Tony Wu (tony.wu.delete@this.usa.com) on June 17, 2022 6:37 pm wrote:
> If Apple has ambitions in STEM, what they should do is make a Mac Pro with an optional Nvidia H100. Ideally,
> the H100 should be in a different form factor than a PCIe card so that it can have an enormous heat sink with
> a cross section of about 7” x 7” (limited by width of chassis). The H100 has no graphics output. It is
> a compute engine only, with 48 Tflops of FP64 and 800 Tflops of FP16/BF16 without sparsity. The only thing
> better than that for STEM, AI and data science applications would be to connect the Apple Silicon to the H100
> with NVLink. Apple could sell this system to every data scientist, AI practitioner, app developer who is using
> their Neural Engine, mechanical/civil engineer and every scientist whose company is willing to pay $30k+ for
> a workstation. Please note that the most expensive configuration of the current Mac Pro is $55k.


Why would a Mac Pro with an H100 be better than a PC with an H100? If you're buying an H100 to do computation on, what supports it becomes less important by definition.

Wouldn't Apple rather try to beat H100 performance themselves?

IMHO it is highly unlikely the Mac Pro supports any full sized PCIe cards, and while I expect it to have some x4 slots I would not be totally shocked if it does not have any slots at all. Apple has already said they will not support 3rd party GPUs with ARM Macs, so the only possible uses for PCIe expansion will be storage and networking.