ๊ธฐ์ฉ๋๋ค. ์ค๋ ํ์ฌ, ํ์จ์ด(firmware)์ ์ฌ์(specification) ๋ชจ๋ ์ค๋น๋ฅผ ๋ง์ณค์ต๋๋ค. ํ๋ซํผ์ ๋ํด์๋ ์ด๋ฏธ ๋ง์๋๋ฆฐ ๋ฐ ์์ต๋๋ค. ๊ณต๊ธ(supply)์ ๋ํด์๋ ๋ง์๋๋ ธ์ต๋๋ค. ์์ง ๋ง์๋๋ฆฌ์ง ์์ ํ ๊ฐ์ง๋ ๋ฐ๋ก ์ํํธ์จ์ด(software)์ ๋๋ค. ๊ทธ๋ผ ์ํํธ์จ์ด์ ๋ํด ์ด์ผ๊ธฐํด ๋ณผ๊น์. ๋ค์ ์ฌ๋ผ์ด๋์ ๋๋ค. ๋ค. ์ฌ์ค Arm์ ๋ฐ์ดํฐ์ผํฐ(data center) ์ํํธ์จ์ด ์ํ๊ณ(ecosystem)์ 15๋ ์ด์ ํฌ์ํด ์์ต๋๋ค. ์ ํฌ๊ฐ ์ํํธ์จ์ด ์ํ๊ณ์ ์ผ๋ง๋ ์ค๋ซ๋์ ํฌ์ํด์๋์ง ๋ชจ๋ ๋ถ๋ค์ด ์์๋์ง๋ ๋ชจ๋ฅด๊ฒ ์ต๋๋ค๋ง, ์ด๊ธฐ์๋ Arm๋ง์ด ์ํํธ์จ์ด ์ํ๊ณ์ ํฌ์ํ๊ณ ์์์ต๋๋ค. ๊ทธ๋ฌ๋ค 2019๋ ์ ๋ณํ๊ฐ ์๊ฒผ์ต๋๋ค. ์ ํฌ๋ Arm Neoverse๋ฅผ ์ถ์ํ์ต๋๋ค. ์ ํฌ๊ฐ Arm Neoverse๋ผ๋ ์ปดํจํ ํ๋ซํผ(compute platform)์ ์ถ์ํ์ ๋, ์ด๋ ๊ณ ๊ฐ๋ค์ด ํจ์ฌ ๋ฎ์ ์ง์ ์ฅ๋ฒฝ(barrier to entry)์ผ๋ก ์ ํ์ ์ถ์ํ ์ ์๋๋ก ํด์ฃผ์์ต๋๋ค. ๊ณ ๊ฐ๋ค์ ์ด๋ฅผ ํตํด ์์ฒด ์ค๋ฆฌ์ฝ(silicon)์ ๊ตฌ์ถํ๊ณ ๊ณตํต ํ๋ซํผ(common platform)์ ์ค์ฌ์ผ๋ก ๊ฒฐ์งํ๊ธฐ ์์ํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ๊ฒ์ด ๋ฐ๋ก ์ํํธ์จ์ด ํ๋ผ์ดํ (software flywheel)์ด ๋์๊ฐ๊ธฐ ์์ํ ๊ณ๊ธฐ์์ต๋๋ค. ์์๋ค์ํผ, ๊ธฐ์ ์ ๋ ๊ธฐ์ ๋ค์ด Neoverse๋ฅผ ์ฑํํ๊ธฐ ์์ํ๋ฉด์, ๊ทธ๋ค์ Neoverse์ ๋ง์ถฐ ์ํํธ์จ์ด๋ฅผ ์ต์ ํํ๊ธฐ ์์ํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ Neoverse๋ฅผ ์ฑํํ๋ ๊ธฐ์ ์ ๋ ๊ธฐ์ ๋ค์ด ๋์ด๋ ์๋ก, ๊ทธ ํ๋ผ์ดํ ์ ๋ ๋น ๋ฅด๊ฒ ํ์ ํ๊ธฐ ์์ํ์ต๋๋ค. ์ค๋๋ AWS, ๊ตฌ๊ธ, ๋ฉํ, ๋ง์ดํฌ๋ก์ํํธ, ์ค๋ผํด, ์๋น๋์์ ๊ฐ์ ๊ธฐ์ ๋ค์ด ๋ชจ๋ ์ ํฌ์ ํจ๊ป ์ํํธ์จ์ด ์ํ๊ณ(software ecosystem)์ ํฌ์ํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ๊ฒ์ด ๋ฐ๋ก ์ ํฌ๊ฐ ์ํํธ์จ์ด ๋ถ์ผ์์ ์๋นํ ์ฑ๊ณผ(traction)๋ฅผ ๊ฑฐ๋ ์ ์์๋ ์๋๋ ฅ์ด์์ต๋๋ค. ์ ํฌ๋ ํจ๊ป Arm์ ๋๋ถ๋ถ์ ์ต์ ์ํํธ์จ์ด ํจํค์ง(software package)์์ ์ผ๋ฑ ์๋ฏผ(first-class citizen)์ผ๋ก ๋ง๋ค์์ต๋๋ค. ํนํ ์ ํฌ์ AI ์ํํธ์จ์ด ์ํ๊ณ(AI software ecosystem)์ ๊ฒฝ์ฐ, ์ ํฌ๋ ์ผ๋ฑ ์๋ฏผ์ผ ๋ฟ๋ง ์๋๋ผ, ์ํํธ์จ์ด๊ฐ Arm์์ ์ ์๋ํ๋ ๊ฒ์ ๋์ด, ์ํํธ์จ์ด๊ฐ Arm์์ ๊ฐ์ฅ ์ ์๋ํฉ๋๋ค. ๊ทธ ์ด์ ๋ ๋งค์ฐ ๊ฐ๋จํฉ๋๋ค. AI ๋ถ์ผ์์ Arm ์ํํธ์จ์ด ์ํ๊ณ์ Arm ์ํคํ ์ฒ๋ ์ค๋๋ AI๋ฅผ ์ง์ํ๋ ์ฃผ์ CPU ์ํคํ ์ฒ(CPU architecture)์ ๋๋ค. ์ค์ ๋ก ์ ํฌ๊ฐ ๊ธฐ์ ์ ๋ ๊ธฐ์ ๋ค๊ณผ ํจ๊ป ํด์จ ๋ ธ๋ ฅ ๋๋ถ์, ์ค๋๋ ์๋ง ๊ฐ์ ๊ธฐ์ ๋ค์ด ํด๋ผ์ฐ๋(cloud)์์ Arm ๊ธฐ๋ฐ์ผ๋ก ์ํํธ์จ์ด(software)๋ฅผ ์ด์ํ๊ณ ์์ต๋๋ค. ์ด๋ ์ ํฌ๊ฐ ์ด๋ฏธ ์ ์ธ๊ณ ๋ฐ์ดํฐ์ผํฐ(data center)์ ๊ณต๊ธํ 12์ต 5์ฒ๋ง ๊ฐ ์ด์์ Arm Neoverse ์ฝ์ด(core)์์ ๊ตฌ๋๋ฉ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๋ฌํ ์ฑ์ฅ์ ๋์ฑ ๊ฐ์ํ๋๊ณ ์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ๊ทธ ์ฑ์ฅ ๊ณก์ ์ ๋๋ค. ๋ณด์๋ค์ํผ, ๋ฐ์ดํฐ์ผํฐ์์ Arm์ ๊ทธ์ ์ ์๋ํฉ๋๋ค. ์ด ์ ์ด ํต์ฌ์ ๋๋ค. ์ ๊ฐ ์ด ์ ์ ์ถฉ๋ถํ ์ ์ค๋ช ํ๋์ง ๋ชจ๋ฅด๊ฒ ์ต๋๋ค. ๊ทธ๋์ ์ํํธ์จ์ด ๋ถ์ผ์์ ๊ฒฝํ์ด ํ๋ถํ ๋ถ์ ๋ฌด๋๋ก ๋ชจ์๊ฒ ์ต๋๋ค. ํด ์ฌ๋ธ(Paul Saab)๋ ๋ฉํ(Meta)์ ์ธํ๋ผ(infrastructure) ๋ถ์ผ์์ 18๋ ์ด์ ๊ทผ๋ฌดํ์ผ๋ฉฐ, ํ์ฌ์์ ๊ฐ์ฅ ์ค๋ ๊ทผ๋ฌดํ ์ง์ ์ค ํ ๋ช ์ ๋๋ค. ๊ทธ๋ถ์ ํ๋์ ์คํ ๋ฆฌ์ง(flash storage) ๋์ ๋ถํฐ IPV6 ๊ตฌํ์ ์ด๋ฅด๊ธฐ๊น์ง ์๋ง์ ์ ๋ฌด๋ฅผ ์ฑ ์์ ธ ์ค์ จ์ต๋๋ค. ํ์ฌ๋ ํนํ ์ธํ๋ผ(infrastructure) ๋ด์์ AI์ ํจ์จ์ฑ์ ๋์ด๋ ๋ฐ ์ฃผ๋ ฅํ๊ณ ๊ณ์ญ๋๋ค. ๊ทธ ์ธ์ฐ์ผ๋ก ์ ํฌ๋ ์๋ก๋ฅผ ์๊ฒ ๋์๊ณ , ์ด์ ํด ์ฌ๋ธ ๋์ ๋ชจ์๊ฒ ์ต๋๋ค. |
| Paul Saab: Thank you. | **Paul Saab:** ๊ฐ์ฌํฉ๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit Thank you. Thanks for being here. Thanks for being here. | **Mohamed Awad:** ๊ฐ์ฌํฉ๋๋ค. ์ค๋ ์ด ์๋ฆฌ์ ์ฐธ์ํด ์ฃผ์ ์ ๋๋จํ ๊ฐ์ฌํฉ๋๋ค. |
| Paul Saab: Thank you for having me. | **Paul Saab:** ์ด๋ํด ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit You've told me the story before, but I really want to hear you guys have had a long history with Arm. -- goes back longer than just a couple of years ago. Can you maybe give everybody a little bit of a history lesson as to kind of how things started? | **Mohamed Awad:** ์ด์ ์ ๋ง์ํด์ฃผ์ ์ ์ด ์์ง๋ง, Arm(์)๊ณผ์ ์ค๋ ์ธ์ฐ์ ๋ํด ์ ๋ง ๋ฃ๊ณ ์ถ์ต๋๋ค. ๋จ์ํ ๋ช ๋ ์ ์ด ์๋๋ผ ํจ์ฌ ๋ ์ค๋ ์ ๋ถํฐ ํจ๊ป ํด์ค์ ๊ฒ์ผ๋ก ์๊ณ ์์ต๋๋ค. ์ด๋ป๊ฒ ์์๋์๋์ง์ ๋ํ ๊ฐ๋ตํ ์ญ์ฌ์ ๋ฐฐ๊ฒฝ์ ๋ชจ๋ ๋ถ๋ค๊ป ์ค๋ช ํด์ฃผ์ค ์ ์์๊น์? |
| Paul Saab: Yes. I think it was like 2014, 2015, we were looking at Arm. We were really excited about the efficiency wins that we were seeing. We were really back then just targeting our hack/PHP platform called HBM. And it was working great. Like we made it work. It was performant and then the market kind of went away for us. We didn't really have a platform anymore. And so we just sort of tabled it. And we ripped all that code out. Everything in the code base was removed. | **Paul Saab:** ๋ค, 2014๋
, 2015๋
์ฏค์ด์๋ ๊ฒ ๊ฐ์์. ์ ํฌ๊ฐ Arm์ ๊ฒํ ํ๊ณ ์์์ ๋์์ฃ . ์ ํฌ๋ ์ ํฌ๊ฐ ํ์ธํ ํจ์จ์ฑ ๊ฐ์ (efficiency wins)์ ๋ํด ์ ๋ง ๊ณ ๋ฌด์ ์ด์์ต๋๋ค. ๋น์ ์ ํฌ๋ ์ฃผ๋ก HBM์ด๋ผ๋ ์ ํฌ์ hack/PHP ํ๋ซํผ์ ๋ชฉํ๋ก ์ผ๊ณ ์์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ทธ ํ๋ซํผ์ ์์ฃผ ์ ์๋ํ์ต๋๋ค. ์ ํฌ๊ฐ ์ ๋๋ก ๊ตฌํํด๋๊ณ , ์ฑ๋ฅ๋ ๋ฐ์ด๋ฌ์ฃ . ๊ทธ๋ฐ๋ฐ ์ ํฌ์๊ฒ๋ ์์ฅ ๊ธฐํ๊ฐ ์ฌ๋ผ์ ธ ๋ฒ๋ ธ์ต๋๋ค. ๋ ์ด์ ํ๋ซํผ์ ์ ์งํ ํ์๊ฐ ์๊ฒ ๋ ๊ฑฐ์ฃ . ๊ทธ๋์ ์ ํฌ๋ ๊ทธ ๊ณํ์ ๋ณด๋ฅํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ๊ด๋ จ ์ฝ๋๋ฅผ ๋ชจ๋ ์ ๊ฑฐํ์ต๋๋ค. ์ฝ๋๋ฒ ์ด์ค(code base)์ ์๋ ๋ชจ๋ ๊ฒ์ด ์ญ์ ๋์์ฃ . |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit Okay. So that was 2014 and 2015. Obviously, something must have changed or you wouldn't be standing here today, right? So kind of where do we go from there? | **Mohamed Awad:** ์ข์ต๋๋ค. 2014๋ ๊ณผ 2015๋ ์ ๊ทธ๋ฌ์ต๋๋ค. ๋ถ๋ช ํ ๋ญ๊ฐ ๋ฐ๋์์ ๊ฒ๋๋ค. ๊ทธ๋ ์ง ์์๋ค๋ฉด ์ค๋ ์ด ์๋ฆฌ์ ๊ณ์์ง ์์๊ฒ ์ฃ ? ๊ทธ๋ผ ๊ทธ ๋ค์์ ์ด๋ป๊ฒ ๋๋ ๊ฑด๊ฐ์? |
| Paul Saab: Well, the story is kind of funny is like we were -- it was like post -- we're just coming out of the COVID bubble, and we had a bunch of people over at the house sitting around socializing and whatever. And I turned to one of my colleagues and I said, hey, I want to build -- I want to port the Arm again. I kind of had this gut feeling that the ecosystem in the world had changed. And if we didn't start then we would be kind of playing catch-up when it actually happened. I didn't even ask my boss here for permission to buy these machines or even to start the project. | **Paul Saab:** ์ด์ผ๊ธฐ๊ฐ ์ข ์ฌ๋ฏธ์๋๋ฐ ๋ง์ด์ฃ , ์ ํฌ๊ฐ ๋ง ์ฝ๋ก๋19 ํฌ๋ฐ๋ฏน(COVID-19 pandemic) ์ํฉ์์ ๋ฒ์ด๋๊ณ ์์ ๋์์ด์. ์ง์ ์ฌ๋ฌ ์ฌ๋์ด ๋ชจ์ฌ์ ๋ด์๋ฅผ ๋๋๊ณ ์์์ฃ . ๊ทธ๋ ๋๋ฃ ์ค ํ ๋ช
์๊ฒ ์ ๊ฐ ์ด๋ ๊ฒ ๋งํ์ต๋๋ค. "์ผ, ์ฐ๋ฆฌ Arm์ ๋ค์ ํฌํ
(porting)ํด์ผ๊ฒ ์ด." ์ ์ง ๋ชจ๋ฅด๊ฒ ์ง๊ฐ์ ์ผ๋ก ์ธ์์ ์ํ๊ณ(ecosystem)๊ฐ ๋ณํ๋ค๋ ๊ฒ์ ๋๊ผ์ต๋๋ค. ๋ง์ฝ ๊ทธ๋ ์์ํ์ง ์์ผ๋ฉด ์ค์ ๋ก ์ํฉ์ด ๋ฅ์ณค์ ๋ ๋ฐ๋ผ์ก๊ธฐ ๊ธ๊ธํ ๊ฒ์ด๋ผ๊ณ ์. ์ ๋ ์ฌ์ง์ด ์์ฌ์๊ฒ ์ด ์ฅ๋น๋ค์ ๊ตฌ๋งคํ๊ฑฐ๋ ํ๋ก์ ํธ๋ฅผ ์์ํ๋ ๊ฒ์ ๋ํด ํ๋ฝ๋ ๋ฐ์ง ์์์ต๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit It's a good thing he approves now. | **Mohamed Awad:** ํด๋ผ์ฐ๋ AI ์ฌ์ ๋ถ ์ ๋ฌด๋๊ป์ ์ด์ ์น์ธํด์ฃผ์ ์ ๋คํ์ ๋๋ค. |
| Paul Saab: I don't really ask him permission for much to do. But yes, so we started -- we decided we found some machines out there. I went to some other colleagues. I said, hey, I want to port to Arm and he actually responded was like I was wondering when you were going to ask me. So we got the machines in, started porting, making great progress, but it was super slow. We only had 8 machines. But we had this vast x86 ecosystem. And I went to the guys and I was like, hey, can we cross compile? And that's what we ended up doing. We ended up like working around the clock. It took us about 90 days, 5 engineers, and we had a full complete port, full system ready, but then we ran into another problem. We had no silicon to buy. And this -- and Santosh referenced this that like we looked at every partner. And I think this is about the time you and I started talking. | **Paul Saab:** ๊ทธ์๊ฒ ๋ฑํ ํ๋ฝ์ ๋ฐ์์ผ ํ ์ผ์ ๋ง์ง ์์ต๋๋ค๋ง, ๋ค, ๊ทธ๋์ ์ ํฌ๋ ๋ช๋ช ์ฅ๋น๋ฅผ ์ฐพ์๋ด๊ธฐ๋ก ๊ฒฐ์ ํ๊ณ ์์
์ ์์ํ์ต๋๋ค. ๋ค๋ฅธ ๋๋ฃ๋ค์๊ฒ ๊ฐ์ Arm์ผ๋ก ํฌํ
(porting)ํ๊ณ ์ถ๋ค๊ณ ๋งํ๋๋, ๊ทธ ๋๋ฃ๋ '์ธ์ ์ฏค ๋ฌผ์ด๋ณผ๊น ํ๋ค'๋ ๋ฐ์์ ๋ณด์ด๋๊ตฐ์. ๊ทธ๋์ ์ฅ๋น๋ฅผ ๋ค์ฌ์ ํฌํ ์์ ์ ์์ํ๊ณ , ์๋นํ ์ง์ ์ ๋ณด์์ง๋ง ์๋๊ฐ ๋๋ฌด ๋๋ ธ์ต๋๋ค. ์ฅ๋น๋ 8๋๋ฐ์ ์์๊ฑฐ๋ ์. ํ์ง๋ง ์ ํฌ์๊ฒ๋ ๋ฐฉ๋ํ x86 ์ํ๊ณ(ecosystem)๊ฐ ์์์ต๋๋ค. ๊ทธ๋์ ๋ด๋น์๋ค์๊ฒ ๊ฐ์ 'ํฌ๋ก์ค ์ปดํ์ผ(cross compile)์ด ๊ฐ๋ฅํ ๊น์?'๋ผ๊ณ ๋ฌผ์๊ณ , ๊ฒฐ๊ตญ ์ ํฌ๋ ๊ทธ๋ ๊ฒ ํ์ต๋๋ค. ๋ฐค๋ฎ์์ด ์ผํ๊ฒ ๋์์ฃ . ์ฝ 90์ผ์ด ๊ฑธ๋ ธ๊ณ , 5๋ช ์ ์์ง๋์ด๊ฐ ํฌ์ ๋์ด ์๋ฒฝํ ํฌํ ๊ณผ ์ ์ฒด ์์คํ ์ ์ค๋นํ์ง๋ง, ๋ ๋ค๋ฅธ ๋ฌธ์ ์ ๋ถ๋ชํ์ต๋๋ค. ๊ตฌ๋งคํ ์ค๋ฆฌ์ฝ(silicon)์ด ์์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด ๋ถ๋ถ์ Santosh๋ ์ธ๊ธํ๋ฏ์ด, ์ ํฌ๋ ๋ชจ๋ ํํธ๋๋ฅผ ๊ฒํ ํ์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ์๊ฐ์๋ ์ด๋์ฏค ์ ํฌ๊ฐ ๋ํ๋ฅผ ์์ํ๋ ๊ฒ ๊ฐ์ต๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit So you'd say the market was a little bit underserved maybe for what you guys were. | **Mohamed Awad:** ๊ทธ๋ผ, ๊ท์ฌ์ ์๋น์ค/์ ํ์ ๋ํด ์์ฅ์ด ๋ค์ ๋ฏธ์ถฉ์กฑ(underserved) ์ํ์๋ค๊ณ ๋ณด์๋ ๊ฑด๊ฐ์? |
| Paul Saab: I think underserved is an understatement. | **Paul Saab:** underserved (์์๊ฐ ์ถฉ์กฑ๋์ง ๋ชปํ) ์์ฅ์ด๋ผ๊ณ ๋งํ๊ธฐ์๋ ๊ทธ ์ํฉ์ ๋ค ๋ด์๋ด์ง ๋ชปํ๋ค๊ณ ๋ด ๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit But let's go back to the 90 days, 5 people. I mean, really -- it's okay. So I'm going to take your word for it. It was 90 days, 5 people, but that's just getting the code working. Like now you've got to now operationalize and get it performing. Like how is that going? | **Mohamed Awad:** ํ์ง๋ง 90์ผ, 5๋ช ์ด๋ผ๋ ๊ทธ ๋ถ๋ถ์ผ๋ก ๋ค์ ๋์๊ฐ ๋ณด์ฃ . ์, ์ข์ต๋๋ค. ๋ง์ํ์ ๋๋ก ๋ฐ์๋ค์ด๊ฒ ์ต๋๋ค. 90์ผ, 5๋ช ์ผ๋ก ์งํ๋์๋ค๊ณ ํ์ จ๋๋ฐ, ๊ทธ๊ฑด ๋จ์ํ ์ฝ๋๋ฅผ ์๋์ํค๋ ๋ฐ ๊ทธ์ณค์ ๋ฟ์ ๋๋ค. ์ด์ ๋ ๊ทธ๊ฑธ ์ด์ํ(operationalize)ํ๊ณ ์ ๋๋ก ์ฑ๋ฅ์ ๋ด๋๋ก ํด์ผ ํฉ๋๋ค. ๊ทธ ๋ถ๋ถ์ ์ด๋ป๊ฒ ์งํ๋๊ณ ์์ต๋๊น? |
| Paul Saab: It's still a small team. I mean it's a lot of very devoted people bringing the systems up. From the time we finished that initial port in 2022, it took us about 2.5 years to actually get some sort of production worthy systems in that were TCO effective, performance per watt. And it was still a very small team. And even today, it's really a small team that's focused on hyper optimizing. It started off with -- once those performance systems landed, it was really just one engineer until a few more came in. But that engineer never had written a single line of Neon, never written a single line of SVE and single-handedly took some of our most precious workloads and made them work on Arm. | **Paul Saab:** ์ฌ์ ํ ์๊ท๋ชจ ํ์
๋๋ค. ์์คํ
์ ๊ตฌ์ถํ๋ ๋ฐ ๋ง์ ํ์ ์ ์ธ ์ธ๋ ฅ๋ค์ด ์ฐธ์ฌํ๊ณ ์์ต๋๋ค. 2022๋ ์ ์ด๊ธฐ ํฌํ (porting)์ ์๋ฃํ ์์ ๋ถํฐ, ์ด์์ ๋น์ฉ(TCO: Total Cost of Ownership) ํจ์จ์ ์ด๊ณ ์ํธ๋น ์ฑ๋ฅ(performance per watt)์ด ๋ฐ์ด๋ ์์ฐ ์์ค์ ์์คํ ์ ์ค์ ๋ก ๊ตฌ์ถํ๊ธฐ๊น์ง ์ฝ 2.5๋ ์ด ๊ฑธ๋ ธ์ต๋๋ค. ๊ทธ๋ผ์๋ ๋ถ๊ตฌํ๊ณ ์ฌ์ ํ ๋งค์ฐ ์์ ํ์ด์์ต๋๋ค. ์ค๋๋ ์๋ ์ด๊ณ ๋ ์ต์ ํ(hyper optimizing)์ ์ง์คํ๋ ํ์ ์ฌ์ ํ ์๊ท๋ชจ์ ๋๋ค. ์ฒ์์๋ ๊ณ ์ฑ๋ฅ ์์คํ ์ด ๋์ ๋ ํ ๋จ ํ ๋ช ์ ์์ง๋์ด๋ก ์์ํ๊ณ , ๊ทธ ํ์ ๋ช ๋ช ์ด ๋ ํฉ๋ฅํ์ต๋๋ค. ํ์ง๋ง ๊ทธ ์์ง๋์ด๋ Neon ์ฝ๋ ํ ์ค, SVE ์ฝ๋ ํ ์ค๋ ์์ฑํด ๋ณธ ์ ์ด ์์์์๋ ๋ถ๊ตฌํ๊ณ , ์ ํฌ์ ๊ฐ์ฅ ์ค์ํ ์ํฌ๋ก๋(workload) ์ค ์ผ๋ถ๋ฅผ ํ๋ก ๋ด๋นํ์ฌ Arm์์ ์๋ํ๋๋ก ๋ง๋ค์์ต๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit And how is it performing now generally, like on typical workloads? Like how should we think about the performance in the general? | **Mohamed Awad:** ๊ทธ๋ ๋ค๋ฉด ํ์ฌ ์ ๋ฐ์ ์ผ๋ก ์ด๋ค ์ฑ๊ณผ๋ฅผ ๋ณด์ด๊ณ ์์ต๋๊น? ์๋ฅผ ๋ค์ด, ์ผ๋ฐ์ ์ธ ์ํฌ๋ก๋(workload)์์๋ ์ด๋ค ์ฑ๋ฅ์ ๋ณด์ด๊ณ ์๋์? ์ ๋ฐ์ ์ธ ์ฑ๋ฅ์ ๋ํด ์ ํฌ๊ฐ ์ด๋ป๊ฒ ์ดํดํ๋ฉด ์ข์๊น์? |
| Paul Saab: We're seeing performance that is basically equal to anything you can buy in the market today at massive performance per watt improvements. | **Paul Saab:** ์ ํฌ๋ ํ์ฌ ์์ฅ์์ ๊ตฌ๋งคํ ์ ์๋ ์ด๋ค ์ ํ๊ณผ๋ ๊ธฐ๋ณธ์ ์ผ๋ก ๋๋ฑํ ์ฑ๋ฅ์ ๋ณด์ด๋ฉด์๋, ์ํธ๋น ์ฑ๋ฅ (performance per watt) ๋ฉด์์๋ ์์ฒญ๋ ๊ฐ์ ์ ์ด๋ฃจ๊ณ ์์ต๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit That's great. That's great. Okay. My light is going to start blinking in a minute here. So I'm not going to keep you on stage too long. But I just -- first of all, I want to say thank you. But before I let you go, I guess one question for you. If somebody is out there thinking about, hey, because there are tens of thousand companies that are using Arm already, but there's still a few that aren't. What sort of advice or guidance would you get? And what would be your kind of recommendation to them? | **Mohamed Awad:** ๋ค, ์์ฃผ ์ข์ ๋ง์์ด์ญ๋๋ค. ์ข์ต๋๋ค. ์ ์๊ฐ์ด ์ผ๋ง ๋จ์ง ์์ ๊ณง ๋ถ์ด ๊น๋นก์ผ ๊ฒ ๊ฐ์ต๋๋ค. ๊ทธ๋์ ์ค๋ ๋ถ์ก์ ๋์ง๋ ์๊ฒ ์ต๋๋ค. ํ์ง๋ง ์ฐ์ ๊ฐ์ฌ์ ๋ง์์ ๋๋ฆฌ๊ณ ์ถ์ต๋๋ค. ๊ฐ์๊ธฐ ์ ์, ํ ๊ฐ์ง ์ง๋ฌธ์ ๋๋ฆฌ๊ณ ์ถ์ต๋๋ค. ์ด๋ฏธ ์๋ง ๊ฐ์ ๊ธฐ์ ์ด Arm์ ์ฌ์ฉํ๊ณ ์์ง๋ง, ์์ง ์ฌ์ฉํ์ง ์๋ ๊ธฐ์ ๋ค๋ ๋ถ๋ช ํ ์์ต๋๋ค. ๋ง์ฝ ๊ทธ๋ฐ ๊ธฐ์ ๋ค์ด Arm ๋์ ์ ๊ณ ๋ฏผํ๊ณ ์๋ค๋ฉด, ์ด๋ค ์กฐ์ธ์ด๋ ์ง์นจ์ ์ฃผ์๊ฒ ์ต๋๊น? ๊ทธ๋ค์๊ฒ ์ด๋ค ๊ถ๊ณ ๋ฅผ ํ์๊ฒ ์ต๋๊น? |
| Paul Saab: I think small focus teams doing the port. But like if I were starting the port today, I would be using an LLM. I mean what we're -- what I'm seeing some of the engineers that are now optimizing even existing Arm accelerated code, they're using LLMs to even boost those by 10% or 20%. So the barrier to entry today like porting to Arm is, I would say, close to 0. because like the LLM is just going to do it for you. I don't even write any handwritten code anymore myself. It's just all LLM, all test cases all across the board. So like there's no excuse to port to Arm today. | **Paul Saab:** ์๊ท๋ชจ ์ ๋ดํ์ด ํฌํ (porting) ์์ ์ ์ํํ๊ณ ์๋ค๊ณ ์๊ฐํฉ๋๋ค. ํ์ง๋ง ๋ง์ฝ ์ ๊ฐ ์ค๋ ํฌํ ์์ ์ ์์ํ๋ค๋ฉด, LLM (๊ฑฐ๋ ์ธ์ด ๋ชจ๋ธ)์ ์ฌ์ฉํ ๊ฒ๋๋ค. ์ ๊ฐ ๋ณด๊ธฐ์๋ ์ผ๋ถ ์์ง๋์ด๋ค์ด ํ์ฌ ๊ธฐ์กด์ Arm ๊ฐ์ํ(accelerated) ์ฝ๋๊น์ง ์ต์ ํ(optimizing)ํ๊ณ ์๋๋ฐ, ์ด๋ค์ LLM์ ํ์ฉํด์ ์ฑ๋ฅ์ 10% ๋๋ 20%๊น์ง ํฅ์์ํค๊ณ ์์ต๋๋ค. ๊ทธ๋์ ์ค๋๋ Arm์ผ๋ก์ ํฌํ ์ง์ ์ฅ๋ฒฝ(barrier to entry)์ ๊ฑฐ์ 0์ ๊ฐ๊น๋ค๊ณ ํ ์ ์์ต๋๋ค. ์๋ํ๋ฉด LLM์ด ๊ทธ ์์ ์ ์์์ ๋ค ํด์ฃผ๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ ๋ ๋ ์ด์ ์ง์ ์์ผ๋ก ์ฝ๋๋ฅผ ์์ฑํ์ง ์์ต๋๋ค. ๋ชจ๋ ๊ฒ์ด LLM์ผ๋ก ์ฒ๋ฆฌ๋๊ณ , ๋ชจ๋ ํ ์คํธ ์ผ์ด์ค(test cases)๋ ์ ๋ฐ์ ์ผ๋ก ๋ค LLM์ ํตํด ์ด๋ฃจ์ด์ง๋๋ค. ๋ฐ๋ผ์ ์ค๋๋ Arm์ผ๋ก ํฌํ ํ๋ ๊ฒ์ ๋ํด ๋ณ๋ช ์ ์ฌ์ง๊ฐ ์์ต๋๋ค. |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit Excellent. Thanks, Well, that was inspiring. I mean, Paul and I have obviously known each other for a little while. And the tenacity and what I hear around is once Paul gets something in his mind, it just kind of happens. So I appreciate all the support, Paul. Thank you. We're so part of the partnership that we've had with you and with Meta more broadly. So thank you very much. What I love about that story is that they had a need, the market was underserved. And together, we worked together to go address it. The reality is the opportunity for the AGI CPU is broad. The software is ready, and we have a great product. And that's why we're seeing such great customer traction. We're seeing it in multiple areas. If you think about companies like Cerebras and Positron and Rebellions, they're joining Meta and OpenAI by using Arm AGI CPU for things like managing head nodes that they're building or managing accelerators they're building, so a head node type use case or also for agentic orchestration and fan-out. These are specific use cases that they're looking at. And then in the cloud, we see companies like SAP and SK Telecom and Cloudflare who are actively using or planning on deploying Arm as part of their infrastructure. These are just a few of the customers that are planning on using Arm AGI CPU. But rather than me tell you let's listen to them. [Presentation] | **Mohamed Awad:** ํ๋ฅญํฉ๋๋ค. ๊ฐ์ฌํฉ๋๋ค. ์ ๋ง ๊ณ ๋ฌด์ ์ด๋ค์. ํด๊ณผ ์ ๋ ๋ฌผ๋ก ์ค๋ซ๋์ ์๊ณ ์ง๋์ต๋๋ค. ์ ๊ฐ ๋ฃ๊ธฐ๋ก๋, ํด์ด ํ๋ฒ ๋ง์๋จน์ผ๋ฉด ์ด๋ค ์ผ์ด๋ ๋๊ธฐ ์๊ฒ ํด๋ธ๋ค๊ณ ํฉ๋๋ค. ํด, ๋ชจ๋ ์ง์์ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ ํฌ๋ ํด๊ณผ, ๊ทธ๋ฆฌ๊ณ ๋ ๋์๊ฐ ๋ฉํ(Meta)์ ๋งบ์ ํํธ๋์ญ(partnership)์ ์ค์ํ ์ผ์์
๋๋ค. ๋ค์ ํ๋ฒ ์ง์ฌ์ผ๋ก ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ๊ทธ ์ด์ผ๊ธฐ์์ ์ ๊ฐ ๊ฐ์ฅ ์ข์๋ ์ ์, ๊ทธ๋ค์๊ฒ๋ ํ์๊ฐ ์์๊ณ ์์ฅ์ ์ ๋๋ก ์ถฉ์กฑ๋์ง ๋ชปํ๋ค๋ ๊ฒ์
๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ ํจ๊ป ํ๋ ฅํ์ฌ ๊ทธ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ์ต๋๋ค. ํ์ค์ ์ผ๋ก AGI CPU์ ๊ธฐํ๋ ๋งค์ฐ ๊ด๋ฒ์ํฉ๋๋ค. ์ํํธ์จ์ด(software)๋ ์ค๋น๋์๊ณ , ์ ํฌ๋ ํ๋ฅญํ ์ ํ์ ๊ฐ์ง๊ณ ์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ์ ํฌ๊ฐ ์ด๋ ๊ฒ ํ๋ฅญํ ๊ณ ๊ฐ ์ ์น(customer traction)๋ฅผ ๋ณด๊ณ ์๋ ์ด์ ์
๋๋ค. ์ฌ๋ฌ ๋ถ์ผ์์ ์ด๋ฌํ ํ์์ ๋ชฉ๊ฒฉํ๊ณ ์์ต๋๋ค. Cerebras, Positron, Rebellions ๊ฐ์ ๊ธฐ์
๋ค์ ์ดํด๋ณด๋ฉด, ์ด๋ค์ Meta์ OpenAI์ฒ๋ผ Arm AGI CPU๋ฅผ ํ์ฉํ์ฌ ์์ฒด์ ์ผ๋ก ๊ตฌ์ถ ์ค์ธ ํค๋ ๋
ธ๋(head node)๋ ๊ฐ์๊ธฐ(accelerator)๋ฅผ ๊ด๋ฆฌํ๋ ๋ฐ ์ฌ์ฉํ๊ณ ์์ต๋๋ค. ์ฆ, ํค๋ ๋
ธ๋ ์ ํ์ ์ฌ์ฉ ์ฌ๋ก(use case)๋ ์์ด์ ํธ ์ค์ผ์คํธ๋ ์ด์
(agentic orchestration) ๋ฐ ํฌ์์(fan-out) ๋ฑ์ ํ์ฉํ๊ณ ์๋ ๊ฒ์ด์ฃ . ์ด๋ค์ด ํ์ฌ ๊ฒํ ํ๊ณ ์๋ ๊ตฌ์ฒด์ ์ธ ์ฌ์ฉ ์ฌ๋ก๋ค์
๋๋ค. ํด๋ผ์ฐ๋(cloud) ๋ถ์ผ์์๋ SAP, SKํ ๋ ์ฝค, Cloudflare์ ๊ฐ์ ๊ธฐ์ ๋ค์ด Arm์ ์์ฌ ์ธํ๋ผ(infrastructure)์ ์ผ๋ถ๋ก ์ ๊ทน์ ์ผ๋ก ์ฌ์ฉํ๊ฑฐ๋ ๋ฐฐํฌํ ๊ณํ์ ๊ฐ์ง๊ณ ์์ต๋๋ค. ์ด๋ค์ Arm AGI CPU๋ฅผ ๋์ ํ ๊ณํ์ ๊ฐ์ง ์ฌ๋ฌ ๊ณ ๊ฐ์ฌ๋ค ์ค ์ผ๋ถ์ ๋ถ๊ณผํฉ๋๋ค. ์ ๊ฐ ์ง์ ์ค๋ช ๋๋ฆฌ๋ ๊ฒ๋ณด๋ค, ๊ทธ๋ค์ ์ด์ผ๊ธฐ๋ฅผ ์ง์ ๋ค์ด๋ณด์์ฃ . |
| Mohamed Awad: Executive Vice President of Cloud AI Business Unit I just want to say thanks again to all of our customers and some partners that are supporting us here today. The support we've gotten has really just been incredible. We built Arm AGI CPU for you, and we're so pleased with the response. You see Arm AGI CPU has been designed from the ground up to make sure that performance scales and power stays predictable. That's the superpower, performance, scale and efficiency. And it's resonating with our partners. You see that's a very different approach than it's taken by x86. They've burdened their execution -- they are burdened with execution overhead and legacy feature support. They chose to focus on things like modularity, support for lots of different markets and esoteric use cases. We are ruthlessly focused on improving efficiency and reducing latency. Ultimately, this is about architectural philosophy. We're not strapped to the past. We are not strapped to the past. Listen, we don't support Lotus Notes, okay? We just don't do it. We're focused on exactly and only what the AGI at data center needs, performance, scale and efficiency. Let me take you through that in a little more detail. It starts with performance. And performance for us is all about doing more work for every clock cycle. This has always been an area. Great IPC has always been an area where Arm has shined. How much work do you get done every single cycle? Our AGI CPU absolutely shines here. Now what we see is that legacy CPUs, they sometimes try to compete on this vector by doing things like increasing the frequency, going to boost modes. But here's the reality. When you increase the frequency, what else do you increase? Power. That's a problem. These boost modes are not sustainable across long periods of time. They're not sustainable across a chip. With Arm AGI CPU, what we give you is full performance sustainably all the time. And ultimately, that means scale. We linearly scale across cores, and our memory and I/O subsystem is specifically designed to be matched to those cores so that we can continue to feed them, 6 gigabytes per second of memory bandwidth to every single core. In order to scale, what we see some of these legacy architectures do is multi-threading, right? What happens when you do multi-threading? You throw 2 jobs at the same core. That's how they get to a high thread count or try to get to a lot of devices. But the reality on that is your I/O and your bandwidth, that doesn't double. So you've just moved the bottleneck elsewhere. And oh, by the way, the CPU needs to be burdened with managing that back and forth. And so your performance degrades, you end up starving your processes. What we see over and over again is that data center operators have to overprovision their data centers by 30% or more to deal with this lack of nonlinear scaling. This is an actual thing that happens. We take pride in not having to do that. There's actually a great demo of this out on the show floor. I encourage you all to check it out after the keynote. And then finally, we have this maniacal focus on efficiency. Obviously, that's always been Mark's -- that's always been Arm's hallmark. It's always been something that we've been great at. We're leveraging all those techniques and methods and experience that we've built up over the decades around building incredibly efficient processors, incredibly efficient technology. And we're packaging that all up in a custom design specifically for this use case. AGI CPU is purpose-built without that legacy overhead because it all comes back to performance, scale and efficiency. It's my efficiency bullet. At the end of the day, no wasted cycles, no stranded compute, no wasted power or silicon, and we're super proud of that. Let's look at what it means in practice. I'm going to show you the results, and they kind of speak for themselves. First, let's talk about sustained performance. What you see here is the performance that you can expect to achieve consistently. So this is consistent performance. No performance throttling because you're over power budget, no memory or I/O contention. This is the sort of performance you're going to see. You can see with AGI CPU, it's world-class. You've got world-class performance, you can take to the bank. Next, let's talk about scale. How many threads or agents can you run in each rack? How much compute do you actually support with a fixed power budget with a fixed physical footprint. Remember those racks I showed you earlier, there you go. That's where we land. And of course, there's efficiency, performance per watt. What's going on with my screens. They're flipping all over the place. Can you go back, please? Go back one more. So what you're seeing here, these are -- all of these charts are with SMT disabled. So these are single-threaded cores for us, single-threaded cores for them. So no multi-threading whatsoever, okay? I told you what I thought about multi-threading, which is why we elected to show it to you this way. But oftentimes, what we hear is that multi-threading is going to improve that middle chart. It's going to allow for more scalability. Multi-threading is going to improve the performance per watt. Let's take a look at what happens if we turn multi-threading on, okay? See, first of all, your performance goes down. That's the chart on the left. And the reason why the performance goes down is because you can't just add more work and expect performance to be the same. So that's pretty self-explanatory. And in this particular case, again, we've held it at kind of based on the memory and the I/O bandwidth available, kind of where you land. That second one, the sustained threads per rack. The reality is that because of the limitations on the device and all of the bottlenecks, you end up in a scenario where you can't actually use all of those threads. Many are left idle. And then finally, performance per watt. Yes, there is a small improvement there, but not enough to change the calculus. At the end of the day, the results are clear. This is a killer product and Arm is a class of its own. Performance, scale and efficiency. I'll say it one more time. This is what the Arm AGI CPU is built for. And the impact on the AI data center is going to be profound. Let me turn it back to Rene. Thank you. | **Mohamed Awad:** ์ค๋ ์ ํฌ๋ฅผ ์ง์ํด์ฃผ์๋ ๋ชจ๋ ๊ณ ๊ฐ๋ถ๋ค๊ณผ ํํธ๋๋ถ๋ค๊ป ๋ค์ ํ๋ฒ ๊น์ด ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ ํฌ์๊ฒ ๋ณด๋ด์ฃผ์ ์ฑ์์ ์ ๋ง ๋๋จํ์ต๋๋ค. ์ ํฌ๋ ๊ณ ๊ฐ ์ฌ๋ฌ๋ถ์ ์ํด Arm AGI CPU๋ฅผ ๊ฐ๋ฐํ์ผ๋ฉฐ, ๊ทธ ๋ฐ์์ ๋งค์ฐ ๋ง์กฑํ๊ณ ์์ต๋๋ค. ์์๋ค์ํผ, Arm AGI CPU๋ ์ฑ๋ฅ ํ์ฅ์ฑ(performance scalability)์ ํ๋ณดํ๊ณ ์ ๋ ฅ ํจ์จ์ฑ(power efficiency)์ ์์ธก ๊ฐ๋ฅํ๊ฒ ์ ์งํ๋๋ก ์ฒ์๋ถํฐ ์ค๊ณ๋์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ์ ํฌ์ ํต์ฌ ๊ฐ์ , ์ฆ ์ฑ๋ฅ, ํ์ฅ์ฑ, ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด๋ ์ ํฌ ํํธ๋์ฌ๋ค์๊ฒ๋ ํฐ ๋ฐํฅ์ ์ป๊ณ ์์ต๋๋ค. ์์๋ค์ํผ, ์ด๋ x86 ์ํคํ ์ฒ์๋ ๋งค์ฐ ๋ค๋ฅธ ์ ๊ทผ ๋ฐฉ์์ ๋๋ค. x86์ ์คํ ์ค๋ฒํค๋(execution overhead)์ ๋ ๊ฑฐ์ ๊ธฐ๋ฅ(legacy feature) ์ง์์ผ๋ก ์ธํด ๋ถ๋ด์ ์๊ณ ์์ต๋๋ค. ๊ทธ๋ค์ ๋ชจ๋์ฑ(modularity), ๋ค์ํ ์์ฅ ์ง์, ๊ทธ๋ฆฌ๊ณ ํน์ํ(esoteric) ์ฌ์ฉ ์ฌ๋ก(use cases) ๋ฑ์ ์ค์ ์ ๋๊ธฐ๋ก ์ ํํ์ต๋๋ค. ์ ํฌ๋ ํจ์จ์ฑ ํฅ์๊ณผ ์ง์ฐ ์๊ฐ(latency) ๋จ์ถ์ ์ฒ ์ ํ๊ฒ ์ง์คํ๊ณ ์์ต๋๋ค. ๊ถ๊ทน์ ์ผ๋ก, ์ด๋ ์ํคํ ์ฒ ์ฒ ํ์ ๊ดํ ๋ฌธ์ ์ ๋๋ค. ์ ํฌ๋ ๊ณผ๊ฑฐ์ ์ฝ๋งค์ฌ ์์ง ์์ต๋๋ค. ์ ๋๋ก ๊ณผ๊ฑฐ์ ์ฝ๋งค์ฌ ์์ง ์์ต๋๋ค. ์, ์ ํฌ๋ ๋กํฐ์ค ๋ ธ์ธ (Lotus Notes)๋ฅผ ์ง์ํ์ง ์์ต๋๋ค. ์ ํฌ๋ ๊ทธ๋ ๊ฒ ํ์ง ์์ต๋๋ค. ์ ํฌ๋ ๋ฐ์ดํฐ์ผํฐ์ AGI์ ํ์ํ ๊ฒ, ์ฆ ์ฑ๋ฅ, ํ์ฅ์ฑ(scale), ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ์๋ง ์ ํํ ์ง์คํ๊ณ ์์ต๋๋ค. ์ด ๋ถ๋ถ์ ๋ํด ์ข ๋ ์์ธํ ์ค๋ช ํด ๋๋ฆฌ๊ฒ ์ต๋๋ค. ๋จผ์ ์ฑ๋ฅ๋ถํฐ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ ํฌ์๊ฒ ์ฑ๋ฅ์ด๋ ๋ชจ๋ ํด๋ก ์ฌ์ดํด(clock cycle)๋น ๋ ๋ง์ ์์ ์ ์ฒ๋ฆฌํ๋ ๊ฒ์ ์๋ฏธํฉ๋๋ค. ์ด ๋ถ๋ถ์ ํญ์ ์ค์ํ ์์ญ์ด์์ต๋๋ค. ๋ฐ์ด๋ IPC(Instructions Per Cycle)๋ Arm์ด ํญ์ ๋๊ฐ์ ๋ํ๋๋ ๋ถ์ผ์ ๋๋ค. ๋งค ์ฌ์ดํด๋ง๋ค ์ผ๋ง๋ ๋ง์ ์์ ์ ์ฒ๋ฆฌํ ์ ์๋๊ฐ ํ๋ ๊ฒ์ด์ฃ . ์ ํฌ AGI CPU๋ ์ด ๋ถ๋ถ์์ ๋จ์ฐ ๋๋ณด์ ๋๋ค. ๊ธฐ์กด ์ค์์ฒ๋ฆฌ์ฅ์น(CPU)๋ค์ ๋๋๋ก ์ฃผํ์(frequency)๋ฅผ ๋์ด๊ฑฐ๋ ๋ถ์คํธ ๋ชจ๋(boost mode)๋ฅผ ์ฌ์ฉํ๋ ๋ฐฉ์์ผ๋ก ์ด ๋ถ์ผ์์ ๊ฒฝ์ํ๋ ค๊ณ ํฉ๋๋ค. ํ์ง๋ง ํ์ค์ ์ด๋ ์ต๋๋ค. ์ฃผํ์๋ฅผ ๋์ด๋ฉด ๋ฌด์์ด ํจ๊ป ๋์ด๋ ๊น์? ๋ฐ๋ก ์ ๋ ฅ(power)์ ๋๋ค. ๊ทธ๊ฒ ๋ฌธ์ ์ ๋๋ค. ์ด๋ฌํ ๋ถ์คํธ ๋ชจ๋๋ ์ฅ์๊ฐ ์ง์ ๊ฐ๋ฅํ์ง ์์ผ๋ฉฐ, ์นฉ(chip) ์ ์ฒด์์ ์ง์๋ ์๋ ์์ต๋๋ค. Arm AGI CPU๋ฅผ ํตํด ์ ํฌ๋ ํญ์ ์ง์ ๊ฐ๋ฅํ ์์ ํ ์ฑ๋ฅ์ ์ ๊ณตํฉ๋๋ค. ๊ถ๊ทน์ ์ผ๋ก ๊ทธ๊ฒ์ ํ์ฅ์ฑ(scale)์ ์๋ฏธํฉ๋๋ค. ์ ํฌ๋ ์ฝ์ด(core) ์ ๋ฐ์ ๊ฑธ์ณ ์ ํ์ ์ผ๋ก ํ์ฅ๋๋ฉฐ, ์ ํฌ์ ๋ฉ๋ชจ๋ฆฌ(memory) ๋ฐ I/O ์๋ธ์์คํ (subsystem)์ ํด๋น ์ฝ์ด์ ๋ง์ถฐ ํน๋ณํ ์ค๊ณ๋์ด ๊ฐ ์ฝ์ด์ ์ด๋น 6๊ธฐ๊ฐ๋ฐ์ดํธ(GB/s)์ ๋ฉ๋ชจ๋ฆฌ ๋์ญํญ(memory bandwidth)์ ์ง์์ ์ผ๋ก ๊ณต๊ธํ ์ ์์ต๋๋ค. ํ์ฅํ๊ธฐ ์ํด ์ผ๋ถ ๊ธฐ์กด ์ํคํ ์ฒ(architecture)๋ค์ด ์ฌ์ฉํ๋ ๋ฐฉ์์ ๋ฉํฐ ์ค๋ ๋ฉ(multi-threading)์ ๋๋ค, ๊ทธ๋ ์ฃ ? ๋ฉํฐ์ค๋ ๋ฉ(multi-threading)์ ํ๋ฉด ์ด๋ค ์ผ์ด ๋ฒ์ด์ง๊น์? ํ๋์ ์ฝ์ด์ ๋ ๊ฐ์ ์์ ์ ํ ๋นํ๊ฒ ๋ฉ๋๋ค. ๊ทธ๋ ๊ฒ ํด์ ๋์ ์ค๋ ๋ ์(thread count)๋ฅผ ๋ฌ์ฑํ๊ฑฐ๋ ๋ ๋ง์ ๋๋ฐ์ด์ค(device)๋ฅผ ํ์ฉํ๋ ค๊ณ ํ์ฃ . ํ์ง๋ง ํ์ค์ ์ ์ถ๋ ฅ(I/O)๊ณผ ๋์ญํญ(bandwidth)์ด ๋ ๋ฐฐ๋ก ๋์ด๋์ง ์๋๋ค๋ ๊ฒ๋๋ค. ๊ฒฐ๊ตญ ๋ณ๋ชฉ ํ์(bottleneck)๋ง ๋ค๋ฅธ ๊ณณ์ผ๋ก ์ฎ๊ธฐ๋ ์ ์ด์ฃ . ๊ฒ๋ค๊ฐ CPU๋ ๊ทธ ์๋ค ๊ฐ๋ค ํ๋ ์์ ์ ๊ด๋ฆฌํ๋ ๋ถ๋ด๊น์ง ์ ธ์ผ ํฉ๋๋ค. ๊ทธ ๊ฒฐ๊ณผ ์ฑ๋ฅ์ด ์ ํ๋๊ณ , ํ๋ก์ธ์ค(process)์ ์์์ด ๋ถ์กฑํด์ง๋ ์ด๋ฅธ๋ฐ 'ํ๋ก์ธ์ค ๊ธฐ์ ํ์'์ด ๋ฐ์ํฉ๋๋ค. ์ ํฌ๊ฐ ๋ฐ๋ณต์ ์ผ๋ก ๋ชฉ๊ฒฉํ๋ ๊ฒ์, ๋ฐ์ดํฐ์ผํฐ ์ด์์๋ค์ด ์ด๋ฌํ ๋น์ ํ์ ํ์ฅ์ฑ(nonlinear scaling) ๋ถ์กฑ ๋ฌธ์ ๋ฅผ ํด๊ฒฐํ๊ธฐ ์ํด ๋ฐ์ดํฐ์ผํฐ๋ฅผ 30% ์ด์ ๊ณผ๋ํ๊ฒ ํ๋ก๋น์ ๋(overprovision)ํด์ผ ํ๋ค๋ ์ ์ ๋๋ค. ์ด๋ ์ค์ ๋ก ๋ฒ์ด์ง๋ ์ผ์ ๋๋ค. ์ ํฌ๋ ๊ทธ๋ ๊ฒ ํ ํ์๊ฐ ์๋ค๋ ์ ์ ์๋ถ์ฌ์ ๋๋๋๋ค. ์ฌ์ค, ์ ์ํ์ฅ์๋ ์ด์ ๊ด๋ จ๋ ํ๋ฅญํ ๋ฐ๋ชจ๊ฐ ์ค๋น๋์ด ์์ต๋๋ค. ๊ธฐ์กฐ์ฐ์ค(keynote)์ด ๋๋ ํ ์ฌ๋ฌ๋ถ ๋ชจ๋ ํ์ธํด ๋ณด์๊ธฐ๋ฅผ ๊ถํฉ๋๋ค. ๋ง์ง๋ง์ผ๋ก, ์ ํฌ๋ ํจ์จ์ฑ์ ๊ทน๋๋ก ์ง์คํ๊ณ ์์ต๋๋ค. ์์๋ค์ํผ, ์ด๋ ํญ์ Arm์ ํน์ง์ด์๊ณ , ์ ํฌ๊ฐ ํญ์ ์ํด์๋ ๋ถ๋ถ์ด๊ธฐ๋ ํฉ๋๋ค. ์ ํฌ๋ ์์ญ ๋ ๊ฐ ์ถ์ ํด ์จ ๋งค์ฐ ํจ์จ์ ์ธ ํ๋ก์ธ์(processor)์ ๊ธฐ์ ์ ๊ตฌ์ถํ๋ ๋ฐ ํ์ํ ๋ชจ๋ ๊ธฐ์ , ๋ฐฉ๋ฒ, ๊ฒฝํ์ ํ์ฉํ๊ณ ์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด ๋ชจ๋ ๊ฒ์ ์ด ํน์ ์ฌ์ฉ ์ฌ๋ก(use case)์ ๋ง์ถฐ ํน๋ณํ ๋ง์ถคํ ์ค๊ณ(custom design)๋ก ํตํฉํ๊ณ ์์ต๋๋ค. AGI CPU๋ ์ฑ๋ฅ, ํ์ฅ์ฑ(scale), ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ์ผ๋ก ๊ท๊ฒฐ๋๊ธฐ ๋๋ฌธ์, ๊ธฐ์กด์ ๋ถํ์ํ ์ค๋ฒํค๋(overhead) ์์ด ํน์ ๋ชฉ์ ์ ์ํด ์ค๊ณ๋์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ์ ํฌ ํจ์จ์ฑ์ ํต์ฌ์ ๋๋ค. ๊ฒฐ๊ตญ, ๋ญ๋น๋๋ ์ฌ์ดํด(cycle)์ด๋ ๊ณ ๋ฆฝ๋๋ ์ปดํจํ (compute) ์์, ๋ญ๋น๋๋ ์ ๋ ฅ์ด๋ ์ค๋ฆฌ์ฝ(silicon)์ด ์ ํ ์์ต๋๋ค. ์ ํฌ๋ ์ด์ ๋ํด ๋งค์ฐ ์๋์ค๋ฝ๊ฒ ์๊ฐํฉ๋๋ค. ์ด๊ฒ์ด ์ค์ ๋ก ์ด๋ค ์๋ฏธ๋ฅผ ๊ฐ์ง๋์ง ์ดํด๋ณด๊ฒ ์ต๋๋ค. ๊ฒฐ๊ณผ๋ฅผ ๋ณด์ฌ๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ด ๊ฒฐ๊ณผ๋ค์ด ๋ชจ๋ ๊ฒ์ ๋งํด์ค ๊ฒ๋๋ค. ๋จผ์ , ์ง์์ ์ธ ์ฑ๋ฅ(sustained performance)์ ๋ํด ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ฌ๊ธฐ์ ๋ณด์๋ ๊ฒ์ ์ฌ๋ฌ๋ถ์ด ๊พธ์คํ ๋ฌ์ฑํ ์ ์๋ค๊ณ ๊ธฐ๋ํ ์ ์๋ ์ฑ๋ฅ์ ๋๋ค. ์ฆ, ์ด๋ ์ผ๊ด๋ ์ฑ๋ฅ(consistent performance)์ ๋๋ค. ์ ๋ ฅ ์์ฐ(power budget) ์ด๊ณผ๋ก ์ธํ ์ฑ๋ฅ ์ ํ(performance throttling)๋ ์๊ณ , ๋ฉ๋ชจ๋ฆฌ๋ I/O ๊ฒฝํฉ(contention)๋ ์์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก ์ฌ๋ฌ๋ถ์ด ๊ฒฝํํ๊ฒ ๋ ์ฑ๋ฅ์ ๋๋ค. AGI CPU๋ฅผ ํตํด ์ธ๊ณ ์ต๊ณ ์์ค์ ์ฑ๋ฅ์ ํ์ธํ์ค ์ ์์ต๋๋ค. ํ์คํ ๋ฏฟ์ ์ ์๋ ์ธ๊ณ ์ต๊ณ ์์ค์ ์ฑ๋ฅ์ ์ ๊ณตํฉ๋๋ค. ๋ค์์ผ๋ก, ํ์ฅ์ฑ(scale)์ ๋ํด ์ด์ผ๊ธฐํด ๋ณด๊ฒ ์ต๋๋ค. ๊ฐ ๋(rack)์์ ์ผ๋ง๋ ๋ง์ ์ค๋ ๋(threads)๋ ์์ด์ ํธ(agents)๋ฅผ ์คํํ ์ ์์๊น์? ๊ณ ์ ๋ ์ ๋ ฅ ์์ฐ๊ณผ ๊ณ ์ ๋ ๋ฌผ๋ฆฌ์ ๊ณต๊ฐ(physical footprint) ๋ด์์ ์ค์ ๋ก ์ผ๋ง๋ ๋ง์ ์ปดํจํ (compute)์ ์ง์ํ ์ ์์๊น์? ์ด์ ์ ๋ณด์ฌ๋๋ ธ๋ ๋๋ค์ ๊ธฐ์ตํ์์ฃ ? ๋ฐ๋ก ๊ทธ๊ฒ๋๋ค. ์ ํฌ์ ์ญ๋์ด ๋ฐ๋ก ์ฌ๊ธฐ์ ์์ต๋๋ค. ๋ฌผ๋ก ์ด์ฃ , ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ, ์ฆ ์ํธ๋น ์ฑ๋ฅ(performance per watt)๋ ์ค์ํฉ๋๋ค. ์ ํ๋ฉด์ด ์ ์ด๋ฌ์ฃ ? ๊ณ์ ๋์ด๊ฐ๋๋ค. ์ฃ์กํฉ๋๋ค. ํ ์ฅ๋ง ๋ค๋ก ๋๊ฒจ์ฃผ์๊ฒ ์ด์? ๋ค, ํ ์ฅ ๋์. ์ง๊ธ ๋ณด์๋ ์ฐจํธ๋ค์ ๋ชจ๋ SMT(Simultaneous Multi-threading)๋ฅผ ๋นํ์ฑํํ ์ํ์ ๋๋ค. ์ฆ, ์ ํฌ ์ฝ์ด๋ ์๋๋ฐฉ ์ฝ์ด๋ ๋ชจ๋ ์ฑ๊ธ ์ค๋ ๋(single-threaded) ์ฝ์ด๋ผ๋ ๋ป์ ๋๋ค. ๋ฉํฐ ์ค๋ ๋ฉ(multi-threading)์ ์ ํ ์ฌ์ฉํ์ง ์์์ต๋๋ค. ์ ๊ฐ ๋ฉํฐ ์ค๋ ๋ฉ์ ๋ํด ์ด๋ป๊ฒ ์๊ฐํ๋์ง ๋ง์๋๋ ธ๊ณ , ๊ทธ๋์ ์ ํฌ๋ ์ด๋ฐ ๋ฐฉ์์ผ๋ก ๋ณด์ฌ๋๋ฆฌ๊ธฐ๋ก ๊ฒฐ์ ํ์ต๋๋ค. ํ์ง๋ง ์ข ์ข ๋ฉํฐ ์ค๋ ๋ฉ์ด ๊ฐ์ด๋ฐ ์ฐจํธ๋ฅผ ๊ฐ์ ํ๊ณ , ๋ ๋์ ํ์ฅ์ฑ(scalability)์ ์ ๊ณตํ๋ฉฐ, ์ํธ๋น ์ฑ๋ฅ(performance per watt)์ ํฅ์์ํฌ ๊ฒ์ด๋ผ๊ณ ๋ค ํฉ๋๋ค. ๊ทธ๋ผ ๋ฉํฐ ์ค๋ ๋ฉ์ ์ผฐ์ ๋ ์ด๋ค ์ผ์ด ์ผ์ด๋๋์ง ํ๋ฒ ๋ณผ๊น์? ๋ณด์๋ค์ํผ, ์ฐ์ ์ฑ๋ฅ์ด ๋จ์ด์ง๋๋ค. ๊ทธ๊ฒ์ด ์ผ์ชฝ์ ์๋ ์ฐจํธ์ ๋๋ค. ์ฑ๋ฅ์ด ์ ํ๋๋ ์ด์ ๋ ๋จ์ํ ์์ ๋์ ๋๋ฆฐ๋ค๊ณ ํด์ ๋์ผํ ์ฑ๋ฅ์ ๊ธฐ๋ํ ์ ์๊ธฐ ๋๋ฌธ์ ๋๋ค. ์ด๋ ๋งค์ฐ ์๋ช ํ ์ฌ์ค์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ด ํน์ ์ฌ๋ก์์๋, ๋ค์ ๋ง์๋๋ฆฌ์ง๋ง, ์ฌ์ฉ ๊ฐ๋ฅํ ๋ฉ๋ชจ๋ฆฌ(memory)์ I/O ๋์ญํญ(I/O bandwidth)์ ๊ธฐ๋ฐ์ผ๋ก ์ด๋ ์ ๋์ ๊ฒฐ๊ณผ์ ๋๋ฌํ๋์ง๋ฅผ ์ ์งํ์ต๋๋ค. ๋ ๋ฒ์งธ๋ ๋๋น ์ง์ ์ค๋ ๋(sustained threads per rack)์ ๋๋ค. ํ์ค์ ์ฅ์น(device)์ ์ ์ฝ๊ณผ ๋ชจ๋ ๋ณ๋ชฉ ํ์(bottlenecks)์ผ๋ก ์ธํด, ์ค์ ๋ก ๊ทธ ๋ชจ๋ ์ค๋ ๋(thread)๋ฅผ ์ฌ์ฉํ ์ ์๋ ์ํฉ์ ์ด๋ฅด๊ฒ ๋๋ค๋ ๊ฒ์ ๋๋ค. ๋ง์ ์ค๋ ๋๊ฐ ์ ํด ์ํ๋ก ๋จ๊ฒ ๋ฉ๋๋ค. ๊ทธ๋ฆฌ๊ณ ๋ง์ง๋ง์ผ๋ก ์ํธ๋น ์ฑ๋ฅ(performance per watt)์ ๋๋ค. ๋ค, ๊ฑฐ๊ธฐ์๋ ์ฝ๊ฐ์ ๊ฐ์ ์ด ์์์ง๋ง, ์ ์ฒด์ ์ธ ํ๋๋ฅผ ๋ฐ๊ฟ ๋งํผ์ ์๋๋๋ค. ๊ฒฐ๋ก ์ ์ผ๋ก, ๊ฒฐ๊ณผ๋ ๋ช ํํฉ๋๋ค. ์ด๊ฒ์ ํ๊ธฐ์ ์ธ ์ ํ(killer product)์ด๋ฉฐ, Arm์ ๊ทธ ์์ฒด๋ก ๋ ๋ณด์ ์ธ ์กด์ฌ์ ๋๋ค. ์ฑ๋ฅ, ๊ท๋ชจ, ๊ทธ๋ฆฌ๊ณ ํจ์จ์ฑ์ ๋๋ค. ๋ค์ ํ๋ฒ ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ด๊ฒ์ด ๋ฐ๋ก Arm AGI CPU๊ฐ ์ค๊ณ๋ ๋ชฉ์ ์ ๋๋ค. ๊ทธ๋ฆฌ๊ณ AI ๋ฐ์ดํฐ ์ผํฐ์ ๋ฏธ์น๋ ์ํฅ์ ์์ฒญ๋ ๊ฒ์ ๋๋ค. ์ด์ ๋ฅด๋ค์๊ฒ ๋ง์ดํฌ๋ฅผ ๋๊ธฐ๊ฒ ์ต๋๋ค. ๊ฐ์ฌํฉ๋๋ค. |
| CEO & Director: Thank you, Mohamed, and thank you, Paul, and your LLM agent that's going to do all the conversions for us. So we've shared a lot with you today, and I'm grateful for your patience and time. If there were just a few things to take away from this morning, I think it starts here. performance per watt, which translates to performance per rack. When you look at an x86 equivalent structure, same power delivery, 36 kilowatts, 2x the performance in the same power. That's what you need to remember. For those of you who are paying for that power, there's another number you need to remember. If you think about 1 gigawatt of capacity and you think about the CapEx associated with that extra power you're spending at the sake of performance, it's up to $10 billion of CapEx. Obviously, these are serious numbers. So again, the takeaway from the Arm AGI CPU is 2x performance per watt, probably more than 2x -- now you heard a number of comments in the videos, including Santosh, that when you embark on the kind of engagement and partnership we're talking about, while a day like this and an event like this is wonderful and amazing and we're talking about a great product, it's really not about the day, but it's about the future and commitment to a road map. So we are committing to future generations of this product. Arm AGI CPU 2 is coming out soon as is Arm AGI CPU 3. As you heard in the videos again, these are multigenerational engagements. We are investing a lot. Our customers are investing a lot. The ecosystem is investing a lot. We are absolutely committed to a road map and a future around this product line. In addition, we will continue the CSSs around these products. And as Mohamed mentioned, one of the big benefits of the CSSs are the speed it allows our customers to get to market. It also enables a lot of benefit for us as well. So the CSS road map will continue. So I want to close a little bit around what we think the financial opportunity is for Arm. So before this day, our business has been IP and IP compute subsystems. And we have been doing extremely well in that business, far better than what we had talked to investors about 2.5 years ago when we did our roadshow for the IPO, we're actually ahead of that. When we look at the AI data center business, that represents today about a $3 billion TAM. And now I'm just talking about roughly the royalties. So I mentioned on one of the earnings calls that the cloud AI business will probably be our largest business in a few years. And this is really driven by all of the growth that Mohamed talked about, the deployment of 1.25 billion Neoverse cores and forward. When we think about our business going forward, the Arm AGI CPU and as Mohamed mentioned, we have committed customers, Meta, OpenAI, Cloudflare, SAP, F5, customer you saw in the video. When we think forward about what is the market opportunity for this business, it is a dramatic sea change for the opportunity. When we look at what's going on with Agentic AI, the growth of CPUs, the benefit that power-efficient CPUs bring to the data center, we think this represents about $100 billion TAM for us in the future. So today, it is all about the Arm AGI CPU. But there will be some tomorrows. And don't ask me about tomorrow today, but there will be some tomorrows. And we think this opportunity to take the work we've done across all of the markets, as you've heard in the video from edge to cloud, from milliwatts to gigawatts, we think we have an opportunity to address greater than a $1 trillion TAM by the end of the decade. So we've got some work to do, but I couldn't be more proud of what our company has achieved, grateful to the ecosystem that helps us achieve it and the customers that are now committed to buy our product. I want to close by saying that we stand on the shoulders of our ecosystem. None of this is possible without the ecosystem that we have nurtured for 35-plus years, many of you who are here today and watching on video. Thank you for attending today. Arm is everywhere, and we appreciate your support. | **CEO & Director:** ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ฅผ ์ํด ๋ชจ๋ ๋ณํ์ ์ฒ๋ฆฌํด ์ค LLM ์์ด์ ํธ์๊ฒ๋ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ค๋ ์ฌ๋ฌ๋ถ๊ณผ ๋ง์ ๊ฒ์ ๊ณต์ ํ์ผ๋ฉฐ, ์ธ๋ด์ฌ์ ๊ฐ์ง๊ณ ์๊ฐ์ ํ ์ ํด ์ฃผ์ ์ ์ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ์ค๋ ๋ง์๋๋ฆฐ ๋ด์ฉ ์ค ๋ช ๊ฐ์ง๋ง ๊ธฐ์ตํ์ ๋ค๋ฉด, ํต์ฌ์ ๋ฐ๋ก ์ด๊ฒ์ ๋๋ค. ์ํธ๋น ์ฑ๋ฅ(performance per watt)์ด๋ฉฐ, ์ด๋ ๊ณง ๋๋น ์ฑ๋ฅ(performance per rack)์ผ๋ก ์ด์ด์ง๋๋ค. x86 ๊ธฐ๋ฐ์ ๋๋ฑํ ๊ตฌ์กฐ๋ฅผ ์ดํด๋ณด๋ฉด, ๋์ผํ ์ ๋ ฅ ๊ณต๊ธ, ์ฆ 36ํฌ๋ก์ํธ(kW)์์ 2๋ฐฐ์ ์ฑ๋ฅ์ ๋ฐํํฉ๋๋ค. ์ด๊ฒ์ด ์ฌ๋ฌ๋ถ์ด ๊ธฐ์ตํด์ผ ํ ํต์ฌ์ ๋๋ค. ๊ทธ ์ ๋ ฅ ๋น์ฉ์ ์ง๋ถํ๋ ๋ถ๋ค๊ป๋ ๊ธฐ์ตํด์ผ ํ ๋ ๋ค๋ฅธ ์ซ์๊ฐ ์์ต๋๋ค. 1๊ธฐ๊ฐ์ํธ(GW)์ ์ฉ๋์ ๊ณ ๋ คํ๊ณ , ์ฑ๋ฅ์ ์ํด ์ถ๊ฐ๋ก ์ง์ถํ๋ ์ ๋ ฅ๊ณผ ๊ด๋ จ๋ ์๋ณธ ์ง์ถ(CapEx)์ ์๊ฐํด๋ณด์ญ์์ค. ์ด๋ ์ต๋ 100์ต ๋ฌ๋ฌ์ ๋ฌํ๋ ์๋ณธ ์ง์ถ์ด ๋ฐ์ํ ์ ์์ต๋๋ค. ๋ถ๋ช ํ, ์ด ์์น๋ค์ ๋งค์ฐ ์ค์ํฉ๋๋ค. ๋ค์ ๋ง์๋๋ฆฌ์ง๋ง, Arm AGI CPU์ ํต์ฌ์ ์ํธ๋น 2๋ฐฐ์ ์ฑ๋ฅ ํฅ์์ด๋ฉฐ, ์๋ง๋ 2๋ฐฐ ์ด์์ผ ๊ฒ์ ๋๋ค. ์ฐํ ์๋ฅผ ๋น๋กฏํด ์์์์ ์ฌ๋ฌ ์๊ฒฌ์ ๋ค์ผ์ จ๊ฒ ์ง๋ง, ์ ํฌ๊ฐ ๋ ผ์ํ๋ ์ด๋ฌํ ์ข ๋ฅ์ ํ๋ ฅ๊ณผ ํํธ๋์ญ(partnership)์ ์์ํ ๋, ์ค๋๊ณผ ๊ฐ์ ๋ , ๊ทธ๋ฆฌ๊ณ ์ด๋ฌํ ํ์ฌ๊ฐ ํ๋ฅญํ๊ณ ๋๋๊ณ , ์ ํฌ๊ฐ ๋ฉ์ง ์ ํ์ ๋ํด ์ด์ผ๊ธฐํ๊ณ ์์ง๋ง, ์ ๋ง๋ก ์ค์ํ ๊ฒ์ ์ค๋ ํ๋ฃจ๊ฐ ์๋๋ผ ๋ฏธ๋์ ๋ก๋๋งต(road map)์ ๋ํ ์ฝ์์ ๋๋ค. ๋ฐ๋ผ์ ์ ํฌ๋ ์ด ์ ํ์ ๋ฏธ๋ ์ธ๋์ ์ ๋ ํ๊ณ ์์ต๋๋ค. Arm AGI CPU 2๊ฐ ๊ณง ์ถ์๋ ์์ ์ด๋ฉฐ, Arm AGI CPU 3๋ ๋ง์ฐฌ๊ฐ์ง์ ๋๋ค. ์์์์ ๋ค์ ๋ค์ผ์ จ๊ฒ ์ง๋ง, ์ด๋ ์ฌ๋ฌ ์ธ๋์ ๊ฑธ์น ํ๋ ฅ์ ๋๋ค. ์ ํฌ๋ ๋ง์ ํฌ์๋ฅผ ํ๊ณ ์์ต๋๋ค. ์ ํฌ ๊ณ ๊ฐ๋ค๋ ๋ง์ ํฌ์๋ฅผ ํ๊ณ ์์ผ๋ฉฐ, ์ํ๊ณ(ecosystem) ์ ๋ฐ์์๋ ๋ง์ ํฌ์๊ฐ ์ด๋ฃจ์ด์ง๊ณ ์์ต๋๋ค. ์ ํฌ๋ ์ด ์ ํ๊ตฐ์ ๋ํ ๋ก๋๋งต๊ณผ ๋ฏธ๋์ ๋ํด ํ๊ณ ํ ์์ง๋ฅผ ๊ฐ์ง๊ณ ์์ต๋๋ค. ๋ํ, ์ด ์ ํ๋ค์ ๋ํ CSS(Compute Subsystem)๋ฅผ ๊ณ์ ์ด์ด๊ฐ ๊ฒ์ ๋๋ค. Mohamed๊ฐ ์ธ๊ธํ๋ฏ์ด, CSS์ ํฐ ์ฅ์ ์ค ํ๋๋ ๊ณ ๊ฐ๋ค์ด ์์ฅ์ ๋ ๋น ๋ฅด๊ฒ ์ง์ถํ ์ ์๋๋ก ํด์ฃผ๋ ์๋์ ๋๋ค. ์ด๋ ๋ํ ์ ํฌ์๊ฒ๋ ๋ง์ ์ด์ ์ ์ ๊ณตํฉ๋๋ค. ๋ฐ๋ผ์ CSS ๋ก๋๋งต์ ๊ณ์๋ ๊ฒ์ ๋๋ค. ์ด์ ๋ง๋ฌด๋ฆฌํ๋ฉด์ Arm์๊ฒ ์์ด ์ฌ์ ์ ๊ธฐํ(financial opportunity)๊ฐ ๋ฌด์์ธ์ง์ ๋ํด ๋ง์๋๋ฆฌ๊ณ ์ถ์ต๋๋ค. ์ง๊ธ๊น์ง ์ ํฌ ์ฌ์ ์ IP(Intellectual Property)์ IP ์ปดํจํธ ์๋ธ์์คํ ์ด์์ต๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ ๊ทธ ์ฌ์ ์์ ๋งค์ฐ ์ ํด์์ผ๋ฉฐ, 2๋ ๋ฐ ์ IPO(๊ธฐ์ ๊ณต๊ฐ) ๋ก๋์ผ๋ฅผ ์งํํ ๋ ํฌ์์๋ค์๊ฒ ๋ง์๋๋ ธ๋ ๊ฒ๋ณด๋ค ํจ์ฌ ๋ ์ํ๊ณ ์์ต๋๋ค. ์ค์ ๋ก ์ ํฌ๋ ๊ทธ ๋ชฉํ๋ฅผ ๋ฐ์ด๋๊ณ ์์ต๋๋ค. AI ๋ฐ์ดํฐ์ผํฐ ์ฌ์ ์ ์ดํด๋ณด๋ฉด, ํ์ฌ ์ฝ 30์ต ๋ฌ๋ฌ ๊ท๋ชจ์ TAM(Total Addressable Market)์ ๋ํ๋ ๋๋ค. ์ง๊ธ๋ถํฐ ๋ก์ดํฐ(royalties)์ ๋ํด ๋๋ต์ ์ผ๋ก ๋ง์๋๋ฆฌ๊ฒ ์ต๋๋ค. ์ด์ ์ค์ ๋ฐํ(earnings call) ์ค ํ ๋ฒ ๋ง์๋๋ ธ๋ฏ์ด, ํด๋ผ์ฐ๋ AI ์ฌ์ ์ ๋ช ๋ ์์ ์ ํฌ์ ๊ฐ์ฅ ํฐ ์ฌ์ ์ด ๋ ๊ฒ์ ๋๋ค. ์ด๋ ๋ชจํ๋ฉ๋๊ฐ ์ธ๊ธํ๋ ๋ชจ๋ ์ฑ์ฅ, ์ฆ 12์ต 5์ฒ๋ง ๊ฐ์ Neoverse ์ฝ์ด ๋ฐฐํฌ(deployment)์ ๊ทธ ์ดํ์ ์ฑ์ฅ์ ํ์ ์ ๊ฒ์ ๋๋ค. ์์ผ๋ก ์ ํฌ ์ฌ์ ์ ์๊ฐํด๋ณด๋ฉด, Arm AGI CPU์ ๊ด๋ จํ์ฌ ๋ชจํ๋ฉ๋๊ฐ ์ธ๊ธํ๋ฏ์ด, ์ ํฌ๋ ์ด๋ฏธ ํ์ ๋ ๊ณ ๊ฐ๋ค์ ํ๋ณดํ๊ณ ์์ต๋๋ค. Meta, OpenAI, Cloudflare, SAP, F5 ๋ฑ ๋น๋์ค์์ ๋ณด์ จ๋ ๊ณ ๊ฐ๋ค์ด์ฃ . ์ด ์ฌ์ ์ ์์ฅ ๊ธฐํ(market opportunity)์ ๋ํด ์์ผ๋ก ์๊ฐํด๋ณด๋ฉด, ์ด๋ ๊ธฐํ์ ์์ด ๊ทน์ ์ธ ๋๋ณํ(sea change)์ ๋๋ค. ์์ด์ ํธ AI, CPU์ ์ฑ์ฅ, ๊ทธ๋ฆฌ๊ณ ์ ๋ ฅ ํจ์จ์ ์ธ CPU๊ฐ ๋ฐ์ดํฐ์ผํฐ์ ๊ฐ์ ธ๋ค์ฃผ๋ ์ด์ ๋ค์ ์ข ํฉ์ ์ผ๋ก ๊ณ ๋ คํด๋ณผ ๋, ์ด๋ ๋ฏธ๋์ ์ ํฌ์๊ฒ ์ฝ 1,000์ต ๋ฌ๋ฌ ๊ท๋ชจ์ ์ด ์ ํจ ์์ฅ(TAM: Total Addressable Market)์ ์๋ฏธํ๋ค๊ณ ์๊ฐํฉ๋๋ค. ํ์ฌ๋ Arm AGI CPU๊ฐ ๋ชจ๋ ๊ฒ์ ์ค์ฌ์ด์ง๋ง, ๋ถ๋ช ํ ๋ฏธ๋๋ ๊ณ์ํด์ ํผ์ณ์ง ๊ฒ์ ๋๋ค. ์ค๋ ๋น์ฅ ๊ทธ ๋ฏธ๋์ ๋ํด ์์ธํ ๋ฌป์ง๋ ๋ง์ญ์์ค. ํ์ง๋ง ๋ถ๋ช ํ ๊ทธ ๋ฏธ๋๋ ์กด์ฌํฉ๋๋ค. ๊ทธ๋ฆฌ๊ณ ์ ํฌ๋ ์์์์ ๋ณด์ จ๋ฏ์ด ์ฃ์ง(edge)๋ถํฐ ํด๋ผ์ฐ๋(cloud)๊น์ง, ๋ฐ๋ฆฌ์ํธ(milliwatts)๋ถํฐ ๊ธฐ๊ฐ์ํธ(gigawatts)๊น์ง, ๋ชจ๋ ์์ฅ์์ ์ ํฌ๊ฐ ์ด๋ค๋ธ ์ฑ๊ณผ๋ฅผ ๋ฐํ์ผ๋ก ์ด๋ฒ 10๋ ๋ง๊น์ง 1์กฐ ๋ฌ๋ฌ๊ฐ ๋๋ ์ด ์ ํจ ์์ฅ(TAM)์ ๊ณต๋ตํ ์ ์๋ ๊ธฐํ๋ฅผ ๊ฐ์ง๊ณ ์๋ค๊ณ ๋ด ๋๋ค. ๋ฌผ๋ก ์์ง ํด์ผ ํ ์ผ์ด ๋จ์์์ต๋๋ค๋ง, ์ ํฌ ํ์ฌ๊ฐ ๋ฌ์ฑํ ์ฑ๊ณผ์ ๋ํด ๋ํ ๋์ ์์ด ์๋์ค๋ฝ์ต๋๋ค. ๋ํ ์ ํฌ์ ์ฑ๊ณผ๋ฅผ ๋๋ ์ํ๊ณ(ecosystem)์ ์ ํฌ ์ ํ ๊ตฌ๋งค๋ฅผ ์ฝ์ํด์ฃผ์ ๊ณ ๊ฐ๋ถ๋ค๊ป๋ ๊น์ด ๊ฐ์ฌ๋๋ฆฝ๋๋ค. ๋ง์ง๋ง์ผ๋ก ๋ง์๋๋ฆฌ๊ณ ์ถ์ ๊ฒ์, ์ ํฌ๋ ์ ํฌ ์ํ๊ณ์ ์ด๊นจ ์์ ์ ์๋ค๋ ์ ์ ๋๋ค. ์ค๋ ์ด ์๋ฆฌ์ ์ฐธ์ํด ์ฃผ์ ๋ง์ ๋ถ๋ค๊ณผ ์์์ ํตํด ์์ฒญํ๊ณ ๊ณ์ ๋ถ๋ค์ ํฌํจํ์ฌ, ์ ํฌ๊ฐ 35๋ ์ด์ ์ก์ฑํด ์จ ์ํ๊ณ ์์ด๋ ์ด ๋ชจ๋ ๊ฒ์ด ๋ถ๊ฐ๋ฅํ์ ๊ฒ์ ๋๋ค. ์ค๋ ์ฐธ์ํด ์ฃผ์ ์ ๊ฐ์ฌํฉ๋๋ค. Arm์ ์ด๋์๋ ์์ผ๋ฉฐ, ์ฌ๋ฌ๋ถ์ ์ฑ์์ ๊ฐ์ฌ๋๋ฆฝ๋๋ค. |
๋ค์์ ํด๋น ํธ๋์คํฌ๋ฆฝํธ์ ๋ํ ์์ฝ์
๋๋ค.
* **์ ์ ํ ์ถ์ ๋ฐ ์ฑ๋ฅ ์ฐ์ ๊ฐ์กฐ:** Arm์ AGI CPU๋ฅผ ๊ณต์ ์ถ์ํ๋ฉฐ x86 ๋๋น ์ํธ๋น ์ฑ๋ฅ(performance per watt)์์ 2๋ฐฐ์ ์ฐ์๋ฅผ ๊ฐ์กฐํ๊ณ ,