NVIDIA HGX H100 8-GPU 

Specification NVIDIA HGX H100 8-GPU
Product Name NVIDIA HGX H100 8-GPU
Platform Type Accelerated Computing Platform
GPU 8 × NVIDIA H100 Tensor Core GPU
GPU Architecture NVIDIA Hopper
GPU Form Factor SXM5
GPU Memory per GPU 80GB HBM3
Total GPU Memory 640GB HBM3
Memory Bandwidth per GPU 3.35 TB/s
Total GPU Memory Bandwidth 26.8 TB/s
NVLink Generation 4th Generation NVLink
NVLink Bandwidth per GPU 900 GB/s bidirectional
NVSwitch 4 × 3rd Generation NVIDIA NVSwitch
GPU Interconnect Full NVLink / NVSwitch Fabric
Host Interface 8 × PCIe Gen5 x16
PCIe Bandwidth Up to 128 GB/s per x16 interface
FP8 Tensor Performance 32 PFLOPS*
INT8 Tensor Performance 32 POPS*
FP16 / BF16 Tensor Performance 16 PFLOPS*
TF32 Tensor Performance 8 PFLOPS*
FP32 Performance 540 TFLOPS*
FP64 Performance 270 TFLOPS*
FP64 Tensor Core 540 TFLOPS*
Typical Power ~5,600W
Baseboard Weight ~24 kg
Target Market AI / HPC / Data Center
Typical Server Configuration 4-GPU or 8-GPU HGX H100