Together Compute Deleted Its Megakernels: Whole-Model Kernels Overcomplicate Optimization
simran_s_arora · x · 2026-11-17
The Together Compute team reflects on helping introduce some of the first AI megakernels a year ago, noting that while whole-model megakernels are a helpful abstraction, they overcomplicate systems and are not always an ideal optimization target, so they deleted them this year.
The post dives into two contrasting paradigms of ML optimization, Torch Compile and Megakernels, and walks through how the team understood and pushed past these limitations.
Related event: Together AI Drops Its Own Megakernels; New Compiler Is 28% Faster on B200(7 posts)→
More from Infra
- Rumor: Jensen Huang reportedly mulling a brand-new semiconductor startup with billions in day-one funding — Redo · 2026-11-17
- Researcher Praises Together's Torch Compile vs Megakernel Tradeoff Study — simran_s_arora · 2026-11-17
- Kalshi to Launch Stock Trading and AI Compute Futures, Reportedly Raising at $65B Valuation — nima_owji · 2026-11-17
- SpaceX uniquely owns the ride to orbit, the global network, and the AI computers at once — ccerrato147 · 2026-11-17
- Nvidia earnings could be good for NVDA but bad for the rest of the AI trade — pdamodaran · 2026-11-17
- TensorFold Runs Local LLMs at 874 tok/s With Native Metal/CUDA Inference — AIFlow_ML · 2026-11-17