Lisa Su, chairwoman and CEO of Advanced Micro Devices (AMD), delivers the opening keynote code astatine Computex 2024, Taiwan's premier tech expo, successful Taipei connected June 3, 2024.
I-Hwa Cheng | Afp | Getty Images
AMD launched a caller artificial-intelligence spot connected Thursday that is taking nonstop purpose astatine Nvidia's information halfway graphics processors, known arsenic GPUs.
The Instinct MI325X, arsenic the spot is called, volition commencement accumulation earlier the extremity of 2024, AMD said Thursday during an lawsuit announcing the caller product. If AMD's AI chips are seen by developers and unreality giants arsenic a adjacent substitute for Nvidia's products, it could enactment pricing unit connected Nvidia, which has enjoyed astir 75% gross margins portion its GPUs person been successful precocious request implicit the past year.
Advanced generative AI specified arsenic OpenAI's ChatGPT requires monolithic information centers afloat of GPUs successful bid to bash the indispensable processing, which has created request for much companies to supply AI chips.
In the past fewer years, Nvidia has dominated the bulk of the information halfway GPU market, but AMD is historically successful 2nd place. Now, AMD is aiming to instrumentality stock from its Silicon Valley rival oregon astatine slightest to seizure a large chunk of the market, which it says volition beryllium worthy $500 cardinal by 2028.
"AI request has really continued to instrumentality disconnected and really transcend expectations. It's wide that the complaint of concern is continuing to turn everywhere," AMD CEO Lisa Su said astatine the event.
AMD didn't uncover caller large unreality oregon net customers for its Instinct GPUs astatine the event, but the institution has antecedently disclosed that some Meta and Microsoft bargain its AI GPUs and that OpenAI uses them for immoderate applications. The institution besides did not disclose pricing for the Instinct MI325X, which is typically sold arsenic portion of a implicit server.
With the motorboat of the MI325X, AMD is accelerating its merchandise docket to merchandise caller chips connected an yearly docket to amended vie with Nvidia and instrumentality vantage of the roar for AI chips. The caller AI spot is the successor to the MI300X, which started shipping precocious past year. AMD's 2025 spot volition beryllium called MI350, and its 2026 spot volition beryllium called MI400, the institution said.
The MI325X's rollout volition pit it against Nvidia's upcoming Blackwell chips, which Nvidia has said volition commencement shipping successful important quantities aboriginal adjacent year.
A palmy motorboat for AMD's newest information halfway GPU could gully involvement from investors that are looking for further companies that are successful enactment to payment from the AI boom. AMD is lone up 20% truthful acold successful 2024 portion Nvidia's banal is up implicit 175%. Most manufacture estimates accidental Nvidia has implicit 90% of the marketplace for information halfway AI chips.
AMD banal fell 3% during trading connected Thursday.
AMD's biggest obstacle successful taking marketplace stock is that its rival's chips usage their ain programming language, CUDA, which has go modular among AI developers. That fundamentally locks developers into Nvidia's ecosystem.
In response, AMD this week said that it has been improving its competing software, called ROCm, truthful that AI developers tin much easy power much of their AI models implicit to AMD's chips, which it calls accelerators.
AMD has framed its AI accelerators arsenic much competitory for usage cases wherever AI models are creating contented oregon making predictions alternatively than erstwhile an AI exemplary is processing terabytes of information to improve. That's partially owed to the precocious representation AMD is utilizing connected its chip, it said, which allows it to server Meta's Llama AI exemplary faster than immoderate Nvidia chips.
"What you spot is that MI325 level delivers up to 40% much inference show than the H200 connected Llama 3.1," said Su, referring to Meta's large-language AI model.
Taking connected Intel, too
While AI accelerators and GPUs person go the astir intensely watched portion of the semiconductor industry, AMD's halfway concern has been cardinal processors, oregon CPUs, that laic astatine the bosom of astir each server successful the world.
AMD's information halfway income during the June 4th much than doubled successful the past twelvemonth to $2.8 billion, with AI chips accounting for lone astir $1 billion, the institution said successful July.
AMD takes astir 34% of full dollars spent connected information halfway CPUs, the institution said. That's inactive little than Intel, which remains the brag of the marketplace with its Xeon enactment of chips. AMD is aiming to alteration that with a caller enactment of CPUs, called EPYC 5th Gen, that it besides announced connected Thursday.
Those chips travel successful a fig of antithetic configurations ranging from a low-cost and low-power 8-core spot that costs $527 to 192-core, 500-watt processors intended for supercomputers that outgo $14,813 per chip.
The caller CPUs are peculiarly bully for feeding information into AI workloads, AMD said. Nearly each GPUs necessitate a CPU connected the aforesaid strategy successful bid to footwear up the computer.
"Today's AI is truly astir CPU capability, and you spot that successful information analytics and a batch of those types of applications," Su said.
English (US) ·