Home » Skynet Smiles: AMD Unveils Monster Intuition MI300 Hybrid AI Accelerator

Skynet Smiles: AMD Unveils Monster Intuition MI300 Hybrid AI Accelerator

by Genzo
0 comment

This web site could earn affiliate commissions from the hyperlinks on this web page. Phrases of use.

At AMD’s CES keynote this week, Chairman and CEO Dr. Lisa Su seemingly saved the most effective for final. On the very finish, she held up an absolute monster of an APU, describing it as the primary knowledge heart chip that mixes a CPU, GPU, and reminiscence on a single bundle. It’s referred to as the AMD Intuition MI300, and it packs a whopping 146 billion transistors–46 billion greater than present in Intel’s refined knowledge heart chip, Ponte Vechio. AMD says the MI300 might be coming within the latter half of 2023, so for now it simply needed to point out it off a bit.

The chip is totally huge; when she held it aloft, it was the dimensions of her palm. The MI300 sports activities 24 Zen 4 cores together with its CDNA 3 GPU structure and 128GB of high-bandwidth reminiscence (HBM3). AMD didn’t point out what number of CUs had been getting used on the GPU facet. The CPU and GPU chiplets are 3D-stacked on high of a base die that connects them to the eight stacks of HBM. The design makes use of 9 5nm chiplets 3D-stacked onto 4 6nm chiplets. It’s unclear how the CPU and GPU chips are divided amongst the chiplets and AMD hasn’t revealed it but. The 4 6nm base dies are energetic interposers that deal with communications and I/O between the chiplets and the reminiscence stacks.

AMD says this built-in strategy permits for a major enhance in efficiency in addition to effectivity. For example, Dr. Su says it destroys the corporate’s earlier technology product, the Intuition MI250X. General it affords 8X the AI efficiency together with a 5X uplift in performance-per-watt. As well as, she stated the MI300 can scale back the coaching time for AI fashions from months to only weeks. This might theoretically save corporations tens of millions of {dollars} in electrical energy prices. The mixing of the CPU and GPU additionally permits them to work on the identical knowledge concurrently. Moreover, for the reason that HBM3 is built-in, it’d be capable of perform with out separate DRAM modules.

Though the corporate’s 4th Gen Epyc “Genoa” chips are broadly accessible and high-volume, that most likely received’t be the case with the MI300. AMD didn’t reveal pricing as a result of it’s not releasing it any time quickly, nevertheless it is not going to be low cost. It’s at present slated to energy the 2-exaflop El Capitan supercomputer at Lawrence Nationwide Laboratory. This pc might be powered on by the finish of 2023 and will affords double the efficiency of the world’s present quickest supercomputer, Frontier, which can also be AMD-powered. Frontier was the primary supercomputer to interrupt the exaflop barrier. Now it feels like El Capitan could possibly be the primary to 2 exaflops.

Now learn:


You may also like

Leave a Comment