Nvidia announces its most powerful GPU, the Blackwell Ultra, built for training agentic AI

At the Computex 2025 event in Taipei, Taiwan, Nvidia announced Blackwell Ultra, the company's latest GPU built for large-scale AI reasoning, agentic AI and advanced inference workloads.

The Blackwell Ultra powers the GB300 NVL72 system, a fully liquid-cooled computing rack that utilizes 72 Blackwell Ultra GPUs and 36 Nvidia Grace CPUs. Nvidia says this system delivers 1.5x more performance than its predecessor, the GB200 NVL72, and increases Blackwell's revenue opportunities by 50x compared to those built with Nvidia Hopper.

Along with the HGX B300 NVL16, which uses the NVLink Fusion, Nvidia says the new system features 11x faster inference on large language models, 7x more compute, and 4x larger memory when compared to the Hopper generation.

Jensen Huang, founder and CEO of Nvidia, said:

“AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance, We designed Blackwell Ultra for this moment — it’s a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference.”

Nvidia says it has built the new Blackwell Ultra platform specifically for building agentic AI and physical AI systems. For starters, Agentic AI uses reasoning and iterative planning to solve complex problems. It takes a multi-step approach to plan and take action. Physical AI, on the other hand, allows companies to virtually train robots and autonomous vehicles at scale before they're deployed in the real world.

The Blackwell Ultra systems can be integrated with Nvidia's Spectrum-X Ethernet and the Quantim-X800 InfiniBand platforms, with 800Gb/s of data throughput for each GPU in the system, connected via the new ConnectX-8 SuperNIC network accelerator, which was also announced at the event.

For better inter-GPU communication bandwidth, ConnectX-8 integrates 48 lanes of PCIe Gen6 with an integrated PCIe Gen6 switch, consolidating GPU-to-GPU and GPU-to-NIC communication into a single, high-performance device, switching from previously used dedicated PCIe switches.

Blackwell Ultra products will be available from Nvidia partners Cisco, Dell Technologies, Hewlett-Packard Enterprise, Lenovo, and Supermicro starting from the second half of 2025.

Blackwell-powered cloud instances will also be available on AWS, Microsoft Azure, Google Cloud, Oracle Cloud, as well as GPU cloud providers CoreWeave, Crusoe, Lambda, Nebius, Nscale, Yotta, and YTL. The GB300 NVL72 system will also be available on Nvidia's cloud platform, DGX Cloud.

Source

Hope you enjoyed this news post.

Thank you for appreciating my time and effort posting news every day for many years.

News posts... 2023: 5,800+ | 2024: 5,700+ | 2025 (till end of April): 1,811

RIP Matrix | Farewell my friend