NVIDIA introduces the Blackwell Ultra GPU, part of the Blackwell architecture family, designed for AI factories and real-time AI services, promising enhanced performance, scalability, and efficiency for training and reasoning. The GPU features a dual-reticle design, HBM3E memory, fifth-generation Tensor Cores, and NVFP4 precision format for accelerated computing.
The Blackwell Ultra GPU is composed of two reticle-sized dies connected using NV-HBI interconnect technology, manufactured using TSMC 4NP with 208B transistors, providing peak performance per square millimeter. The GPU offers 640 fifth-generation Tensor Cores with NVFP4 compute, shared L2 cache, and 160 Streaming Multiprocessors for efficient AI workloads.
The fifth-generation Tensor Cores in Blackwell Ultra enhance performance and memory efficiency with 256 KB of Tensor Memory per SM, supporting dual-thread-block MMA for higher throughput and faster inference. The GPU also introduces NVFP4, a 4-bit floating-point format that balances accuracy, efficiency, and performance for low-precision AI inference, boosting performance by 1.5x compared to previous GPUs.
Blackwell Ultra accelerates softmax in the attention layer by doubling SFU throughput for faster compute in transformer models, improving reasoning with lower latency and compute costs. The enhanced attention mechanism enables faster AI reasoning, lower costs, and higher system efficiency, particularly beneficial for models with large context windows.
Memory capacity in Blackwell Ultra is scaled up with 288 GB of HBM3e per GPU, offering 3.6x more on-package memory than previous models. The GPU supports complete model residence, extended context lengths, and improved compute efficiency for diverse AI workloads, enabling the processing of multi-trillion-parameter models with high bandwidth memory features.
The Blackwell Ultra GPU is built for scale with fifth-generation NVIDIA NVLink for GPU-to-GPU communication, NVLink-C2C for communication with NVIDIA Grace CPU, and PCIe Gen6 interface for host connectivity. The GPU features advanced scheduling, security, and reliability features for enterprise-grade operations, including enhanced work scheduler, multi-instance GPU partitioning, and NVIDIA Remote Attestation Service engine.
Blackwell Ultra GPUs are designed to transform AI infrastructure, offering transformative performance from desktop superchips to AI factory racks. The superchip configuration with Grace CPU and Blackwell Ultra GPUs delivers up to 30 PFLOPS dense AI compute and 40 PFLOPS sparse compute, setting a new standard for AI processing. The GPU is fully compatible with CUDA ecosystem and optimized for modern AI frameworks and applications.
Read more at NVIDIA: Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era