Luisa Crawford
Aug 22, 2025 12:22
NVIDIA introduces the Blackwell Ultra GPU, advancing AI capabilities with increased performance, memory, and efficiency, setting a new standard for AI factories.
NVIDIA has announced the release of its latest GPU innovation, the Blackwell Ultra, which promises to revolutionize AI factories by enhancing performance, scalability, and efficiency. As part of the NVIDIA Blackwell architecture family, the Blackwell Ultra GPU integrates advanced silicon innovations, setting a new standard for AI training and reasoning, according to a blog post by NVIDIA.
Innovative Design and Architecture
The Blackwell Ultra GPU features a dual-reticle design, incorporating two reticle-sized dies connected via NVIDIA’s High-Bandwidth Interface (NV-HBI). This design enables a significant increase in performance while supporting the familiar CUDA programming model. With 208 billion transistors, the GPU outperforms its predecessor, the NVIDIA Hopper, offering a leap in processing capabilities.
Enhanced Tensor Cores and NVFP4 Precision
NVIDIA’s fifth-generation Tensor Cores, integrated into the Blackwell Ultra, provide substantial improvements in AI compute power. These cores support the new NVFP4 precision format, offering nearly FP8-equivalent accuracy with a reduced memory footprint. The GPU’s dense NVFP4 compute capability reaches 15 petaFLOPS, a 1.5x increase over the original Blackwell GPU, making it ideal for large-scale AI inference and model deployment.
Memory and Bandwidth Improvements
The Blackwell Ultra is equipped with 288 GB of high-bandwidth memory (HBM3E), significantly increasing memory capacity compared to previous models. This enhancement is crucial for handling multi-trillion-parameter models and facilitating high-concurrency inference in AI factories. The GPU also boasts a bandwidth of 8 TB/s, enhancing data processing speeds.
Advanced Interconnect and Integration
Incorporating NVIDIA’s fifth-generation NVLink technology, the Blackwell Ultra supports seamless GPU-to-GPU communication, with a bidirectional bandwidth of 1.8 TB/s. This feature, along with NVLink-C2C for coherent interconnects to NVIDIA Grace CPUs, enhances the scalability of AI systems, allowing for larger and more efficient AI deployments.
Enterprise-Grade Features
Designed with enterprise-grade features, the Blackwell Ultra simplifies operations while strengthening security. It includes advanced scheduling and management capabilities, such as the Enhanced GigaThread Engine and Multi-Instance GPU (MIG) support, allowing for secure multi-tenancy and predictable performance isolation.
Impact on AI Factories
The launch of the Blackwell Ultra GPU marks a significant advancement in AI technology, enabling AI factories to train and deploy models at unprecedented scales and efficiencies. With its groundbreaking architecture and enhanced features, NVIDIA’s latest GPU is set to redefine the capabilities of AI infrastructures, delivering more model instances and faster processing times.
For more information on the Blackwell Ultra GPU and its impact on AI technology, visit the NVIDIA blog.
Image source: Shutterstock
Source: https://blockchain.news/news/nvidia-unveils-blackwell-ultra-gpu-a-leap-in-ai-factory-technology
