NVIDIA HGX AI Supercomputer

The world’s leading AI computing platform.

Purpose-Built for AI and HPC

AI, complex simulations, and massive datasets require multiple GPUs with extremely fast interconnections and a fully accelerated software stack. The NVIDIA HGX™ AI supercomputing platform brings together the full power of NVIDIA GPUs, NVLink®, NVIDIA networking, and fully optimized AI and high-performance computing (HPC) software stacks to provide the highest application performance and drive the fastest time to insights.

Unmatched End-to-End Accelerated Computing Platform

The NVIDIA HGX B200 and HGX B100 integrate NVIDIA Blackwell Tensor Core GPUs with high-speed interconnects to propel the data center into a new era of accelerating computing and generative AI. As a premier accelerated scale-up platform with up to 15X more inference performance than the previous generation, Blackwell-based HGX systems are designed for the most demanding generative AI, data analytics, and HPC workloads.

NVIDIA HGX includes advanced networking options—at speeds up to 400 gigabits per second (Gb/s)—using NVIDIA Quantum-2 InfiniBand and Spectrum™-X Ethernet for the highest AI performance. HGX also includes NVIDIA® BlueField®-3 data processing units (DPUs) to enable cloud networking, composable storage, zero-trust security, and GPU compute elasticity in hyperscale AI clouds.

HGX Stack

Deep Learning Inference: Performance and Versatility

Projected performance subject to change. Token-to-token latency (TTL) = 50 milliseconds (ms) real time, first token latency (FTL) = 5s, input sequence length = 32,768, output sequence length = 1,028, 8x eight-way NVIDIA HGX™ H100 GPUs air-cooled vs. 1x eight-way HGX B200 air-cooled, per GPU performance comparison​.

Real-Time Inference for the Next Generation of Large Language Models

HGX B200 achieves up to 15X higher inference performance over the previous NVIDIA Hopper™ generation for massive models such as GPT-MoE-1.8T. The second-generation Transformer Engine uses custom Blackwell Tensor Core technology combined with TensorRT™-LLM and Nemo™ Framework innovations to accelerate inference for large language models (LLMs) and Mixture-of-Experts (MoE) models. 

Deep Learning Training: Performance and Scalability

Projected performance subject to change. 32,768 GPU scale, 4,096x eight-way HGX H100 air-cooled cluster: 400G InfiniBand (IB) network, 4,096x 8-way HGX B200 air-cooled cluster: 400G IB network.

Next-Level Training Performance

The second-generation Transformer Engine, featuring 8-bit floating point (FP8) and new precisions, enables a remarkable 3X faster training for large language models like GPT-MoE-1.8T. This breakthrough is complemented by fifth-generation NVLink with 1.8TB/s of GPU-to-GPU interconnect, InfiniBand networking, and NVIDIA Magnum IO™ software. Together, these ensure efficient scalability for enterprises and extensive GPU computing clusters.

Accelerating HGX With NVIDIA Networking

The data center is the new unit of computing, and networking plays an integral role in scaling application performance across it. Paired with NVIDIA Quantum InfiniBand, HGX delivers world-class performance and efficiency, which ensures the full utilization of computing resources.

For AI cloud data centers that deploy Ethernet, HGX is best used with the NVIDIA Spectrum-X networking platform, which powers the highest AI performance over Ethernet. It features NVIDIA Spectrum™-X switches and BlueField-3 DPUs to deliver consistent, predictable outcomes for thousands of simultaneous AI jobs at every scale through optimal resource utilization and performance isolation. Spectrum-X also enables advanced cloud multi-tenancy and zero-trust security. As a reference design, NVIDIA has designed Israel-1, a hyperscale generative AI supercomputer built with Dell PowerEdge XE9680 servers based on the NVIDIA HGX eight-GPU platform, BlueField-3 DPUs, and Spectrum-X switches.

Connecting HGX With NVIDIA Networking

  NVIDIA Quantum-2 InfiniBand Platform:

Quantum-2 Switch, ConnectX-7 Adapter, BlueField-3 DPU

NVIDIA Spectrum-X Platform:

Spectrum-4 Switch,
BlueField-3 SuperNIC

NVIDIA Spectrum Ethernet Platform:

Spectrum Switch, ConnectX Adapter, BlueField DPU

Deep Learning Training Best Better Good
Scientific Simulation Best Better Good
Data Analytics Best Better Good
Deep Learning Inference Best Better Good

NVIDIA HGX Specifications

NVIDIA HGX is available in single baseboards with four or eight H200 or H100 GPUs or eight Blackwell GPUs. These powerful combinations of hardware and software lay the foundation for unprecedented AI supercomputing performance.

  HGX B200 HGX B100
GPUs HGX B200 8-GPU HGX B100 8-GPU
Form factor 8x NVIDIA B200 SXM 8x NVIDIA B100 SXM
FP4 Tensor Core 144 PFLOPS 112 PFLOPS
FP8/FP6 Tensor Core 72 PFLOPS 56 PFLOPS
INT8 Tensor Core 72 POPS 56 POPS
FP16/BF16 Tensor Core 36 PFLOPS 28 PFLOPS
TF32 Tensor Core 18 PFLOPS 14 PFLOPS
FP32 640 TFLOPS 480 TFLOPS
FP64 320 TFLOPS 240 TFLOPS
FP64 Tensor Core 320 TFLOPS 240 TFLOPS
Memory Up to 1.5TB Up to 1.5TB
NVIDIA NVLink Fifth generation Fifth generation
NVIDIA NVSwitch™ Fourth generation Fourth generation
NVSwitch GPU-to-GPU bandwidth 1.8TB/s 1.8TB/s
Total aggregate bandwidth 14.4TB/s 14.4TB/s
  * HGX H200
  4-GPU 8-GPU
GPUs HGX H200 4-GPU HGX H200 8-GPU
Form factor 4x NVIDIA H200 SXM 8x NVIDIA H200 SXM
FP8/FP6 Tensor Core 16 PFLOPS 32 PFLOPS
INT8 Tensor Core 16 POPS 32 POPS
FP16/BFLOAT16 Tensor Core 8 PFLOPS 16 PFLOPS
TF32 Tensor Core 4 PFLOPS 8 PFLOPS
FP32 270 TFLOPS 540 TFLOPS
FP64 140 TFLOPS 270 TFLOPS
FP64 Tensor Core 270 TFLOPS 540 TFLOPS
Memory Up to 564GB Up to 1.1TB
NVLink Fourth generation Fourth generation
NVSwitch N/A Third generation
NVSwitch GPU-to-GPU bandwidth N/A 900GB/s
Total aggregate bandwidth 3.6TB/s 7.2TB/s
  * HGX H100
  4-GPU 8-GPU
GPUs HGX H100 4-GPU HGX H100 8-GPU
Form factor 4x NVIDIA H100 SXM 8x NVIDIA H100 SXM
HPC and AI compute (FP64/TF32/FP16/FP8/INT8)* 268TF/4PF/8PF/16PF/16 POPS 535TF/8PF/16PF/32PF/32 POPS
FP8/FP6 Tensor Core 16 PFLOPS 32 PFLOPS
INT8 Tensor Core 16 POPS 32 POPS
FP16/BFLOAT16 Tensor Core 8 PFLOPS 16 PFLOPS
TF32 Tensor Core 4 PFLOPS 8 PFLOPS
FP32 270 TFLOPS 540 TFLOPS
FP64 140 TFLOPS 270 TFLOPS
FP64 Tensor Core 270 TFLOPS 540 TFLOPS
Memory Up to 320GB Up to 640GB
NVLink Fourth generation Fourth generation
NVSwitch N/A Third generation
NVLink Switch N/A N/A
NVSwitch GPU-to-GPU bandwidth N/A 900GB/s
Total aggregate bandwidth 3.6TB/s 7.2TB/s

Find out more about the NVIDIA H200 Tensor Core GPU.