AMD desires individuals to keep in mind that Nvidia’s not the one firm selling AI chips. It’s introduced the provision of latest accelerators and processors geared towards working massive language fashions, or LLMs.
The chipmaker unveiled the Intuition MI300X accelerator and the Intuition M1300A accelerated processing unit (APU), which the corporate stated works to coach and run LLMs. The corporate stated the MI300X has 1.5 occasions extra reminiscence capability than the earlier M1250X model. Each new merchandise have higher reminiscence capability and are extra energy-efficient than their predecessors, stated AMD.
“LLMs proceed to extend in dimension and complexity, requiring huge quantities of reminiscence and compute,” AMD CEO Lisa Su stated. “And we all know the provision of GPUs is the only most essential driver of AI adoption.”
Su stated throughout a presentation that MI300X “is the very best performing accelerator on the earth.” She claimed MI300X is akin to Nvidia’s H100 chips in coaching LLMs however performs higher on the inference aspect — 1.4 occasions higher than H100 when working with Meta’s Llama 2, a 70 billion parameter LLM.
AMD partnered with Microsoft to place MI300X in its Azure digital machines. Microsoft CTO Kevin Scott, a visitor throughout Su’s speech, additionally introduced the Azure ND MI300X virtual machines — first revealed in November — at the moment are out there on preview. Meta additionally introduced it would deploy MI300 processors in its information facilities.
Su stated AMD launched the MI300A APU for information facilities, which she stated are anticipated to develop its complete addressable market to $45 billion. APUs usually mix CPUs and GPUs for quicker processing. AMD stated the MI300A gives higher-performance computing, quicker mannequin coaching, and a 30 occasions vitality effectivity enchancment. In comparison with the H100, AMD stated it has 1.6 occasions the reminiscence capability. It additionally options unified reminiscence, so there is no such thing as a want to maneuver information from totally different units anymore.
MI300A will energy the El Capitan supercomputer constructed by Hewlett Packard Enterprise on the Lawrence Livermore Nationwide Laboratory. El Capitan is taken into account probably the most highly effective supercomputers and is predicted to ship greater than two exaflops of efficiency.
The MI300A APU “is now in manufacturing and is being constructed into information facilities.”
Pricing data was not instantly out there.
Su teased the MI300 chips throughout the Code Convention, saying AMD was excited in regards to the alternative to faucet extra chip customers, not simply from cloud suppliers however from enterprises and startups.
AMD additionally introduced the newest addition to its Ryzen processors, the Ryzen 8040, which might put extra native AI features into cell units. The corporate stated the 8040 sequence gives 1.6 occasions extra AI processing efficiency than earlier fashions and integrates neural processing models (NPUs).
The corporate stated Ryzen 8040 wouldn’t be restricted to AI processing, because it claimed video modifying can be 65 p.c quicker and gaming can be 77 p.c quicker than with competing merchandise like Intel’s chips.
AMD expects producers like Acer, Asus, Dell, HP, Lenovo, and Razer to launch merchandise integrating Ryzen 8040 chips within the first quarter of 2024.
Su stated the subsequent technology of its Strix Level NPUs will probably be launched in 2024.
AMD additionally introduced the Ryzen AI Software program Platform is now broadly out there, which can let builders constructing AI fashions on Ryzen-powered laptops offload fashions into the NPU so the CPU can cut back energy consumption. Customers will get assist for basis fashions just like the speech recognition mannequin Whisper and LLMs like Llama 2.
To energy AI fashions — and make the most of the present hype for the tech — firms like AMD, Nvidia, and Intel have been locked in what’s mainly an AI chip arms race. Thus far, Nvidia captured the biggest market share with its extremely coveted H100 GPUs used to coach fashions like OpenAI’s GPT.