TCAPMINERS — Industrial ASIC Crypto Miners
NVIDIAEnterprise AI Infrastructure

NVIDIA HGX Blackwell GPU Servers — B200 & B300 Specifications & Availability

Production-ready NVIDIA HGX Blackwell platforms for frontier model training, large-scale inference, and HPC. Compare the HGX B200 and HGX B300 side by side and reserve from in-stock inventory.

NVIDIA HGX Blackwell GPU platform
In Stock

HGX B200

8× Blackwell SXM GPUs, 1.5 TB HBM3e, 5th-gen NVLink. The proven workhorse for trillion-parameter training and high-throughput inference.

640 units available
In Stock

HGX B300

8× Blackwell Ultra SXM GPUs, 2.3 TB HBM3e, 800 Gb/s ConnectX-8 networking. Built for agentic AI, reasoning models, and the largest training runs.

128 units available
Architecture
B200
NVIDIA Blackwell
B300
NVIDIA Blackwell Ultra
GPUs per system
B200
8× NVIDIA B200 SXM
B300
8× NVIDIA B300 SXM
GPU memory (per GPU)
B200
192 GB HBM3e
B300
288 GB HBM3e
Total GPU memory
B200
1.5 TB HBM3e
B300
2.3 TB HBM3e
Memory bandwidth (per GPU)
B200
8 TB/s
B300
8 TB/s
FP4 Tensor performance
B200
Up to 144 PFLOPS
B300
Up to 216 PFLOPS
FP8 / INT8 Tensor performance
B200
Up to 72 PFLOPS
B300
Up to 108 PFLOPS
FP16 / BF16 Tensor performance
B200
Up to 36 PFLOPS
B300
Up to 54 PFLOPS
NVLink generation
B200
5th-gen NVLink
B300
5th-gen NVLink
NVLink bandwidth (per GPU)
B200
1.8 TB/s
B300
1.8 TB/s
NVSwitch
B200
4× NVLink Switch
B300
4× NVLink Switch
Networking
B200
8× ConnectX-7 / BlueField-3 (400 Gb/s)
B300
8× ConnectX-8 (800 Gb/s)
Host CPU
B200
Dual Intel Xeon or AMD EPYC
B300
Dual Intel Xeon or AMD EPYC
System RAM
B200
Up to 4 TB DDR5
B300
Up to 4 TB DDR5
Storage
B200
Up to 8× U.2 NVMe (configurable)
B300
Up to 8× E1.S / U.2 NVMe
Form factor
B200
Air or liquid-cooled 8U / 10U
B300
Liquid-cooled 8U recommended
Typical power (system)
B200
~10.2 kW
B300
~14 kW
Cooling
B200
Air or DLC
B300
Direct liquid cooling
Workloads
B200
LLM training, inference, HPC
B300
Frontier LLM training, agentic AI, reasoning models
Availability
B200
640 units in stock
B300
128 units in stock

Deployment & integration

Systems ship as complete HGX baseboards integrated into OEM chassis (Supermicro, Dell, ASUS, QCT, Foxconn). Air-cooled and liquid-cooled SKUs are available. NVLink, NVSwitch, and Quantum-2 / Quantum-X800 InfiniBand fabrics are pre-validated for multi-node scale-out.

Software stack

NVIDIA AI Enterprise, CUDA 12+, TensorRT-LLM, NeMo, Dynamo, and Triton Inference Server are fully supported. Compatible with Kubernetes, Slurm, and the major hyperscaler orchestration layers.

Gallery

Inside the Blackwell platform

From the Blackwell superchip to full DGX SuperPOD deployments — a closer look at the hardware powering frontier AI.

NVIDIA HGX B200 baseboard with liquid cooling
HGX B200 — liquid-cooled baseboard
NVIDIA GB200 NVL72 rack
GB200 NVL72 rack
NVIDIA DGX SuperPOD deployment
DGX SuperPOD deployment
NVIDIA Blackwell superchip close-up
Blackwell superchip
NVIDIA DGX Station mainboard
DGX Station mainboard
NVIDIA HGX Blackwell platform
HGX Blackwell platform

Reserve HGX Blackwell capacity

Inventory moves quickly. Contact our enterprise team for configuration, lead times, logistics, and worldwide deployment support.

Contact enterprise sales