Cursor Autonomously Optimizes NVIDIA CUDA Kernels for 38 Percent Speedup

Cursor

Apr 15, 2026 · Updated Apr 25, 2026

Cursor partnered with NVIDIA to apply a multi-agent system to CUDA kernel optimization, achieving a 38 percent geomean speedup on Blackwell GPUs. This demonstrates that autonomous agents can solve complex hardware engineering tasks that previously required months of manual effort from human experts.

Cursor, an AI-native code editor built by Anysphere, developed a multi-agent system that autonomously optimized 235 CUDA kernels (software that executes math on GPUs) for NVIDIA Blackwell 200 hardware. Over three weeks, the system achieved a 38% geomean speedup, with 19% of solutions delivering improvements greater than 2x.

GPU performance is often limited by the manual effort required to tune individual operations for new hardware. This research proves that multi-agent architectures can navigate the full solution space of low-level assembly. By automating this optimization, providers can significantly reduce inference (running a model) costs and energy consumption.

While this experiment focused on Blackwell kernels, these multi-agent coordination techniques will soon be integrated into the core Cursor product. You can currently use the editor for agentic coding, but expect future updates to handle more autonomous, long-running optimization workflows that fall outside standard training data distributions.

View the full update on cursor.com

Cursor

@cursor_aiApr 14

We've been developing a multi-agent system that builds and maintains complex software autonomously. Recently, we partnered with NVIDIA to apply it to optimizing CUDA kernels. In 3 weeks, it delivered a 38% geomean speedup across 235 problems. https://t.co/0YvbXrzVfe

791.2k

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from Cursor →

Keep reading

Cursor Releases Warp Decode for 1.84x Faster MoE Inference on Blackwell GPUs

Anysphere rebuilt the Mixture of Experts inference path for NVIDIA Blackwell GPUs, achieving 1.84x faster throughput by assigning GPU warps to individual output neurons. This warp decode approach eliminates the data-shuffling overhead typical of expert-centric models while improving output accuracy by 1.4x.

NVIDIAMay 5

NVIDIA Open Sources Internal Agentic Workflows for Supply Chain Optimization

NVIDIA released the open-source agentic workflow it uses internally to optimize its global supply chain in minutes rather than weeks. The system combines GPU-accelerated solvers with multi-agent orchestration to turn natural language logistical problems into mathematically optimized decisions.

Modular Releases Mojo Agent Skills for Writing Portable GPU Kernels

ModularMay 14

Modular Releases Mojo Agent Skills for Writing Portable GPU Kernels

Modular published open-source Mojo agent skills that plug into tools like Claude Code and Cursor to provide specialized knowledge for the Mojo language. These skills allow agents to translate legacy CUDA kernels into portable Mojo code that runs across NVIDIA, AMD, and Apple Silicon hardware.

Qwen Sets 580 TPS Record for Agentic Workloads on Blackwell GPUs

QwenMay 27

Qwen Sets 580 TPS Record for Agentic Workloads on Blackwell GPUs

Qwen achieved a record 580 tokens per second running its Qwen3.5-397B-A17B model on NVIDIA Blackwell GPUs using the TokenSpeed inference engine. The optimization targets agentic workloads, where multi-turn reasoning and tool-calling typically suffer from high latency. By combining a hybrid attention architecture with deep kernel fusion, the system maintains high throughput even as context scales to one million tokens.