AMD CEO Lisa Su: We're seeing 30 times more performance with our Helios systems

Watch on YouTube

Now Playing

Transcript

155 segments

0:00

LOWS, BUT DRIFTING BACK TOWARDS

0:02

THERE. WE'RE BACK ON THE

0:03

CLOSING BELL. LET'S GET OUT TO

0:04

OUR OWN JOHN FORT, WHO JOINS US

0:06

WITH A VERY SPECIAL GUEST, JOHN

0:10

DOM.

0:10

>> THANKS. I AM.

0:11

>> HERE IN SAN FRANCISCO AT

0:12

AMD'S.

0:13

>> ADVANCING AI. EVENT WITH AMD

0:15

CEO.

0:16

>> LISA SU. LISA, THANKS FOR

0:17

HAVING US HERE, JOHN.

0:18

>> GREAT TO BE HERE. THANK YOU

0:20

FOR BEING HERE.

0:21

>> WELL, I GOT TO SET THE STAGE

0:23

A LITTLE BIT. INTEL'S GOT

0:24

EARNINGS THAT THEY'RE

0:25

ANNOUNCING IN JUST A FEW

0:26

MINUTES. WE'VE HAD MASSIVE

0:27

SIGNALS THIS WEEK ON AI

0:29

INFRASTRUCTURE DEMAND FROM

0:30

GOOGLE JUST LAST NIGHT. AND

0:32

THEN MASSIVE PANIC ATTACKS FROM

0:35

THE MARKET OVER KIMMY K THREE.

0:36

AND WHAT THAT MIGHT BE DOING TO

0:40

JUST DEMAND FOR PREMIUM ACROSS

0:43

THE BOARD IN AI, WHETHER IT'S

0:45

ON THE SOFTWARE OR THE HARDWARE

0:46

SIDE. SO WITH THAT AS A

0:47

BACKDROP, TELL ME ABOUT THE

0:50

SIGNIFICANCE OF THIS MOMENT FOR

0:52

HELIOS AND AMD.

0:53

>> YEAH. SO FIRST OF ALL, JOHN,

0:55

ALL OF THOSE THINGS ARE TRUE,

0:58

BUT WHAT WE SEE IS ACTUALLY

1:00

SOMETHING THAT'S MUCH, MUCH

1:01

BROADER. I MEAN, THIS IS LIKE

1:02

AN INCREDIBLE MOMENT FOR AI.

1:04

YOU KNOW, WE LOOK AT OUR

1:07

MARKETS, SPEND TIME WITH OUR

1:08

CUSTOMERS ALL THE TIME. AND,

1:09

YOU KNOW, FROM WHAT WE SEE

1:11

TODAY, THE MARKET FOR HIGH

1:13

PERFORMANCE IN AI COMPUTING

1:14

OVER THE NEXT 4 OR 5 YEARS IS

1:16

GOING TO REACH $2 TRILLION OF

1:18

TAM AND MARKET OPPORTUNITY. SO

1:20

IT'S REALLY HUGE AND TREMENDOUS.

1:22

AND I'M VERY HAPPY TO ANNOUNCE

1:24

THIS MORNING WE ANNOUNCED THAT

1:25

OUR HELIOS RACK SCALE SYSTEM IS

1:28

IN FULL PRODUCTION, AS WELL AS

1:30

OUR NEW VENICE PROCESSORS ARE

1:32

IN FULL PRODUCTION. AND THIS IS

1:34

LIKE A HUGE STEP UP IN TERMS OF

1:37

AI CAPABILITY. I MEAN, WE'RE

1:38

SEEING INCREDIBLE NUMBERS LIKE,

1:39

YOU KNOW, OVER 30 TIMES MORE

1:42

PERFORMANCE WITH OUR NEWEST

1:43

SYSTEMS, HELIOS SYSTEMS,

1:44

COMPARED TO OUR PREVIOUS

1:46

GENERATION. AND WE'RE TALKING

1:47

ABOUT, FROM A CUSTOMER

1:49

STANDPOINT, JUST TREMENDOUS

1:50

INTEREST AND EXCITEMENT. YOU

1:52

KNOW, WE HAD ANTHROPIC WITH US

1:53

OPEN AI, YOU KNOW, META, AT&T,

1:56

CISCO, JUST A NUMBER OF OUR

1:58

CUSTOMERS, YOU KNOW, COMING

2:00

TOGETHER.

2:00

>> YOU'VE LITERALLY BEEN

2:01

BUILDING TOWARD THIS PENSANDO.

2:02

I THINK ABOUT OUR CONVERSATIONS,

2:04

AUSTIN, SAN JOSE, ETC. YOU

2:07

ANNOUNCED THESE MAJOR CUSTOMERS

2:09

FOR HELIOS OUT OF THE GATE. HOW

2:11

MUCH OF THAT IS DEDICATED

2:13

VOLUME? HOW MUCH OF IT IS PILOT

2:15

TESTS?

2:16

>> WELL, I THINK WHAT YOU CAN

2:17

SAY IS WITH A SYSTEM AS COMPLEX

2:19

AS HELIOS, YOU KNOW, PEOPLE

2:21

DON'T DO THIS TYPE OF WORK FOR

2:24

QUOTE UNQUOTE PILOTS BECAUSE

2:26

IT'S A LOT OF WORK WHEN YOU

2:27

HAVE, YOU KNOW, MAJOR

2:29

FOUNDATIONAL MODEL COMPANIES

2:31

NEEDING TO PUT THEIR RESOURCES,

2:33

OPTIMIZE THEIR LATEST MODELS TO

2:35

OUR HARDWARE. SO CERTAINLY, YOU

2:38

KNOW, OPENAI AND META, WE'VE

2:39

TALKED ABOUT MULTI GIGAWATT

2:42

SCALE FOR MI 450 DEPLOYMENTS.

2:45

AND THEN JUST THIS WEEK WE

2:46

ANNOUNCED ANTHROPIC, WHICH WILL

2:49

DO UP TO TWO GIGAWATTS OF MI

2:51

450 IN HELIOS STARTING IN THE

2:53

FIRST HALF OF 27. SO THESE ARE

2:55

LARGE SCALE DEPLOYMENTS. AND

2:57

YOU KNOW, WE'RE WORKING HAND IN

2:59

HAND WITH THE ENGINEERING TEAMS

3:01

OF OUR CUSTOMERS, BOTH HARDWARE,

3:03

SOFTWARE, DATA CENTER. I CAN

3:05

SAY IT'S A TON OF WORK, BUT

3:06

IT'S GOING EXTREMELY WELL. SO

3:08

IT'S A, IT'S A HAPPY DAY FOR US.

3:10

>> I DON'T KNOW THAT YOU'VE

3:11

ANNOUNCED PRICING, BUT I'VE

3:12

SEEN SOME ESTIMATES, GUESSES

3:14

OUT THERE. THIS IS GOING TO BE

3:15

5 MILLION PLUS FOR HELIOS RACK

3:17

THAT YOU'RE PRICING ON THE

3:19

PREMIUM END. HOW MUCH OF THAT

3:21

IS BECAUSE TSMC IS CHARGING

3:22

MORE TO JUST DELIVER CHIPS? HOW

3:24

MUCH OF THAT WOULD JUST BE

3:25

BECAUSE OF THE CONSTRAINED

3:28

ENVIRONMENT AND SUPPLY IN

3:29

GENERAL?

3:29

>> WELL, THE MOST IMPORTANT

3:30

THING THAT I WOULD SAY, JOHN,

3:32

IS TO FOCUS ON TOKENS PER

3:34

DOLLAR. I MEAN, THAT'S WHAT THE

3:36

ENTIRE INDUSTRY IS TALKING

3:37

ABOUT. IT'S TALKING ABOUT

3:38

TOKENS PER DOLLAR. AND FROM OUR

3:41

STANDPOINT, YOU KNOW, WE

3:42

ACTUALLY WENT ON BOTH SIDES.

3:43

OUR HELIOS SYSTEMS HAVE MORE

3:46

PERFORMANCE THAN OUR

3:47

COMPETITION. AND SO, YOU KNOW,

3:48

LET'S CALL IT ANYWHERE FROM 10

3:50

TO 15% PLUS MORE PERFORMANCE.

3:52

AND WE ALSO HAVE LOWER TOKENS

3:54

PER DOLLAR. SO, YOU KNOW, UP TO

3:56

A 30% BENEFIT FOR OUR CUSTOMERS

3:58

IN TERMS OF TOKENS PER DOLLAR.

4:00

THAT'S HUGE. I MEAN, WHEN

4:01

YOU'RE TALKING ABOUT INFERENCE,

4:02

YOU KNOW, EVERYONE IS LOOKING

4:03

FOR THE OPPORTUNITY TO USE AI

4:06

IN MORE AND MORE APPLICATIONS.

4:08

AND IT IS SUCH AN SUCH A PLACE

4:10

WHERE EVERYONE'S TOKEN BUDGETS

4:12

ARE RAMPING UP, YOU KNOW,

4:13

INCLUDING OUR OWN, BY THE WAY,

4:14

EVERY MONTH OUR TOKEN BUDGETS

4:15

ARE GOING UP, BUT WE'RE ALSO

4:17

SEEING JUST TREMENDOUS

4:19

PRODUCTIVITY COME OUT OF THAT.

4:20

AND SO, YOU KNOW, THIS IS AN

4:22

OPPORTUNITY TO OPTIMIZE LIKE

4:23

NEW TECH IS THERE SO THAT YOU

4:25

CAN ADO

Interactive Summary

Ask follow-up questions or revisit key timestamps.

AMD CEO Lisa Su discusses the significant growth in the AI computing market, projecting it to reach a $2 trillion opportunity over the next five years. During the 'Advancing AI' event, Su announced that AMD's new Helios rack-scale systems and Venice processors are in full production, offering substantial performance improvements over previous generations. Major industry players like OpenAI, Meta, and Anthropic are already integrating these systems into their large-scale, multi-gigawatt deployments, focusing on maximizing efficiency through a better tokens-per-dollar ratio.