'When will the mega kernel psyop stop': CUDA graphs are a strict super-set
realJosiahClark · x · 2026-11-10
A sharp rebuttal to the mega kernel hype: a CUDA graph containing co-designed kernels expresses a strict super-set of what a single mega kernel can do. Evidence: cuGraphAddBatchMemOpNode (analog of cuStreamWaitValue64) enables SM-free waiting, whereas a mega kernel must burn SM cycles spinning on synchronization.
Follow-up discussion covers whether a mega kernel could call out to the copy engine or write into DMA MMIO space — device-updatable kernels exist in theory, but it's not practical.
Related event: Megakernel Route Sparks Heated Debate in CUDA Community(6 posts)→