Hacker News Reader

Comment by rhaps0dy | Hacker News Reader

Semi-relatedly and unfortunately, AI research is locked-in to NVIDIA (on Linux) because all the big lienar algebra and autodifferentiation frameworks work only with CUDA.

View on HN

AMD is working on ROCm https://rocm.github.io/ It will be an open source answer to CUDA and I hope this will now get more support.

"We are excited to present ROCm, the first open-source HPC/Hyperscale-class platform for GPU computing that’s also programming-language independent. We are bringing the UNIX philosophy of choice, minimalism and modular software development to GPU computing. The new ROCm foundation lets you choose or even develop tools and a language run time for your application."

Does anyone here know how fqr along this effort is? Is it usable today (in the sense that one could develop TF backend which uses ROCm)?

Not for long. Some of them actually work with AMD now, and with NVIDIA now banning use of GeForce cards for AI research (https://news.ycombinator.com/item?id=15983587), NVIDIA is going to lose that market very quickly.

This is totally not my field so it may very well be the stupidest question ever: is there technically anything that prevents implementing CUDA on top of non-NVIDIA hardware?

CUDA is made by nvidia

Not sure why you were down voted. CUDA is developed by NVIDIA.

CUDA is a registered trademark by NVIDIA - https://trademarks.justia.com/850/30/cuda-85030071.html

Here is the CUDA webpage

https://developer.nvidia.com/cuda-zone

Because it's a pointless answer that basically just repeats part of the question, providing absolutely no new information.

Well I think it makes a short answer NVIDIA doesn't let their proprietary CUDA technology work on non-NVIDIA devices.

No there isn't, Google developed a CUDA frontend for clang/llvm which targets PTX (nvidias virtual machine instruction set), in principle someone could come along and implement a different backend for this frontend targeting for example AMDs gpu instruction architecture or something different. The other ingredient missing is a reimplementation of CUDAs runtime libraries.

The last part is particularly important. Most deep learning libs rely on things like cuDNN which implement by hand basic operations very efficiently.

The ROCm initiative by AMD already "transpile" most CUDA kernels to being AMD compatible, but it seems performance is not yet comparable in real word benchmarks.

AMD is working on a tool to convert CUDA code: https://github.com/ROCm-Developer-Tools/HIP

We are working on an OpenCL based layer that enables AMD GPUs, see these links for more info https://developer.codeplay.com/computecppce/latest/getting-s... https://github.com/tensorflow/tensorflow/issues/22

Reply on news.ycombinator.com