Nihal Pasham

🦀 Rust Tech | @Nvidia | Make general purpose GPU programming accessible 🖖 Disclaimer: The views, opinions expressed are my own (not my employer's)

[Nihal Pasham @npashi]

Finally able to talk about what I've been heads-down on for 6 months at @nvidia 🦀⚡

We just open-sourced cuda-oxide — an experimental rustc backend that lets you write CUDA kernels in pure Rust.

No DSLs. No FFI. No source-to-source step. Single source.

Short🧵👇

[Leo Alt @leonardoalt]

We can now fully rewrite most software in @leanprover and prove it correct:

- Compiler module rewrite (AI) from Rust to Lean
- Full FFI integration
- All unit and integration tests pass
- Formal spec and proofs!!
- Under 20h wall time (unnoticed pauses)

https://t.co/u601dZ8wph

[Nihal Pasham @npashi]

@Nekrolm @roeschinc @nvidia @Nekrolm - the index_2d bug-fix landed. The latest vecadd example uses the more ergonomic api - get_mut_indexed

and if you'd like to take a look at the entire fix - https://t.co/PssDX87Q8I, feel free to reopen if you have thoughts.

Thanks for flaggin this again.

[Nihal Pasham @npashi]

@Nekrolm @roeschinc @nvidia I'll push out a branch to help reason about this more concretely

[Nihal Pasham @npashi]

@hazle111753854 @nvidia @VaivaswathaN We haven't done an actual comparison yet but so far the experience is - its on par with Cuda C++.

Ps: needs thorough validation though.

[Nihal Pasham @npashi]

Finally able to talk about what I've been heads-down on for 6 months at @nvidia 🦀⚡

We just open-sourced cuda-oxide — an experimental rustc backend that lets you write CUDA kernels in pure Rust.

No DSLs. No FFI. No source-to-source step. Single source.

Short🧵👇

[Nihal Pasham @npashi]

@ibuildthecloud 😅

[Nihal Pasham @npashi]

@shareastronomy Apparently communities are being deprecated.

[Nihal Pasham @npashi]

I know our community won’t be around come June (as ‘X’ has other plans).

But before we go — I thought I’d drop something I’ve been working on for a while here.

[Nihal Pasham @npashi]

Finally able to talk about what I've been heads-down on for 6 months at @nvidia 🦀⚡

We just open-sourced cuda-oxide — an experimental rustc backend that lets you write CUDA kernels in pure Rust.

No DSLs. No FFI. No source-to-source step. Single source.

Short🧵👇

[Nihal Pasham @npashi]

@JamesTervit @nvidia @VaivaswathaN This is great. Please feel free. Would love to hear feedback.

[Nihal Pasham @npashi]

@HiSohan 😆

[Nihal Pasham @npashi]

@dr_sensor Just a small correction. Kernel code doesn’t support async/await. But we can do async kernel launches. We have a few examples that use the #tokio runtime to launch GPU work asynchronously.

[Nihal Pasham @npashi]

@mr_r0b0t @entrepeneur4lyf @nvidia @VaivaswathaN @perplexity_ai Interesting, I didn’t know they had one. Would you happen to have a link?

[Nihal Pasham @npashi]

@WebDevCaptain @nvidia @VaivaswathaN @grok I’ll lend @grok a hand here - https://t.co/572j2R26hy

[Nihal Pasham @npashi]

@VaivaswathaN 💯 - I can say this after having worked with it over the past few months. Pliron addresses all of cuda-oxide's needs - extensibility, rust-native and a breeze to debug. Love it!

[Nihal Pasham @npashi]

@AstraKernel
@rustoftheday
@rustaceans_rs
@ThisWeekInRust

[Nihal Pasham @npashi]

46 worked examples shipped — async MLP, cross-crate kernels, Rust ↔ C++/CCCL device FFI.

Still alpha + under active development. So, expect bugs, missing features, API churn — we think its a good start.

⭐ https://t.co/ongDJBi7Ew
📖 https://t.co/OpYHBUd8YP

[Nihal Pasham @npashi]

🌟 Highlights:

🧱 Custom rustc → PTX backend
⚡ Generics, closures, structs, enums on the GPU
🔧 Full intrinsics: warp, shared mem, atomics, clusters
📈 GEMM (naive): 868 TFLOPS
🌐 Composable async, .await on GPU work
🦀 Pliron pipeline: Rust → MIR → Pliron → LLVM → PTX

[Nihal Pasham @npashi]

@Nekrolm @roeschinc @nvidia But I do think, it is actually more ergonomic (i.e. rusty) option if perf is not an issue. However, on another note - I'm working on a oxide-native MIR Inliner pass, which should solve the perf problem.

[Nihal Pasham @npashi]

@Nekrolm @roeschinc @nvidia Tried that first. But perf cost adds up quickly for trivial index arithmetic -- at the method-call boundary the no-param `get_mut()` doesn't inline, so the index is computed twice plus call overhead. So, went with ThreadIndex version.