NVIDIa HGX BLACKWELL

The HGX B300 is the flagship of NVIDIA’s Blackwell platform, combining cutting-edge GPU and CPU technologies to tackle the most demanding AI workloads at scale. It integrates eight NVIDIA Blackwell B200 GPUs tightly coupled with two Grace Blackwell CPUs via high-bandwidth NVLink-Chip2Chip interconnects, delivering a unified architecture with unprecedented compute and memory capacity.

• Optimized for trillion-parameter foundation models, multi-modal AI, and real-time inference.

• Fully supported by the NVIDIA AI Enterprise software suite including NeMo, TensorRT-LLM, and CUDA
   
• Designed for deployment in AI factories, DGX SuperPODs, and hyperscale environments

• Up to 1.5TB of unified memory (combining GPU HBM and CPU LPDDR5X)
   

The HGX B300 sets a new benchmark in AI infrastructure, enabling breakthroughs in large-scale model training, inference, and AI-driven scientific research with superior efficiency and scalability.

The HGX B200 is part of NVIDIA’s Blackwell platform, purpose-built to power the next wave of large-scale AI, including trillion-parameter models, generative AI, and high-performance computing. It integrates eight NVIDIA Blackwell B200 GPUs connected via 4th-generation NVLink® and NVSwitch, delivering exceptional compute density, efficiency, and scalability.

• Designed for multi-GPU training, inference, and RAG workloads
   
• Fully compatible with NVIDIA AI Enterprise, including NeMo, CUDA, and TensorRT-LLM
   
• Scalable into large AI infrastructures like DGX SuperPOD

•  Optimized for FP8/FP6 precision and advanced Transformer Engine performance 

•  Up to 1.4TB of total GPU memory 

The HGX B200 enables organizations to accelerate innovation, reduce time to insight, and support next-gen AI applications with unmatched performance per watt and per rack.