Prime Highlight
- Nvidia has introduced its new Rubin computing architecture, which CEO Jensen Huang described as the company’s most advanced AI hardware platform to date.
- The launch reflects surging global demand for AI computing, as Nvidia prepares Rubin to replace the current Blackwell architecture later this year.
Key Facts
- Rubin delivers major performance gains, running 5 times faster than Blackwell for AI trainingand nearly five times faster for inference, with peak performance of 50 petaflops.
- The platform includes six integrated chips, featuring a new Rubin GPU, upgraded networking and NVLink interconnect, and a new “Vera” CPUdesigned for advanced AI reasoning tasks.
Background
Nvidia on Tuesday introduced its new Rubin computing architecture, calling it its most advanced AI hardware platform so far. CEO Jensen Huang unveiled the technology at the Consumer Electronics Show, saying the system is already in full production and will see a wider rollout in the second half of the year.
The Rubin platform will replace Nvidia’s Blackwell architecture, which had earlier taken over from Hopper and Lovelace. The company said the new design responds to the rapidly growing demand for computing power as artificial intelligence models become larger and more complex.
Rubin systems will support nearly all major cloud providers. Nvidia has lined up partnerships with firms such as Anthropic, OpenAI, and Amazon Web Services. The architecture will also run on Hewlett Packard Enterprise’s Blue Lion supercomputer and the upcoming Doudna system at Lawrence Berkeley National Laboratory.
Named after astronomer Vera Rubin, the platform includes six different chips that work together. At the centre is the Rubin GPU, supported by an upgraded BlueField networking system and a faster NVLink interconnect. Nvidia has also added a new “Vera” CPU that targets advanced AI reasoning tasks.
The company said the new design also tackles storage bottlenecks. Modern AI systems use large memory caches, which strain existing infrastructure. Nvidia has introduced a new external storage tier that connects directly to the compute hardware, allowing firms to scale storage more easily.
In performance tests, Nvidia said Rubin delivers major gains. The architecture runs about 3.5 times faster than Blackwell for training AI models and nearly five times faster for inference. Peak performance can reach 50 petaflops, while power efficiency improves by up to eight times for inference workloads.
The launch comes as global spending on AI infrastructure accelerates. On a recent earnings call, Huang said the industry may invest between $3 trillion and $4 trillion in AI hardware and facilities over the next five years.
Read Also: Toyota Raises Offer for Toyota Industries by 15% as Takeover Gains Momentum