AI BriefingJane Street and NVIDIA Release Mamba-2 cuTile Kernels, Achieving Up to 1.82x Speedup on B300
·2026.10.26 20:00
Key point
cuTile kernels that boost Mamba-2 forward propagation speed by up to 1.82x on B300 GPUs have been released.
Details
- Jane Street and NVIDIA implemented Mamba-2 forward propagation using three cuTile Python kernels, achieving performance 1.14 to 1.82 times faster than Triton on NVIDIA B300
- The performance gap widens as the state size increases, with a 1.82x improvement at state 512 and a 1.39x improvement at state 64 for a 16K token baseline
- The
mamba2_chunk_forwardand backward kernelmamba2_chunk_backwardwere released under the MIT license in the TileGym repository - Optimizations for the Blackwell architecture include latency hints, A/B operand swapping, global memory layout changes, and the application of an autotuner
- This is a kernel-level forward propagation benchmark; backward propagation and overall training speed improvements have not been measured
- Available from PyPI
tilegym1.5.1 onwards, requiring a CUDA 13.1+ and Python 3.10~3.14 environment
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
gpu-unlock Linux module doubles FP/BF16 instruction rate on Nvidia 3000 series GPUs
Nvidia OpenShell: Secure Runtime for Autonomous AI Agent Fleets