NVIDIA B300

Manufacturer specification reference

Architecture

Blackwell UltraNVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Category

Individual acceleratorNVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Accelerator memory

Not verified for this recordPer accelerator; generic value not normalized across HGX and DGX documentation.NVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Memory type

HBM3eNVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Memory bandwidth

Not verified for this recordConfiguration-dependent GPU memory bandwidth; not the eight-GPU aggregate or NVLink bandwidth.NVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Power

Not verified for this recordA universal B300 module TDP is not established by the reviewed sources. Do not use a complete DGX server power rating as GPU TDP.NVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Precision formats

FP4, FP8, FP6, INT8, FP16, BF16, TF32, FP32, FP64NVIDIA Blackwell Ultra datasheet (external source)

Cooling / integration

NVIDIA documents system-specific cooling. Verify the exact server and module configuration before comparison.NVIDIA Blackwell Ultra datasheet (external source) · Introduction to NVIDIA DGX B300 System (external source) · NVIDIA HGX Platform (external source) · CUDA GPU Compute Capability (external source)

Configuration boundaries

• The Blackwell Ultra datasheet's HGX column lists 270 GB per GPU, 7.7 TB/s and up to 1100 W; current DGX B300 documentation lists 288 GB per GPU. These are retained as variant/source differences.
• Null numeric fields mean no universal generic value was verified, not zero.
• A system-specific B300 entry can be added once the intended configuration is selected.

Deployment considerations

• Resolve the exact OEM/module configuration and usable memory before estimating model fit.
• Keep each variant's memory, bandwidth and TDP together; mixing fields from different systems produces an invalid comparison.

Software ecosystem

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