Announcing Burn-Wgpu: New Deep Learning Cross-Platform GPU Backend

Over the past year, I've been hard at work on Burn. Now that the architecture has stabilized, I'm happy to announce the release of Burn version 0.8.0, featuring a brand new GPU backend built entirely in Rust using wgpu.

This release brings numerous improvements, but the most significant one is undoubtedly the introduction of the new GPU backend. We've discussed the project's motivation and details in our latest blog post, so feel free to check it out: https://burn-rs.github.io/blog/cross-platform-gpu-backend.

I want to express my gratitude to the community for contributing to the development of wgpu. I believe that using it with Burn will be highly beneficial. While our primary focus was on ensuring correctness and setting up the GPU computing pipeline, I'm thrilled to see that we've achieved reduced memory usage compared to the previous LibTorch backend. It's worth mentioning that the GPU backend is still currently slower than CUDA implementations, but I remain optimistic about the potential for further performance optimizations. I'm curious to see if the community will come up with creative speed-ups to push the boundaries even further.

For more details about the latest release see the release notes: https://github.com/burn-rs/burn/releases/tag/v0.8.0.