About

Overview

The NVIDIA HGX B200 is at the forefront of AI supercomputing, integrating eight cutting-edge Blackwell B200 GPUs into a single baseboard. With each GPU delivering 180 GB of ultra-fast HBM3e memory, the B200 aggregate memory reaches a staggering 1.44 TB. This setup empowers enterprises and data centers to run large-scale AI and HPC workloads efficiently.

 

Key Features:

  • Unmatched AI and HPC Performance: Built on NVIDIA's latest Blackwell architecture, the HGX B200 leverages fifth-generation NVLink to provide 1.8 TB/s GPU-to-GPU bandwidth, ensuring minimal latency and high throughput during intensive AI model processing.

  • Massive Unified Memory: The total memory bandwidth of 8 TB/s per GPU allows for fully loaded large language models and complex simulations, drastically reducing data processing times.

  • Enhanced Versatility: Supports a range of precision formats, from FP4 and FP8 to FP64 for diverse workloads, making it versatile across AI and scientific computations.

  • Advanced Networking: The integration with NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet allows for seamless scalability in large cluster environments, accommodating thousands of GPUs for supercomputing power.

Whether in training extensive LLMs or powering data science pipelines, the HGX B200 is engineered for efficiency and performance, positioning itself as the backbone for next-gen AI data centers.

Characteristics
Architecture NVIDIA Blackwell
GPU Quantity 8x NVIDIA B200 Tensor Core GPUs
Form Factor SXM (commonly SXM6)
GPU Memory Capacity 180 GB HBM3e per GPU
Total Memory Capacity Approximately 1.44 TB for the 8-GPU system
Memory Bandwidth Up to 8 TB/s per GPU
FP4 Tensor Core 144 PFLOPS
FP8/FP6 Tensor Core 72 PFLOPS
INT8 Tensor Core 72 PetaOPS
FP16/BF16 Tensor Core 36 PFLOPS
TF32 Tensor Core 18 PFLOPS
FP32 600 TFLOPS
FP64 Tensor Core 296 TFLOPS
NVLink 5th Generation
NVSwitch 4th Generation
NVLink Bandwidt 1.8 TB/s per GPU
Total Aggregate Bandwidth 14.4 TB/s across the baseboard
Cooling Options Supports both air and liquid cooling
Max Thermal Design Power (TDP) Typically around 1000 W per GPU
Total Platform Power Consumption Approximately 10.2 kW for the complete system
Transformer Engine 2nd Generation, supports FP4 and FP6
Decompression Engine Enhances data processing speeds and efficiency
Multi-Instance GPU (MIG) Up to 7 instances per GPU
Media Processing Equipped with NVDEC and NVJPEG engines for video and image processing
Networking Support Offers NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet platforms
Network Speed Up to 400 Gb/s per port
Carbon Emission Reduction Achieves a 24% reduction compared to HGX H100, with significant improvements in energy efficiency for inference tasks
Overview

The NVIDIA HGX B200 is at the forefront of AI supercomputing, integrating eight cutting-edge Blackwell B200 GPUs into a single baseboard. With each GPU delivering 180 GB of ultra-fast HBM3e memory, the B200 aggregate memory reaches a staggering 1.44 TB. This setup empowers enterprises and data centers to run large-scale AI and HPC workloads efficiently.

 

Key Features:

  • Unmatched AI and HPC Performance: Built on NVIDIA's latest Blackwell architecture, the HGX B200 leverages fifth-generation NVLink to provide 1.8 TB/s GPU-to-GPU bandwidth, ensuring minimal latency and high throughput during intensive AI model processing.

  • Massive Unified Memory: The total memory bandwidth of 8 TB/s per GPU allows for fully loaded large language models and complex simulations, drastically reducing data processing times.

  • Enhanced Versatility: Supports a range of precision formats, from FP4 and FP8 to FP64 for diverse workloads, making it versatile across AI and scientific computations.

  • Advanced Networking: The integration with NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet allows for seamless scalability in large cluster environments, accommodating thousands of GPUs for supercomputing power.

Whether in training extensive LLMs or powering data science pipelines, the HGX B200 is engineered for efficiency and performance, positioning itself as the backbone for next-gen AI data centers.

Characteristics
Architecture NVIDIA Blackwell
GPU Quantity 8x NVIDIA B200 Tensor Core GPUs
Form Factor SXM (commonly SXM6)
GPU Memory Capacity 180 GB HBM3e per GPU
Total Memory Capacity Approximately 1.44 TB for the 8-GPU system
Memory Bandwidth Up to 8 TB/s per GPU
FP4 Tensor Core 144 PFLOPS
FP8/FP6 Tensor Core 72 PFLOPS
INT8 Tensor Core 72 PetaOPS
FP16/BF16 Tensor Core 36 PFLOPS
TF32 Tensor Core 18 PFLOPS
FP32 600 TFLOPS
FP64 Tensor Core 296 TFLOPS
NVLink 5th Generation
NVSwitch 4th Generation
NVLink Bandwidt 1.8 TB/s per GPU
Total Aggregate Bandwidth 14.4 TB/s across the baseboard
Cooling Options Supports both air and liquid cooling
Max Thermal Design Power (TDP) Typically around 1000 W per GPU
Total Platform Power Consumption Approximately 10.2 kW for the complete system
Transformer Engine 2nd Generation, supports FP4 and FP6
Decompression Engine Enhances data processing speeds and efficiency
Multi-Instance GPU (MIG) Up to 7 instances per GPU
Media Processing Equipped with NVDEC and NVJPEG engines for video and image processing
Networking Support Offers NVIDIA Quantum-2 InfiniBand and Spectrum-X Ethernet platforms
Network Speed Up to 400 Gb/s per port
Carbon Emission Reduction Achieves a 24% reduction compared to HGX H100, with significant improvements in energy efficiency for inference tasks
NVIDIA by NVIDIA

NVIDIA HGX B200 8x GPU Baseboard - 8x B200 SXM2 180 GB HBM3e

SKU: 935-26287-00A0-000 Mfr #: 935-26287-00A0-000

$497,323.00
$497,323.00
FREE Delivery on Orders $3,000+
* NO SALES TAX Except in CA
** Prices and stock might change without notice due to high market volatility.
Designed for the next era of generative AI and HPC, it features eight B200 SXM GPUs, each equipped with 180 GB HBM3e memory, del...  Read more
Designed for the next era of generative AI and HPC, it features eight B200 SXM GPUs, each equipped with 180 GB HBM3e memory, delivering up to 144 PFLOPS FP4 and 72 PFLOPS FP8 Tensor Core performance. This infrastructure is purpose-built for real-time inference and training of trillion-parameter models, offering 15x faster inference than previous generations.

Condition: NEW

Request Personalized Quote from our dedicated sales manager - Get Up To 5% Extra Discount!
Warranty & Support
Covered under a 3-YEAR EXTENDED WARRANTY from the manufacturer & SMTTR.
Shipping & Fulfillment
All products ship from verified U.S. distributors or SMTTR testing facility
Returns & Refunds
Returns accepted within 30 days Learn more

Shop with confidence! SMTTR is an authorized NVIDIA Partner Reseller

Let customers speak for us

WHY CHOOSE SMTTR

AI Solutions & Hardware Powered by NVIDIA Certified Professionals

TRUSTED BY

ENGINEERS FROM LEADING COMPANIES & OUR PARTNERS