CUDA Tile

Headlines · 1

1. Oct 19
   Jane Street and NVIDIA Release Mamba-2 cuTile Kernels

   Jane Street engineers and NVIDIA implemented three custom kernels for the Mamba-2 forward pass in CUDA Tile (cuTile), a Python-based DSL and parallel programming model for tile-level NVIDIA GPU programming. NVIDIA reports that the kernels were 1.14x to 1.82x faster than an open-source Triton implementation in tests on a single NVIDIA B300 SXM6. The results measure kernel performance only; end-to-end training performance will also depend on backward-pass implementations.

   NVIDIA Technical Blog · 🔥 0

Experience and discussion from the community

Share my CUDA Tile experienceAsk about CUDA Tile

All

cuda-rust•4d Pinned

cuda-rust: NVIDIA's CUDA platform for Rust, providing host-side runtime crates and GPU kernel programming models in idiomatic Rust.

What problem it solves Rust developers need host-side access to the CUDA runtime and tools for expressing GPU kernels while preserving Rust's ownership and borrowing rules. The platform provides separate programming models for workloads that require direct thread-level control and workloads naturally expressed through tensor operations. Key features Shared CUDA runtime — Host-side crates provide d…

github.comhttps://github.com/NVIDIA/cuda-rust