AMD is coming for Nvidia’s AI crown in an enormous approach with the launch of its new Intuition processor, which it claims can do the work of a number of GPUs.
CEO Lisa Su referred to as the Intuition MI300X “probably the most advanced factor we’ve ever constructed.” She held up the chip, which is concerning the measurement of a drink coaster, at an occasion on Tuesday in San Francisco.
Weighing in at 146 billion transistors, the MI300X comes with as much as 192GB of high-bandwidth HBM3 reminiscence shared by each the CPU and GPU. It has a complete of 13 chiplets on the die. The chip additionally has a reminiscence bandwidth of 5.2 TB/s, which is 60% sooner than Nvidia’s H100.
The chip consists of Zen CPU cores and AMD’s next-generation CDNA 3 GPU structure. The big quantity of reminiscence is the actual promoting level, based on Su.
“When you take a look at the business immediately, you typically see that, to begin with, the mannequin sizes are getting a lot bigger. And also you really need a number of GPUs to run the newest giant language fashions,” she stated. “With MI300X, you may cut back the variety of GPUs for this, and as mannequin sizes proceed rising, this can turn out to be much more necessary. So with extra reminiscence, extra reminiscence bandwidth, and fewer GPUs wanted.”
AMD says the design of the MI300X makes it eight instances extra highly effective than the present MI250X utilized in Frontier (the world’s quickest supercomputer) and 5 instances extra power environment friendly. It is going to be used within the two-plus exaFLOP El Capitan system that can be constructed subsequent 12 months at Lawrence Livermore Nationwide Labs.
As a part of the announcement, Su additionally unveiled the AMD Intuition platform, a server reference design primarily based on specs from the Open Compute Venture that makes use of eight MI300X GPUs for generative AI coaching and inference workloads.
This implies enterprises and hyperscalers can use the Intuition Platform to place MI300X GPUs in present OCP server racks.
“We’re truly accelerating prospects’ time to market and lowering total improvement prices, whereas making it very easy to deploy the MI300X into their present AI ramp and server development,” Su stated.
New cloud CPU
In different information, AMD mentioned its fourth-gen EPYC 97X4 processor, code-named Bergamo. This processor is particularly designed for cloud environments in that it has many cores for digital machines to run. Bergamo comes with 128 cores with hyperthreading, so a dual-socket system can have as much as 512 digital CPUs.
Su talked about how cloud-native workloads are “born within the cloud.” They’re designed to take full benefit of latest cloud computing frameworks, and so they primarily run as microservices. The design of those processors is completely different from common goal computing – Bergamo processors are smaller and really throughput-oriented, therefore the various core design.
“Bergamo leverages all the platform infrastructure that we already developed for Genoa. And it helps the identical subsequent gen reminiscence and the identical IO capabilities. But it surely permits us with this design level to broaden to 128 cores per socket for management efficiency and power effectivity within the cloud,” stated Su.
Each the MI300X and Bergamo will start sampling within the third quarter.
Copyright © 2023 IDG Communications, Inc.
