NVIDIA has announced the Blackwell Ultra GPU architecture, its most powerful data center GPU to date, featuring 20,000 CUDA cores and 192GB of HBM4 memory.
Key specifications
The Blackwell Ultra delivers 20 petaflops of FP8 inference performance, making it 4x faster than the H100 for large language model inference workloads.
Target market
NVIDIA is targeting hyperscalers and enterprise AI deployments. The GPU will be available in both PCIe and SXM form factors.
Pricing and availability
Cloud availability through AWS, Azure, and Google Cloud is expected in Q3 2025, with on-premise systems shipping to select partners in Q4.