I have been trying to use Jax on AMD/ROCM since 2023 and still can't because of the countless ROCM bugs and terrible response time of their team.
I wish AMD could succeed, I really do. I'm all for competition and having someone stand up to NVIDIA.
But 3 months I found that on AMD MI250x linear algebra operations like svd or eigh are 10x to 40x times slower than on NVIDIA A100, a gpu with 20% of the MI250x stated TFLOPs
I reported it, and I got a vague response. 3 months later the bug is still largely there, and AMD still does not offer a clear message.
I understand AMD is prioritising their ML customers who only use matrix multiplications, but as AMD somehow managed to convince France to buy HPC supercomputers from them, they should be honest and commit to a decent support.
Papers today are longer than ever and full of jargon and symbols. They depend on chains of computer programs that generate data, and clean up data, and plot data, and run statistical models on data. These programs tend to be both so sloppily written and so central to the results that it’s contributed to a replication crisis, or put another way, a failure of the paper to perform its most basic task: to report what you’ve actually discovered, clearly enough that someone else can discover it for themselves.
That's incorrect.
If you work with mid-sized neural networks and MCMC sampling, allocations start to play a significant role (And Flux.jl is bad at preallocation).
Prealloc.jl does not work properly. Zygote.jl adds even more allocations to the mix...
Jax/XLA completely solves this problem. Yes, it's annoying that you have to work with a static graph but if your problem fits the description... it's great.
I wish AMD could succeed, I really do. I'm all for competition and having someone stand up to NVIDIA.
But 3 months I found that on AMD MI250x linear algebra operations like svd or eigh are 10x to 40x times slower than on NVIDIA A100, a gpu with 20% of the MI250x stated TFLOPs I reported it, and I got a vague response. 3 months later the bug is still largely there, and AMD still does not offer a clear message.
I understand AMD is prioritising their ML customers who only use matrix multiplications, but as AMD somehow managed to convince France to buy HPC supercomputers from them, they should be honest and commit to a decent support.