Cursor and NVIDIA Ship a Multi-Agent System That Optimizes CUDA Kernels — 38% Faster

Cursor's multi-agent coding system, built in partnership with NVIDIA, achieved a 38% geomean speedup across 235 CUDA kernel optimization problems — a concrete benchmark for what agentic coding can do at the hardware level.

Cursor has been quietly developing a multi-agent system capable of autonomously optimizing CUDA kernels, and the results are striking: a 38% geometric mean speedup across 235 benchmark problems, achieved in partnership with NVIDIA, as announced by @cursor_ai.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.