News for hackers. Stuff that matters.

A comment on MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUDA Moat Still Alive. Parent comment

Re:MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUD... by nemothekid on Monday December 23, 2024 @12:07AM (#42491940)
>Hyperscalers are also more self-sufficient at software: they have entire teams working on PyTorch, Jax, and writing kernels.

None of this matters because AMD drivers are broken. No one is asking AMD to write a PyTorch backend. The idea that AMD will have twice the silicon performance than nvidia to make up the performance loss for bad software is a pipedream.

[ Parent | on HN ]
Re:MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUD... by sangnoir on Monday December 23, 2024 @01:39AM (#42492318)
> None of this matters because AMD drivers are broken

Do you honestly think the MI300 has show-stopper driver bugs, or that Meta/Amazon doesn't have a direct line to AMD engineers?

[ Parent | on HN ]
Re:MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUD... by nemothekid on Monday December 23, 2024 @07:20PM (#42498649)
>Do you honestly think the MI300 has show-stopper driver bugs

Yes

>Meta/Amazon doesn't have a direct line to AMD engineers?

I don't even think AMD engineers have a direct line to AMD.

[ Parent | on HN ]
Re:MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUD... by sangnoir on Monday December 23, 2024 @07:57PM (#42498830)
The fine article says they had the ears of VPs and Principal Engineers, just to help with a benchmark. Meta/Amazon will get white-glove service
[ Parent | on HN ]
Re:MI300X vs. H100 vs. H200 Benchmark Part 1: Training – CUD... by david-gpu on Monday December 23, 2024 @06:47AM (#42493729)
> None of this matters because AMD drivers are broken

How do you know that the problems arise from broken drivers rather than broken hardware? Real world GPU drivers are full of workarounds for hardware bugs.

[ Parent | on HN ]