5 ms·
> The problem for them is the CUDA moat. Compared to the problem of developing a new cutting-edge GPU, building a CUDA compatibility layer is a much smaller pr
by JoshTriplett 2y ago
> The problem for them is the CUDA moat.
Compared to the problem of developing a new cutting-edge GPU, building a CUDA compatibility layer is a much smaller problem. Hire the author of ZLUDA, throw a small team at it, and have a legal department on standby. And separately, there'd also be value in some source-translation projects to help people migrate to some better native framework.
- vidarh 2y agoBoth in terms of hardware and software, they need to be able to support the small set of operations that their training or inference uses, so the problem on both sides is much smaller than both a full GPU and a full CUDA replacement.
- codedokode 2y agoWhy support proprietary libraries? Isn't it better to make an open-source library. Also as for AMD, as I understand, they are unwilling to make ML libraries for consumer-grade GPUs and GPUs built into CPUs.
- JoshTriplett 2y agoAn Open Source compatibility layer for CUDA to allow people to run on non-NVIDIA GPUs is a first step; it removes the lock-in between GPU and library. Once people can run all their existing software on a different GPU, they can then consider adopting a better standard to build on top of. The reverse approach, of trying to entice people to move from CUDA to a different library and switch GPUs at the same time, has been tried repeatedly and has not yet succeeded. Trying something different seems warranted.