Cursor Improves Expert Mixture Model Training Throughput by 41%

The AI coding tool Cursor has enhanced the training technology of its Expert Mixture-of-Experts (MoE) model, achieving a 41% increase in overall throughput in an environment with 512 NVIDIA GB300 GPUs. This technology reduces training bottlenecks by overlapping data transfer and computation between GPUs. On the 5th, Cursor released the open-source kernel for Expert Mixture Model training, named 'Mixture-of-Kittens (MoK)', claiming that it recorded performance up to 2.37 times faster than public alternatives in single Expert Mixture Layer tests. The Expert Mixture structure includes multiple expert networks and only calls a subset based on the input, allowing for an increase in model size while reducing computational load. However, when experts are distributed across multiple GPUs, data movement and computation wait times can increase. This technology focuses on reducing wait times by processing data transfer and computation simultaneously. No announcements with the same name were confirmed on Cursor's official blog or GitHub, and the figures of 41% and 2.37 times are based on reports from Dongchal Biting. Cursor previously stated that it improved the performance of Expert Mixture Layers for Blackwell GPUs by 3.5 times and overall training speed by 1.5 times. This technology could impact blockchain companies utilizing AI agents and code automation tools.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.