During the Bank of America Securities 2022 global technology conference, nvidia CEO huangrenxun finally confirmed that he would choose Intel sapphire rapids Xeon processor to support his AI system based on DGX H100 hopper GPU At the earlier GTC 2022 conference, the company only disclosed that it would use x86 CPU when releasing the DGX H100 new system, without further revealing whether it would choose to cooperate with AMD or Intel.
（via WCCFTech）
Huangrenxun said at the BofA securities 2022 Global Technology Conference on Tuesday:
We have purchased many x86 products and established good partnership with Intel and AMD.
However, given NVIDIA's commitment to making new products applicable to large-scale computers and data center servers around the world, Intel sapphire rapids with excellent single thread performance has become our first choice.
Interestingly, this AI giant switched from Intel to AMD platform solution on the DGX A100 system based on ampere GPU.
As a reference, the DGX A100 is equipped with a dual AMD epyc Rome CPU (up to 64c / 128t SKU).
As for which Xeon SKU from Intel will be used in DGX H100, it is not known at present.
However, even if sapphire rapids only provides the option of 56C / 112t at most, the single thread performance of golden Cove architecture can still be reversed.
On the other hand, if Intel focuses on Multithreading performance, the AMD epyc roadmap is worth looking forward to -- especially the upcoming Zen 4 Genoa product line provides the option of up to 96c / 192T.
Previously, server the home reported that NVIDIA has two platform configurations of DGX H100 around amd / Intel server CPU. As for whether the "exclusive" of sapphire rapids will be abandoned in the future, it still needs time to test.
In terms of specifications, NVIDIA DGX H100 system is equipped with 8 H100 acceleration cards in SXM configuration, supplemented by hbm3 high bandwidth cache up to 640 GB / 32 PFlops peak computing performance.
In contrast, the existing DGX A100 system is equipped with 8 A100 GPUs and 640 GB hbm2e cache, and the AI / int8 peak computing performance is only 5 / 10 PFlops.
Sapphire rapids Xeon CPU provides complete PCIe 5.0 support for the platform, and can provide four-way DisplayPort @ 4K output through DGX independent display adapter.
In addition, the system is equipped with four nvswitch conversion chips, 2TB system memory, 1.9 TB nvme m.2 storage for two systems and 8 built-in 3.84 TB nvme U.2 SSDs.
The DGX H100 chooses to install the AIC expansion card with active cooling on the back of the chassis, and the four GPUs + CPUs are supported by a quiet, maintenance free refrigerant cooling scheme.
The compressor body is also inserted into the DGX chassis, and the nominal peak power consumption is as high as 10.2kw. Finally, according to the plan, DGX H100 is expected to be officially launched in the third quarter of 2022.