Nvidia launches Dynamo software, DeepSeek AI speed is 30 times faster

At the GTC conference on March 18, Nvidia CEO Jensen Huang announced the launch of Dynamo software, aiming to increase DeepSeek's artificial intelligence processing speed by 30 times. This move is a response to the market turmoil caused by DeepSeek's R1 artificial intelligence program. DeepSeek's technology is able to significantly reduce the amount of computation required for each query, raising concerns among investors about the future prospects of Nvidia.

Dynamo software can allocate AI inference tasks to up to 1,000 GPUs in parallel processing, greatly improving query throughput. Huang Renxun said that using the new Blackwell chip, Dynamo can provide 30 times the processing power under the same architecture, allowing AI data centers to process more tokens per second, thereby increasing overall revenue.

In addition, Nvidia released the adjusted DeepSeek R1 version on HuggingFace, using fewer computed bits “FP4” to significantly improve performance without significantly reducing model accuracy. Nvidia also launched the latest version of the Blackwell chip "Ultra", which improves memory capacity and overall performance, and released the DGX Spark, a small personal computer for AI developers, and updated the configuration of the DGX Station desktop computer.

At the conference, Nvidia also presented several other new products and collaboration projects, including improved network switches and software models for medical robots, demonstrating its continued innovation capabilities in AI and computing.