Cohere open-sources megakernel serving engine, up to 1.58x faster than vLLM
cohere · x · 2026-09-09
Cohere has released the first fully-fledged LLM serving system built around a decode megakernel, fully open-source on GitHub.
- Approach: a single persistent CUDA kernel executes the entire decode forward pass, eliminating per-op launch overhead and full-grid barriers to maximize GPU utilization
- Results: on one H100 (BF16), 1.58x faster decode than vLLM at batch size 1 and 1.25x–1.41x end-to-end serving at BS=8, reaching 62% of H100 speed-of-light
- Scope: early research release with an OpenAI-compatible API, tested only on single-H100 (sm90a), CUDA 13+, batch sizes up to 8, built for North Mini Code
Related event: Cohere open-sources Megakernel inference engine, 1.58x faster than vLLM(2 posts)→
More from Infra
- NVIDIA ships CUDA Python 1.0 with stable APIs, making Python first-class for CUDA — PyTorch · 2026-09-09
- Inference is turning GPU compute into a tradable commodity — ArtificialAnlys · 2026-09-09
- Solving Navier-Stokes cost 130B output tokens — up to $18M depending on model pricing — mitsuhiko · 2026-09-09
- Baseten cuts delta weight syncs for frontier models to under 40 seconds — baseten · 2026-09-09
- Dell says DRAM, NAND shortages persist and nearly all leading-node products are constrained — Beth_Kindig · 2026-09-09
- Alphabet's CapitalG backs AI chip startup Celero at $3 billion valuation — dinabass · 2026-09-09