December 1, 2021

AMD Unveils EPYC With 3D V-Cache, Beastly Twin-Die Intuition MI200 GPU For Huge HPC Workloads

For chip corporations, larger revenue margins will be discovered within the information middle market, and that is typically the place we see new improvements manifest first earlier than trickling into the buyer area. To wit, we all know AMD is prepping a Zen 3 refresh with stacked 3D V-cache. Whereas we patiently wait, AMD right this moment additionally simply launched third Gen EPYC processors with 3D V-cache. Mixed with its also-new Intuition MI200 GPU accelerators, AMD claims to be arming information facilities and excessive efficiency computing (HPC) clients with the arsenal they want for the exascale period.

That is the theme of right this moment’s unveiling—the necessity for exascale-capable {hardware} is now upon us, and AMD goals to fill that want with these new large iron merchandise. Let’s begin the dialogue with AMD’s upgraded EPYC server processors, which stay suitable with the SP3 socket (so it is a drop-in improve, primarily).

AMD’s third Gen EPYC lineup primarily based on Zen 3, now with 3D V-cache on board, continues to be being supplied with as much as 64 cores and 128 threads of huge iron muscle. Solely now they’re beefed up with 3 times the L3 cache, equating 804MB of complete cache per socket, for what AMD claims will ship an enormous 50 p.c common uplift throughout focused workloads. Thus far, the determine that is been related to 3D V-cache is 15 p.c, which is how a lot AMD mentioned gaming functions stand to achieve from its eventual Zen 3+ lineup.
AMD EPYC V-Cache Performance
Click on to Enlarge (Supply: AMD)
It is not nearly evaluating with its personal current product stack, although. AMD is claiming a boosted efficiency benefit over Intel’s Xeon processors. For instance, AMD says a 2P EYPC 75F3 system with 32 cores outpaces a 2P Xeon 8362 system (additionally with 32 cores) by as much as 33 p.c in Ansys Mechanical (Finite Aspect Evaluation), as much as 34 p.c in Altair Radioss (Structural Evaluation), and as much as 40 p.c in Ansys CFX (Fluid Dynamics).

AMD’s message is that its up to date EPYC processors are tailor constructed for technical computing workloads. The addition of 3D V-cache permits them to crunch by issues like crash simulations, chemical engineering, design verification, and the such at a a lot quicker tempo.

AMD Unveils Intuition MI200 Accelerators With Up To 128GB Of HBM2e

AMD Instinct MI200

Upgraded EPYC processors are solely a part of the equation. AMD additionally introduced new Intuition MI200 sequence accelerators which might be the primary to assist a multi-die GPU, and in addition the primary to assist a heaping 128GB of HMB2e reminiscence.

The brand new sequence is constructed on AMD’s CNDA 2 structure, and culminates within the Intuition MI250X. In accordance with AMD, the flagship accelerator delivers as much as 4.9X higher efficiency than “aggressive accelerators” for double-precision (FP64) HPC functions, and tops 380 TFLOPs of peak theoretical half-precision (FP16) for AI workloads.

“AMD Intuition MI200 accelerators ship management HPC and AI efficiency, serving to scientists make generational leaps in analysis that may dramatically shorten the time between preliminary speculation and discovery,” mentioned AMD’s Forrest Norrod. “With key improvements in structure, packaging, and system design, the AMD Intuition MI200 sequence accelerators are essentially the most superior information middle GPUs ever, offering distinctive efficiency for supercomputers and information facilities to unravel the world’s most advanced issues.”

AMD Instinct MI200 Performance
Click on to Enlarge (Supply: AMD)
The first competitors right here is NVIDIA and its Ampere structure. In accordance with AMD, the Intuition MI250X gives a “quantum leap in HPC and AI efficiency” over its rival’s A100 product.

As for the specs, the Intuition MI250X wields 220 compute items and 14,080 stream processors. It additionally options the total 128GB HBM2e reminiscence allotment, clocked at 1.6GHz and tied to a 8,192-bit bus, for as much as 3.2TB/s of reminiscence bandwidth. There’s additionally a non-X variant (Intuition MI250) with 208 compute items, 13,312 stream processors, and the identical reminiscence configuration.

AMD Instinct MI200

The multi-die design with AMD’s new 2.5D Elevated Fanout Bridge (EFB) expertise permits 1.8X extra cores and a pair of.7X increased reminiscence bandwidth in comparison with AMD’s earlier gen GPUs, the corporate says. It additionally advantages from as much as eight Infinity Cloth hyperlinks connecting the Intuition MI200 sequence to EPYC CPUs within the node, for a unified CPUK/GPU reminiscence coherency.

AMD’s overarching pitch is that information facilities can mix its newest accelerators with its upgraded EPYC chips and ROCm 5.0 open software program platform to “propel new discovers for the exascale period.” Particularly, AMD factors to issues like tackling local weather change and vaccine analysis, as two doable examples.

The Intuition MI250X and MI250 are each out there within the open-hardware compute accelerator module or OCP accelerator module kind issue, whereas the latter can even be supplied in a PCIe card kind consider OEM servers. Together with the upgraded EPYC chips, all of this may hit the market within the first quarter of subsequent 12 months.

Leave a Reply

Your email address will not be published. Required fields are marked *