Hacker News new | past | comments | ask | show | jobs | submit login

> "Extremely fast"

When people make claims like this, it would be good if they put the benchmarks on the first page. E.g, how does it compare with https://github.com/gfx-rs/wgpu which lets you target Vulkan, Metal, DX, GL or WASM+WebGPU with rust?




I hope this doesn't come off as handwaving, but you're kinda comparing apples to oranges here. CUDA has always been in a class of it's own when it comes to GPU compute, for better and worse, so I think the people out there who want to use this will pretty quickly get an idea of who it's for. Benchmarks would be nice, but I don't really think they'd be germane when comparing a proprietary compute system with a generic cross-platform GPU binding.


What? So you can just claim something is 'fast' with no evidence, but then if it is slower than proprietary things, who cares? That's ridiculous - if you claim something is 'extremely fast' you should back that up with benchmarks.


If it's not relative to anything, than the word "fast" doesn't have much meaning.


"Fast", to me, from a software development perspective can still be meaningful, in the sense of knowing what techniques, patterns, paths, ... enable performant execution, and providing easy and straightforward paths for the user along those. Which, ultimately, leads to high performance in most ways the user will apply the provided framework (in a more general sense of the word). Hope that makes sense.

And it must be OK to claim "fast" as a goal, from the early stages of a project, even before it may be possible to create any meaningful benchmarks. As long as it's discernable for the intended audience the precise stage of development or maturity the project is currently at. Which, I believe, the project in question is communicating just fine ("still in early development").


Would have it made any difference had the parent mentioned OpenCL?


wgpu is a library for running wgsl on GPUs, not Rust.


It turns out that when wgpu is combined with rust-gpu, it can run rust on gpus too

https://github.com/EmbarkStudios/rust-gpu/tree/main/examples...

(on the "builder" directory it builds shaders with the spirv-builder crate)


https://github.com/embarkstudios/rust-gpu would be the closest equivalent, AFAICT.


I wrote some OpenCL code recently, wrapped in a rust program using opencl3. My CL code was pretty slow early on, it's a math heavy problem implemented initially with a bunch of for loops.

I largely "optimized" it by smattering the code with #pragma unroll, which was exceptionally effective for the problem at hand, given that I had tons of statically defined loops that could be easily unrolled.

I know rust has a tendency to aggressively inline and flatten everything, but I'm curious about things like this, where it can be so important in this domain to ensure things are unrolled, as the loop conditionals can otherwise introduce pipeline/concurrency issues.


When they say "extremely fast GPU code" I think they mean relative to Rust's normal home: standard CPUs. So I don't think the claim needs any specific support.


To me, it is clear that "fast GPU code" means GPU code that is faster than other GPU code.

Like a "fast sports car" would obviously have to be faster than other sports cars. You couldn't (truthfully) make that claim if was faster than regular cars but slower than other sports cars.

"Accelerate Rust code by using the GPU" would be a more honest description.


I agree it could be read both ways. I'm just saying what I think their intent is.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: