NVIDIA

Roadmap

NVIDIA HGX B300

Eight-GPU Blackwell Ultra platform with 2.1TB HBM3E for AI reasoning, training and inference

NVIDIA BLACKWELL ULTRA

Specifications (manufacturer data)

NVIDIA HGX B300 integrates eight Blackwell Ultra GPUs with 2.1TB of HBM3E memory for AI reasoning, large-scale model training, agentic systems, high-throughput inference and real-time video generation.

Figures marked TBC / est. are preliminary manufacturer specifications and remain subject to confirmation. Roadmap status and manufacturer specifications do not represent available capacity or a committed delivery date.

Discuss this GPU Browse GPUs available for enquiry

Architecture

Blackwell Ultra

VRAM per card

270GB HBM3E

Memory bandwidth

7.7 TB/s per GPU

FP16 / BF16 tensor

4.5 PFLOPS per GPU (sparse; dense 2.25 PFLOPS)

FP8 tensor

9 PFLOPS per GPU (sparse; dense 4.5 PFLOPS)

FP4 tensor

18 / 14 PFLOPS per GPU (sparse / dense)

TDP

Up to 1,100W per GPU (configurable)

Interconnect

5th-gen NVLink · 1.8 TB/s per GPU

Why it matters

Eight Blackwell Ultra GPUs with 2.1TB total HBM3E memory
144 / 108 PFLOPS total FP4 Tensor Core compute (sparse / dense)
62 TB/s total memory bandwidth and 14.4 TB/s NVLink Switch bandwidth

Built for

Frontier LLM training (100B+ parameters)
Ultra-long-context inference
Large MoE serving at scale

Software stack

CUDA 12.9+

PyTorch

Triton

vLLM

TensorRT-LLM

Compare accelerators

AMD MI350 Series — AMD Instinct MI355X

AMD MI350 Series

AMD Instinct MI355X

288GB HBM3E accelerator with fourth-generation AMD CDNA architecture for generative AI, training, inference and HPC

1400W maximum TBP

Configurations for enquiry

Reference NVIDIA configurations

Discuss the selected GPU, quantity, node design, deployment location and commercial terms with our team.

View enquiry configurations

Evaluate the right accelerator for your workload

Tell us your target model, workload profile and infrastructure requirements to compare accelerator options.

Discuss GPU requirements View Products