The new kernels are designed to optimize training for GLA, Mamba-2 and other linear attention architectures. Researchers Albert Gu and Percy Liang highlighted the release, which leverages Pallas to improve processing efficiency on the v6e hardware to reduce computational overhead.
Tuesday, October 20, 2026
Story
FLAT TPU Kernels Beat H100 by 3x in GLA Training
AI7 days ago
Mayank Mishra has developed a new framework for executing linear attention on Google's TPU hardware to speed up model training. The release, called FLAT, provides Pallas training kernels that enable a single TPU v6e to perform forward and backward passes up to 3x faster than Flash Linear Attention on an Nvidia H100 for GLA, and 1.5x faster for Mamba-2.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.