On: AMD now has more compute on the top 500 than Nvidi...

[pjmlp]

What is the reasonable alternative to CUDA Fortran on AMD?

One example out of many I can point out from CUDA ecosystem.

[amelius → pjmlp]

People use CUDA through a limited number of libraries, for example Torch and Tensorflow, so there isn't a really strong dependence on CUDA for many important applications.

[pjmlp → amelius]

Some people working in machine learning, do use CUDA via Torch and Tensorflow.

[amelius → pjmlp]

Yes, most people in ML, and this field is currently on an exponential growth curve.

[pjmlp → amelius]

And a tiny percentage of why CUDA is as big as it is.

[my123 → pjmlp]

AMD ships a Fortran OpenMP compiler with GPU offloading that works pretty well

[pjmlp → my123]

Made public 6 days ago.

https://www.phoronix.com/news/AMD-Next-Gen-Fortran-Compiler

[my123 → pjmlp]

That's the next gen one. Older one based on classic Flang has been in production since quite a while.