NVIDIA B200 SXM

Manufacturer specification reference

Architecture

BlackwellHGX H100, H200 and B200 reference architecture components (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Category

Individual acceleratorHGX H100, H200 and B200 reference architecture components (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Accelerator memory

180 GBPer B200 SXM accelerator in the reviewed enterprise reference configuration.HGX H100, H200 and B200 reference architecture components (external source)

Memory type

HBM3eHGX H100, H200 and B200 reference architecture components (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Memory bandwidth

8 TB/sUp to peak memory bandwidth per GPU.HGX H100, H200 and B200 reference architecture components (external source)

Power

1,000 WConfigurable up to 1 kW per GPU; excludes CPU, networking and server losses.HGX H100, H200 and B200 reference architecture components (external source)

Precision formats

FP4, FP8, FP6, INT8, FP16, BF16, TF32, FP32, FP64NVIDIA HGX Platform (external source)

Cooling / integration

Verify the selected NVIDIA-certified OEM server cooling configuration; module specifications alone do not determine facility cooling requirements.HGX H100, H200 and B200 reference architecture components (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Configuration boundaries

• Selected capacity is current B200 SXM, not a GB200 GPU or early architecture marketing capacity.
• No normalized per-GPU throughput ranking is supplied from HGX aggregate figures.

Deployment considerations

• Use the actual allocatable memory of the chosen module and MIG profile rather than early architecture capacity announcements.
• Do not divide baseboard throughput by GPU count and present the result as independently benchmarked GPU throughput.

Software ecosystem

CUDA. Confirm the exact model, runtime release, server integration and vendor support matrix before evaluating a deployment.