nVidia H100 (14592CudaCores/456TensorCores, 80GB HBM2e, PCI-E16g5, 2sloty, 350W), pasivní chlazení, pouze pro výrobu

FP64                               26 teraFLOPS
FP64 Tensor Core           51 teraFLOPS
FP32                               51 teraFLOPS
TF32  Tensor Core         756 teraFLOPS*
BFLOAT16 Tensor Core 1,513 teraFLOPS*
FP16 Tensor Core          1,513 teraFLOPS*
FP8 Tensor Core            3,026 teraFLOPS*
INT8 Tensor Core           3,026 TOPS*
GPU memory                   80GB
GPU memory bandwidth  2TB/s
Decoders                         7 NVDEC 7 JPEG
Max thermal design power (TDP)        300-350W (configurable)
Multi-instance GPUs         Up to 7 MIGs @ 10GB each
Form factor                       PCIe  dual-slot  air-cooled
Interconnect NVLink:      NVLink:  600GB/s
                                       PCIe Gen5: 128GB/s
Server options                  Partner and NVIDIA-Certified Systems with1–8 GPUs
NVIDIA Enterprise            Included

* With sparsity.

Product website:

https://www.pny.com/nvidia-h100