Since NVIDIA owns huggingface now and huggingface has the excellent Candle [1] crate for inference on Rust, this seems like a good step towards nice native Rust kernels.
Nobody cares if kernels are written in Rust. Kernels were meant to be written in C, but if you want to go more high-level try Triton or a similar DSL that nicely abstract tile sizes etc.
In this age of LLM written everything which has softly killed my motivation for learning Rust somewhat, this has revived my interest if not only for the fact the LLMs haven't yet been trained on this yet!
I strongly dislike CUDA. Once you have allowed that proprietary cr*p into your C++ codebase, it is very hard to get rid, and you end up with code that is either tied to a single vendor or an #ifdef hell, probably both.
The best way to program GPUs is face up to the reality that they are not the same machine as the CPU, write your kernels in separate files, and launch them manually, like in Metal, OpenCL, and D3D12, etc.
These days we even have DSLs like Triton that make kernel writing much more ergonomic than anything you would hope to achieve in Rust.
> The best way to program GPUs is face up to the reality that they are not the same machine as the CPU, write your kernels in separate files, and launch them manually
I'm looking forward to trying these when they stabilize! I currently use WGPU for graphics, and cudarc for CUDA.
Note: Cuda-oxide is similar to Cudarc's host component, but uses a rust-style kernel dialect. Advantage: Share structs between host and device. Disadvantage: Trading standard Cuda kernels for a new, WIP dialect.
I haven't tried the tile API yet; looking forward to it.
The last time I checked, Cuda Oxide was Linux only, and required Async; these are why I haven't tried it yet.
cudarc been great for me, because it's easy to look up existing examples and references, and it maps 1-to-1 with what I see. I'm already having a tough time with CUDA itself, a dialect of it makes a tad harder to rely on previous work.
Seems more ergonomic in general though, both approaches they share, compared to cudarc, and less build infrastructure and fiddling with environments, which is great.
what this article tells me is that no one at Nvidia actually cares about this project whatsoever. otherwise, they would have had a person actually write the announcement.
First of all, this is a pre-1.0 release that requires a nightly Rust compiler (if you choose the SIMT track with cuda-oxide) so that one is going to be unstable software.
Secondly, When an issue occurs with a kernel or you want to write your own custom kernel in Rust, now we need to diagnose if the problem came from either cuda-oxide (SIMT), Rust's side, CUDA or Tile (If you decide to choose the Tile track).
Another dependency into the list and course everything is open source except CUDA itself. So any issue that happens on the CUDA level, you are forced to wait for them to fix it.
Damn even Nvidia is putting out fully Claude-written articles.
[1] https://github.com/huggingface/candle
The best way to program GPUs is face up to the reality that they are not the same machine as the CPU, write your kernels in separate files, and launch them manually, like in Metal, OpenCL, and D3D12, etc. These days we even have DSLs like Triton that make kernel writing much more ergonomic than anything you would hope to achieve in Rust.
Isn't that how CUDA code is normally written?
Note: Cuda-oxide is similar to Cudarc's host component, but uses a rust-style kernel dialect. Advantage: Share structs between host and device. Disadvantage: Trading standard Cuda kernels for a new, WIP dialect.
I haven't tried the tile API yet; looking forward to it.
The last time I checked, Cuda Oxide was Linux only, and required Async; these are why I haven't tried it yet.
Seems more ergonomic in general though, both approaches they share, compared to cudarc, and less build infrastructure and fiddling with environments, which is great.
Secondly, When an issue occurs with a kernel or you want to write your own custom kernel in Rust, now we need to diagnose if the problem came from either cuda-oxide (SIMT), Rust's side, CUDA or Tile (If you decide to choose the Tile track).
Another dependency into the list and course everything is open source except CUDA itself. So any issue that happens on the CUDA level, you are forced to wait for them to fix it.