Nvidia taps Intel’s Sapphire Rapids CPU for Hopper-powered DGX H100

A win against AMD as a much bigger war over AI compute plays out


Nvidia has chosen Intel's next-generation Xeon Scalable processor, known as Sapphire Rapids, to go inside its upcoming DGX H100 AI system to showcase its flagship H100 GPU.

Jensen Huang, co-founder and CEO of Nvidia, confirmed the CPU choice during a fireside chat Tuesday at the BofA Securities 2022 Global Technology Conference. Nvidia positions the DGX family as the premier vehicle for its datacenter GPUs, pre-loading the machines with its software and optimizing them to provide the fastest AI performance as individual systems or in large supercomputer clusters.

Huang's confirmation answers a question we and other observers have had about which next-generation x86 server CPU the new DGX system would use since it was announced in March.

The GPU giant has previously promised that the DGX H100 [PDF] will arrive by the end of this year, and it will pack eight H100 GPUs, based on Nvidia's new Hopper architecture. By using its fourth-generation NVLink interconnect to connect the GPUs, the chip designer has claimed that a single system will be capable of delivering 32 petaflops of AI performance using its FP8 format.

Huang confirmed Nvidia's selection of Sapphire Rapids for the DGX H100 while voicing his continued support for x86 CPUs as the company plans to introduce its first Arm-based server CPU, Grace, next year. He also said that Nvidia will use Sapphire Rapids for new supercomputers.

"We buy a lot of x86s. We have great partnerships with Intel and AMD. For the Hopper generation, I've selected Sapphire Rapids to be the CPU for Nvidia Hopper, and Sapphire Rapids has excellent single-threaded performance. And we're qualifying it for hyperscalers all over the world. We're qualifying it for datacenters all over the world. We're qualifying it for our own server, our own DGX. We're qualifying it for our own supercomputers," he said at the Tuesday event.

The selection of Intel's upcoming Sapphire Rapids chip, which has already started shipping to some customers, marks a reversal of sorts for Nvidia after it chose AMD's second-generation Epyc server CPU, code-named Rome, for its DGX A100 system that was introduced in 2020.

This comes after industry publication ServeTheHome reported in mid-April that Nvidia had motherboard designs for both Sapphire Rapids and AMD's upcoming Epyc CPU, code-named Genoa, for the DGX H100 as the GPU giant had not yet decided on which x86 chip it would use.

While Intel will consider this a victory as the semiconductor giant works to regain technology leadership after years of missteps, it's a relatively small win when considering the bigger battle over GPUs and other accelerators that is playing out between Nvidia, Intel, AMD and other companies. It's why, for instance, Intel is making a big bet on its upcoming Ponte Vecchio GPU and why AMD has pushed to become more competitive against Nvidia with its latest Instinct GPUs.

One major reason why Nvidia has decided to build its own Arm-compatible CPU is so it can put a CPU and a GPU together in the same package to significantly speed up the flow of data between the two components to fuel AI workloads and other kinds of demanding applications.

Nvidia plans to introduce its first iteration of this design, called the Grace Hopper Superchip, next year alongside the 144-core, CPU-only Grace Superchip, and we think it's likely that Nvidia will introduce a new kind of DGX system that will use Grace. Intel also plans to introduce a CPU-GPU design for servers with the Falcon Shores XPU in 2024.

During the Tuesday talk, Huang promised that "Grace is going to be an amazing CPU" that will allow the Nvidia to fine-tune everything from the components to the systems to the software. While the GPU giant is designing the Arm-compatible chip to benefit recommender systems and large language models used by so-called hyperscale companies, it will be used for other applications too, according to Huang.

"Grace has the advantage that in every single application domain that we go into, we have the full stack, we have all of the ecosystem all lined up, whether it's data analytics, or machine learning, or cloud gaming, or Omniverse, [or] digital twin simulations. In all of the spaces that we're going to take Grace into, we own the whole stack, so we have the opportunity to create the market for it," he said. ®


Other stories you might like

Biting the hand that feeds IT © 1998–2022