Hacker News Reader

Training on 1k+ v5p and h100 this isn’t true. We get 400 tflops per chip on v5p int8 and 650 fp8 for h100s

H100s are less stable training though.

V5ps are about the same as intel gaudi2s and both are cheaper than h100s price/perf

Gaudi3s are 3-4x faster…

Do you know how Intel Gaudi [1] could be purchased. I see a "Contact us" in their place but any inside or public information about prices and the budget required for bulk purchases would be great.

Also, how Gaudi3 compares to Google TPUs? Thanks! I know that Gaudi3 is not readily available yet.

[1] https://habana.ai/products/gaudi/

Is there enough volume of this chip for it to mean anything to anyone? If there are only a few that no one can’t buy, and they are only better than nvidia at 1 out of 10’000 workloads, can people even use their existence to try to get lower H100 prices?

I am currently analyzing the feasibility and economics of building a cloud service for AI training. This involves "day to day" analysis changes since all the "parts" are moving.

https://www.supermicro.com/en/products/system/AI/4U/SYS-420G...

Reply on news.ycombinator.com