--- format: "aidr-story-markdown/v1" id: "5c7e53bb2228f483c17359fe2ad5aaf46b477df7ec00ea92d89d3c6ae85d4af8" canonical_url: "https://aidr.today/5c7e53bb?lang=en" title: "FLAT TPU Kernels Beat H100 by 3x in GLA Training" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-10-14T01:54:32.000Z" category: "Frameworks" topics: ["google","llm","framework"] source_urls: ["https://marketbrief.now/ai/flat-tpu-kernels-beat-h100-by-3x-in-gla-training-8b7fb7ba","https://huggingnews.com/ai/flat-tpu-kernels-beat-h100-by-3x-in-gla-training-8b7fb7ba"] summary: "Mayank Mishra has released FLAT, a framework with Pallas training kernels that allow 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." --- # FLAT TPU Kernels Beat H100 by 3x in GLA Training > [Open the canonical story]() **Published:** 2026-10-14T01:54:32.000Z **Category:** Frameworks **Topics:** google, llm, framework ## Summary Mayank Mishra has released FLAT, a framework with Pallas training kernels that allow 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\. ## Sources - [Story source]() - [Story source]()