At the GTC conference on March 18th, Nvidia CEO Jensen Huang announced the release of Dynamo, a revolutionary software designed to accelerate the processing speed of DeepSeek's AI by 30 times. This development comes as a strategic response to the market disruption caused by DeepSeek's R1 AI program, which significantly reduces computational requirements per query, raising concerns among investors about Nvidia's future.

Dynamo achieves this performance leap by distributing AI inference tasks across up to 1,000 GPUs for parallel processing, dramatically increasing query throughput. Huang highlighted that when paired with Nvidia's new Blackwell chip, Dynamo delivers 30 times the processing power on the same architecture. This enhancement allows AI data centers to process more tokens per second, ultimately driving higher revenue.

One of Dynamo's standout features is its cost efficiency. The software can increase processing throughput at a rate of $1 per million tokens, enabling service providers to handle more client queries simultaneously or offer more powerful services to individual users at premium rates. Nvidia refers to this high-performance AI service as an "AI factory."

To further optimize DeepSeek's capabilities, Nvidia also released a tuned version of the DeepSeek R1 on HuggingFace. This version utilizes fewer computational bits—specifically "FP4"—resulting in significant performance improvements without compromising model accuracy.

In addition to Dynamo, Nvidia unveiled the latest iteration of its Blackwell chip, dubbed "Ultra," which boasts enhanced memory capacity and overall performance. The company also officially launched the DGX Spark, a compact personal computer tailored for AI developers, and updated the configuration of its DGX Station desktop computer.

The GTC conference served as a platform for Nvidia to showcase its continued innovation in AI and computing. Among the highlights were improved network switches and advanced software models for medical robots, further solidifying Nvidia's position as a leader in the industry.

Key Points

1. Nvidia introduced Dynamo software, offering a 30x speed boost for DeepSeek AI.
2. Dynamo enables parallel processing across 1,000 GPUs, enhancing efficiency and revenue potential.
3. A tuned version of DeepSeek R1 was released on HuggingFace, using FP4 bits for improved performance.
4. Nvidia unveiled the Blackwell Ultra chip and DGX Spark computer at the conference.
5. The company showcased advancements in network switches and medical robotics software models.