Official announcements from companies

Prime Intellect Introduces OpenDiLoCo for 1.1 Billion Parameter AI Training with 90-95% Utilization

Jul 11, 2024, 05:56 PM

Prime Intellect has introduced OpenDiLoCo, an open-source implementation and scaling of DeepMind’s Distributed Low-Communication (DiLoCo) method. This new framework enables globally distributed AI model training and has been tested across three countries with 90-95% compute utilization. The OpenDiLoCo framework allows nodes to sync every 500 steps, significantly reducing the need for close proximity. The team successfully trained a 1.1 billion parameter model, three times the size of the original DeepMind work, using a hybrid code with torch FSDP and hivemind. The training was conducted with a bandwidth of less than 100mb/s. This development marks a significant step towards making decentralized AI training more accessible and efficient.