Where's your evidence that the Rosetta programs are written in a way that will allow such a conversion without using only graphics boards with extremely large numbers of GPU cores?

No evidence.
Mine is only a little demonstration that is possible to avoid CUDA, if necessary...

Possible for SOME programs. No evidence that the Rosetta programs are one of these.

Possible for SOME programs. No evidence that the Rosetta programs are one of these.

In the past there was no evidence that Rosetta runs on Nvidia gpu... :-P

In the past there was no evidence that Rosetta runs on Nvidia gpu... :-P

I remember when it was tried. They ran, but too slowly to be worthwhile.

I remember when it was tried.

Mee to.
But, for example, you have to consider that this thread started 15ys ago.
And 15years, in gpu field, is something like 5 geological eras, both hw and sw.

Code on gpu is a difficult art....

There is A LOT of tools to convert automatically (or semi-automatically) CUDA code to OpenCL/OneApi/SYCL/etc.
From AdaptiveCpp (using C++ as common language between Cuda/OpenCl/Sycl) to SpectralCompute, from HIPIFY to Coriander
But there are 2 problems:
- Nvidia could block your work (as they did with ZLUDA). I know, it's not a good thing for science to block others competitor, but CUDA is closed, so Nvidia can do it.

ZLUDA is back!! And now is focused on AI/Machine Learning.