Can LLMs Write Fast Multi-GPU Kernels? — Simran Arora, Together AI

Puts a measured ceiling on model-written multi-GPU kernels and names where they break — collective ordering, data partitioning, choosing between copy engine, TMA and register transfers — while communication, not compute, becomes the real bottleneck.

• videoTaming the AI Hardware Cambrian Explosion — Abdul Dakkak, Modular
• videoSonarQube + OpenAI: Agentic Development — Killian Carlsen-Phelan, Sonar
• videoThe Autonomous Computer: Infrastructure for Computer Use Agents — Ang Li, Simular
• videoHow Secrets Leak Through AI Agents (and How to Stop It) — Venice AI

• videoTaming the AI Hardware Cambrian Explosion — Abdul Dakkak, ModularAI Engineer
• videoLarge clusters for small models — Daniel Svonava, SuperlinkedAI Engineer
• articleParallelKernelBench: Frontier LLMs can't write fast multi-GPU kernels (yet)Together AI

Checking sign-in…

Loading comments…