Nvidia Announces Its Most Powerful GPU - Blackwell Ultra, Designed for Training Agent AI

At Computex 2025 in Taipei, Taiwan, Nvidia officially introduced Blackwell Ultra - its latest generation GPU, built for large-scale AI tasks, autonomous AI, and advanced inference.

Superior performance with Blackwell Ultra architecture

Blackwell Ultra is the "heart" of the GB300 NVL72 system - a liquid-cooled server cluster integrating 72 Blackwell Ultra GPUs and 36 Nvidia Grace CPUs . According to Nvidia, this system delivers 1.5 times higher performance than the previous generation (GB200 NVL72) and increases revenue opportunities by 50 times compared to systems using Hopper GPUs.

Together with the HGX B300 NVL16 (using NVLink Fusion technology), Nvidia claims that this new system has 11 times faster inference speed on large language models (LLM) , 7 times more powerful computing power and 4 times more memory than the Hopper generation.

Speaking about the newly launched product, Jensen Huang - Founder and CEO of Nvidia - shared:
"AI has taken a giant leap forward - autonomy and inference capabilities require orders of magnitude more computing power. Blackwell Ultra is designed for this moment - a flexible platform that can easily handle AI pretraining, post-training, and inference."

Autonomous AI and Physical AI

Nvidia says Blackwell Ultra is specifically designed to build agentic AI and physical AI . Among them:

1. Autonomous AI uses critical thinking and iterative planning to solve complex problems, taking a multi-step approach to decision making.
2. Physical AI allows companies to train robots and autonomous vehicles virtually at scale before deploying them in the real world.

The Blackwell Ultra system can integrate with the Spectrum-X Ethernet platform and Quantim-X800 InfiniBand , achieving 800Gb/s data transfer rates per GPU thanks to the ConnectX-8 SuperNIC - a network accelerator also launched at the event.

To optimize inter-GPU bandwidth, ConnectX-8 integrates 48 PCIe Gen6 lanes with a PCIe Gen6 switch, consolidating GPU-to-GPU and GPU-to-NIC connectivity into one high-performance device, replacing previous dedicated PCIe switches.

Blackwell Ultra products will be available from Nvidia's traditional partners such as Cisco, Dell, HPE, Lenovo and Supermicro from the second half of 2025 .

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

Micah Soto

Update 26 May 2025