Amdgpu, Rocm And Pytorch

AMDGPU

ROCM

Pytorch

GPU acceleration

Vulkan

OpenGL

DirectML

This article explores utilizing AMDGPU, ROCM, and Pytorch for GPU acceleration. It proposes leveraging Vulkan or OpenGL over ROCM, locating the most current build in Pytorch’s nightly repository, and integrating DirectML with Windows. However, it also cautions that ROCM is not compatible with integrated graphics cards and offers limited compatibility for AMD cards.

Published

October 10, 2023

https://scale-lang.com/

rocm is trash to apu. use vulkan or opengl (alternative backends)

find latest build in pytorch nightly repo

use sudo before invoking amdgpu related commands, otherwise unavailable

use directml (pytorch, tensorflow) instead. need windows.

you need to set the shared graphical memory to a larger value.

rocm does not work for integrated graphic cards (even if you build it and configured the card name), and supported amd cards are limited.