Nvidia Unveils Blackwell Ultra Chips to Power Next Generation of A.I. Reasoning
Jensen Huang, Nvidia CEO, announced the new Blackwell Ultra AI chip during the GTC conference on Tuesday, March 18, 2025. The chip will reach customers through hardware partners in the second half of 2025.
The Blackwell Ultra represents a significant advancement over the standard Blackwell architecture. It delivers 1.5 times more performance and contains 288GB of HBM3e memory, a 50% increase over its predecessor.
Nvidia achieves this boost by using 12-high memory stacks instead of the 8-high stacks found in regular Blackwell chips. This new chip generates up to 15 petaFLOPS of dense 4-bit floating-point performance.
The additional computing power enables reasoning models to operate at ten times the throughput of Nvidia’s previous Hopper generation. Huang emphasized the shift in AI development during his presentation.
“AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance,” he stated. The CEO specifically designed Blackwell Ultra for this new AI era.
Agentic AI moves beyond basic instruction following. These advanced systems can reason, plan, and take independent actions to accomplish specific goals. The Blackwell Ultra reduces response times for complex questions from over a minute to approximately ten seconds.
Nvidia Unveils Blackwell Ultra for AI Dominance
Nvidia will offer two primary versions of the Blackwell Ultra. The GB300 NVL72 combines 72 Blackwell Ultra GPUs with 36 Grace CPUs in a rack-scale solution. The HGX B300 NVL16 provides a server-class alternative for various deployment scenarios.
Major cloud providers such as AWS, Google Cloud, Microsoft Azure, and Oracle will integrate Blackwell Ultra into their offerings. This move helps Nvidia maintain its dominant position against competitors like AMD and Intel.
The company continues its accelerated development cycle with Blackwell Ultra. Huang revealed plans for the Rubin architecture in 2026 and Rubin Ultra in 2027. This roadmap aims to cement Nvidia’s leadership in AI computing as competition intensifies.
The enhanced memory capacity allows running substantially larger AI models on single GPUs. Meta’s Llama 405B can fit on one chip with memory to spare, enabling more efficient AI reasoning operations.
This article was produced by The Rio Times’ automated newsroom system. How we use AI · Report an error · Editorial responsibility: Matthias Camenzind, Editor-in-Chief