NVIDIA B200

The NVIDIA B200 is a data center GPU built on the NVIDIA Blackwell architecture, designed for frontier-scale AI training and inference. It features 192 GB of HBM3e memory with very high memory bandwidth (approximately 8 TB/s).

• Blackwell architecture using a dual-die design connected by a high-bandwidth die-to-die link, presented to software as a single GPU.
• 192 GB HBM3e for extremely large models, long context windows, and large-batch inference without sharding across many devices.
• Second-generation Transformer Engine with support for new low-precision formats including FP4 and FP8.
• Designed for NVIDIA HGX B200 systems and GB200 platforms, with high-speed NVLink for large multi-GPU domains.

Typical use cases: training and serving frontier LLMs and other massive neural networks, high-throughput generative AI inference, and demanding HPC workloads.

Substantial power and cooling requirements; intended for purpose-built HGX/GB200 systems. Confirm platform compatibility, power delivery, and cooling (air or liquid) before ordering. Contact us for lead times.