103μs 降到 18μs 背后，Cursor 为何剑指英伟达，重写 GPU？
Ranking
Overall
82
Content
95
Popularity
N/A
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Cursor open-sourced Mixture-of-Kittens (MoK), a specialized MoE GPU megakernel that integrates token routing, inter-GPU communication, and expert computation. On 512 GB300 GPUs, it reportedly increased per-GPU training throughput by about 41% over Cursor’s prior DeepEP-based production system.
- MoK uses Pull-based forward dispatch, improving NVLink utilization by up to 29% and reducing signaling latency from roughly 103 μs to 18 μs in Cursor’s benchmarks.
- A single megakernel partitions GPU SMs between communication and expert FFN computation, enabling fine-grained overlap without relying solely on competing CUDA streams.
- Minibatching and a fixed-size ring token buffer balance early computation, GEMM efficiency, dynamic expert loads, and memory reuse without repeated CPU intervention.
- Against the fastest public baselines tested on GB300 NVL72, MoK achieved up to 2.37× faster MXFP8 forward execution and 1.78× faster backward execution, though Cursor did not publish a complete component-level ablation.
Sources (1)
103μs 降到 18μs 背后，Cursor 为何剑指英伟达，重写 GPU？
Public signals
N/A
TL;DR - Cursor open-sourced Mixture-of-Kittens (MoK), a specialized MoE GPU megakernel that integrates token routing, inter-GPU communication, and expert computation. On 512 GB300 GPUs, it reportedly increased per-GPU training throughput by about 41% over Cursor’s prior DeepEP-based production system.
- MoK uses Pull-based forward dispatch, improving NVLink utilization by up to 29% and reducing signaling latency from roughly 103 μs to 18 μs in Cursor’s benchmarks.
- A single megakernel partitions GPU SMs between communication and expert FFN computation, enabling fine-grained overlap without relying solely on competing CUDA streams.
- Minibatching and a fixed-size ring token buffer balance early computation, GEMM efficiency, dynamic expert loads, and memory reuse without repeated CPU intervention.
- Against the fastest public baselines tested on GB300 NVL72, MoK achieved up to 2.37× faster MXFP8 forward execution and 1.78× faster backward execution, though Cursor did not publish a complete component-level ablation.