<> Yup.

Deepest apologies. I was thinking about the CPU. Moving fast.

It is a good question, and I suggested that a few years back. I think IBM wants to sell its own AC922 and the wonder is why there are not larger and fatter Power Systems machines with NVSwitches. Gotta figure IBM doesn’t think there is margin in that, but I can’t imagine that this is correct. The other thing is that using Power9 chips would mean burning some of the NVLink ports for the GPU interconnect on the CPU-GPU interconnect, and that would reduce the GPU count in a single image. Maybe NVLink 2.0 is not as important for hosts as is PCI-Express 4.0 and PCI-Express switches to link CPUs to GPUs is, in fact, good enough.

Why doesn’t Nvidia use IBM POWER CPUs in their DGX servers? Are the host CPUs that unimportant?

With POWER9 scale up they could have 8TB in 2 sockets and have NVLink on the CPUs as well with no PCIe switches needed.

Please fix typo 🙂 (yes, i work for nvidia, but am not involved on any PR / marketing, i read this website purely out my own interest)

Very exciting times we are living in for HPC :).
Good article!