Trying to get machine learning lib set up on a 290X

Question

I am having some difficulty finding a library with which to explore machine learning/ai. I have a pair of R9 290x's, and can't seem to find a lib which works well for it.

First I tried array-fire, which has excellent CPU performance, but poor GPU performance for machine learning, as demonstrated on the benchmarks in the machine_learning sample folder.

I looked into rocm and MIOpen, I tried the hip enabled tensorflow but found it is not supported on the 290x generations. I found someone working on llvm-amdgpu suppport for tensorflow as well, but it doesn't look ready yet

I looked into accelerate for haskell, and found an issue regarding the amdgpu backend, but it also looks not ready.

Maybe I haven't been searching broadly enough? But from what I can tell, almost everything runs on cuda, and I can't afford a new GPU for this right now.

Answer 1

At the time you asked the question, AMD did not support Hawaii GPU's with their rocm driver and compute stack. Since then support has been added for these older GPU's.

AMD has made a tensorflow port which installs and functions the same as CUDA tensorflow (amd's port). However it doesn't support anything older than gfx803 (Fiji, such as R9 Fury).

I have an R9 290 and it is works with the latest rocm drivers from AMD's repo, but not with the AMD tensorflow port. This is the error I get:

2018-08-16 12:10:58.529311: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1532] Ignoring visible gpu device (device: 0, name: Hawaii PRO [Radeon R9 290], pci bus id: 0000:01:00.0) with AMDGPU ISA gfx701. The minimum required AMDGPU ISA is gfx803.