NVIDIA HGX A100 datasheet

First-party NVIDIA platform datasheet for the 4/8/16-GPU HGX A100 configurations. The authoritative source for the A100 scale-up numbers and for the NVSwitch GENERATION NAMING: the prose heading reads 'Second-Generation NVIDIA NVSwitch Drives Full-Bandwidth Computing' while the link is described as 'Third-Generation NVIDIA NVLink ... doubles the GPU-to-GPU direct bandwidth to 600 gigabytes per second (GB/s)'. Per-configuration table carries NVIDIA NVLink '3rd generation', NVIDIA NVSwitch 'N/A' at 4 GPUs and '2nd generation' at 8 and 16, GPU-to-GPU Bandwidth 600 GB/s at all three sizes, and Total Aggregate Bandwidth 2.4 / 4.8 / 9.6 TB/s. Also the source for the claim that NVSwitch makes 'the entire node function as a single gigantic GPU'. All Tensor Core rows carry the '* With sparsity' footnote.