NVIDIA's new software promises 30x speed boost for DeepSeek AI
What's the story
NVIDIA's CEO Jensen Huang has revealed plans to make DeepSeek's R1 artificial intelligence (AI) program 30 times faster.
He made the announcement during his keynote address at the company's GTC conference in San Jose, California.
Huang noted that using NVIDIA's Blackwell GPU chips, R1 can process 30 times the information compared to a data center (measured via number of tokens per second), thanks to a new open-source software called NVIDIA Dynamo.
Statement
What did the company say?
Ian Buck, the head of hyperscale and high-performance computing at NVIDIA, explained the prowess of NVIDIA Dynamo. He said, "Dynamo can deliver 30 times more performance in the same number of GPUs in the same architecture for reasoning models like DeepSeek."
Software innovation
Dynamo software: A game changer for AI processing
The Dynamo software, now available on GitHub, distributes the inference work across as many as 1,000 NVIDIA GPU chips.
This way, more work can be done per second of machine time by running tasks in parallel.
Buck explained that service providers can either choose to run more customer queries on DeepSeek or assign more processing power to a single user for a "premium" service.
Business strategy
NVIDIA's approach to investor concerns
NVIDIA's plan of utilizing more chips to increase throughput for AI inference, tackles investor fears that less computing would be utilized overall due to DeepSeek's efficiency in cutting down processing requirements.
Buck said that using Dynamo with Blackwell, AI data centers could generate 50 times as much revenue than the last model, Hopper.
This novel strategy emphasizes NVIDIA's dedication to improving AI processing capabilities while allaying market fears.
Product launch
New version of Blackwell chip out
Along with Dynamo, Huang also unveiled the latest iteration of Blackwell, "Ultra," which improves several features of the current Blackwell 200 model.
The new version boosts the DRAM memory from 192GB of HBM3e high-bandwidth memory, to as much as 288GB.
When combined with NVIDIA's Grace CPU chip, as many as 72 Blackwell Ultras can be configured in the company's NVL72 rack-based computer.
New product
NVIDIA reveals personal computer for AI developers
NVIDIA has officially branded its tiny personal computer for AI developers, previously known as Project Digits, as DGX Spark.
The device uses a version of the Grace-Blackwell combo, dubbed GB10.
NVIDIA is now accepting reservations for the Spark, marking another significant step in its commitment to advancing AI technology and providing innovative solutions for developers in the field.