Skip to content
Hacker News · AI·· 20 h agoAI score60

Together AI's road to TKCC: why it deleted its megakernels and let agents optimize inference

Last year we built megakernels. This year we deleted them: The road to TKCC

AI summary

Together AI describes TKCC, an agent-driven compiler that generates inference optimizations directly from forward-pass source code, reporting up to 1.7x higher TPS than its hand-tuned Llama 1B megakernel and 1,111 TPS on MiniMax M3.1-Flash and 1,205 TPS on DeepSeek v4.1-Flash. The team found that default Torch Inductor kernels plus CUDA Graphs and fine-grained PDL reached 975 TPS, 98% of the megakernel's 990 TPS, without writing a single kernel. Ablations showed that an expert-curated knowledge hub made agents 25-31% slower and that a naive prompt outperformed six expert-written optimization strategies.

Source:Hacker News · AI · together.ai