Nvidia Huang Renxun: DGX H100 AI System Selects Intel Sapphire Rapids Processor

Nvidia has confirmed that it has selected Intel’s next-generation Xeon processor, Sapphire Rapids, to power its upcoming DGX H100 AI system, according to foreign media reports. , which is the world’s first AI platform based on the new NVIDIA H100 Tensor Core GPU.

Nvidia co-founder and CEO Jensen Huang confirmed the news at the BofA Securities 2022 Global Technology Conference (GTC) on Tuesday.Nvidia positions the DGX family as its go-to tool for data center GPUs, pre-installing its software in computers and optimizing itto deliver the fastest artificial intelligence (AI) performance as a single system or in large supercomputer clusters.

In fact, since Nvidia released its next-generation DGX system in March, there has been speculation about which x86 server CPU the new DGX system will use. Now, people finally have the answer.

Nvidia has previously promised that the DGX H100 will be available by the end of this year and will be equipped with eight H100 graphics processing units (GPUs) based on the next-generation Hopper architecture. Nvidia claims that by using its fourth-generation NVLink connectivity technology, AI performance can reach 32 Petaflops (petaflops) at the new FP8 precision.

This time around, Nvidia chose the Sapphire Rapids processor for its DGX H100 system, backing up the company’s promise to “continue to support x86 CPUs,” even though Nvidia plans to launch its first ARM-based server processor, “Grace,” next year. . In addition, Huang Renxun also said that Nvidia will also use Sapphire Rapids to support its new supercomputer.

“We buy a lot of x86 processors, and we have a close partnership with Intel and AMD,” Huang said on Tuesday. “For the Hopper generation, I chose Sapphire Rapids as its CPU, which has excellent single-threaded performance.”

This time, Nvidia chose Intel’s Sapphire Rapids processors. In contrast, Nvidia has opted for AMD’s second-generation Epyc server CPUs (codenamed Rome) in its DGX A100 systems in 2020.

This comes after industry publication ServeTheHome reported in mid-April that Nvidia’s next-generation DGX systems had yet to decide which processor to use. To this end, NVIDIA designed the DGX H100 motherboard for both Intel Sapphire Rapids and AMD Epyc processors.

Currently, Intel is trying to regain its technological leadership in the CPU market. This time around, winning the Nvidia order is clearly a win. But that’s a relatively small victory, considering the battle for bigger GPUs and other accelerators is being fought by Nvidia, Intel, AMD, and others. That’s why Intel is betting big on the upcoming Ponte Vecchio graphics processing unit (GPU), for example, and why AMD is pushing to compete with Nvidia with its latest Instinct GPU.

And one of the main reasons Nvidia decided to build its own ARM-compatible CPU was that it could package the CPU and GPU together to significantly speed up the data flow between the two components to support AI workloads and other types of “demanding” s application.

According to reports, Nvidia plans to launch the first generation of products based on the ARM architecture CPU next year, called “Grace Hopper Superchip”. Nvidia is also expected to introduce a new DGX system that uses Grace. Meanwhile, Intel plans to launch the Falcon Shores XPU in 2024, which will include both x86 CPU cores and Xe GPU cores.