NVIDIA HGX H20: Pinnacle of Hopper Architecture
Unveiling NVIDIA’s HGX H20, a powerhouse driven by the cutting-edge Hopper architecture. With 96 GB of HBM3 memory and a 4.0 TB/s memory bandwidth, this GPU delivers unparalleled performance in AI applications. Diverse Tensor Cores, including INT8, FPB, BF16, and FP16, contribute to up to 296 TFLOPS, complemented by an additional 74 TFLOPS from the TF32 Tensor Core. Beyond raw power, the HGX H20 supports Multi-Instance GPU (MIG) technology, offering versatility with up to seven instances for optimized workload distribution. Enhanced by a 60 MB L2 Cache and a media engine featuring 7 NVDEC and 7 NVJPEG units, this GPU is designed for efficiency and robust multimedia processing. With a 400 W power consumption and a form factor tailored for eight-way HGX configurations, the HGX H20 marks a paradigm shift in GPU processing.
NVIDIA H20: The Next-Generation AI Inference GPU
The H20 is built on the Hopper architecture and features 14,592 CUDA cores. It integrates Tensor Cores optimized for AI workloads and supports the Transformer Engine, enabling highly efficient deep learning acceleration.
For memory, the H20 is equipped with 96GB of HBM3 memory with an ultra-high bandwidth of 4.0TB/s, significantly improving data transfer speeds. It supports NVLink for multi-GPU interconnect and uses the PCIe 5.0 interface.
The power consumption (TDP) of H20 is only 350W, making it much more energy-efficient than the 700W power draw of the H100 while maintaining strong AI compute capabilities. In FP16 precision, the H20 delivers up to 900 TFLOPS, and it also supports FP8 for optimized AI inference.
| Specification | Details |
|---|---|
| GPU Architecture | NVIDIA Hopper |
| GPU Memory | 96 GB HBM3 |
| GPU Memory Bandwidth | 4.0 TB/s |
| INT8 | FPB Tensor Core* |
| BF16 | FP16 Tensor Core® |
| TF32 Tensor Core* | 74 TFLOPS |
| FP32 | 44 TFLOPS |
| FP64 | 1 TFLOPS |
| RT Core | N/A |
| MIG | Up to 7 MIG |
| L2 Cache | 60 MB |
| Media Engine | 7 NVDEC, 7 NVJPEG |
| Power | 400 W |
| Form Factor | 8-way HGX |
| Interconnect | PCIe Gen5 x16: 128 GB/s, NVLink: 900GB/s |