NVIDIA Opts For AMD EPYC Rome CPUs Over Intel Xeon For DGX A100 Ampere Supercomputers
The big tech news this week is the official unveiling of Ampere, following months and months (years, really) of rumors. From the confines of his swank kitchen, NVIDIA CEO Jensen Huang dished up some details about the next-gen GPU, at least as it pertains to the data center high performance computing (HPC) markets. Following the announcement, AMD CEO Dr. Lisa Su revealed another interesting nugget—NVIDIA's DGX A100 servers are packing EPYC "Rome" processors inside rather than Intel Xeon chips.
This is yet another feather for AMD's cap, which has been amassing them ever since Zen arrived (and in particular, the latest iteration, Zen 2).
"Proud to have AMD EPYC 'Rome' CPUs in the new DGX servers. Congrast NVIDIA on the new A100 launch and thanks for selecting AMD as your CPU," Dr. Su said.
What this amounts to is the pairing of two 7-nanometer juggernauts. On the GPU side, Ampere is an absolute beast, both in sheer size at 826mm2, and in its technological prowess—it packs a whopping 54 billion transistors, along with 6,912 FP32 CUDA cores, 432 Tensor cores, 108 streaming multiprocessors, and 40GB of HBM2e memory.
"NVIDIA A100 GPU is a 20x AI performance leap and an end-to-end machine learning accelerator — from data analytics to training to inference," said NVIDIA founder and CEO Jensen Huang. "For the first time, scale-up and scale-out workloads can be accelerated on one platform. NVIDIA A100 will simultaneously boost throughput and drive down the cost of data centers."
Meanwhile, AMD's latest generation EPYC processors supplement the servers with the architectural benefits of Zen 2, which is also produced on a 7nm manufacturing process. One of the big draws is PCI Express 4.0 support. This essentially made EPYC the only viable choice to pair with PCIe 4.0-enabled Ampere GPUs.
It's another bragging right for AMD, as evidenced by Dr. Su's tweet. At the same time, this design win will likely mean a lot less in terms of actual revenue for AMD, since there will only be two 64-core/128-thread EPYC 7742 CPUs in each DGX chassis. You can think of them as basically traffic cops in a mesh of GPUs. We don't anticipate AMD seeing a big spike in revenue over this deal, but we'll see. If nothing else, it is another strong validation for AMD in the high margin data center business.