How to read a GPU specification
Every specification needs a unit and a boundary. “80 GB” is not useful until you know whether it means one GPU or the whole system.
Per device is not aggregate
80 GB of local device memory
640 GB aggregate across eight devices
Read the exact variant
For example, NVIDIA lists H100 SXM with 80 GB of memory and 3.35 TB/s memory bandwidth. Another H100 variant can have different characteristics.
Memory bandwidth describes a different path from network bandwidth. A maximum device power figure is not the total server’s measured draw.
For compute throughput, check the numerical format and whether the number assumes sparsity.
A total is not one allocation
Eight 80 GB GPUs provide 640 GB in aggregate. An unchanged single-GPU program does not automatically get a 640 GB local allocation.
Keep the same discipline when estimating speed. Reading 100 GB at an assumed sustained 2 TB/s takes at least 0.05 seconds before other work.
Try it yourself
Open a manufacturer’s sheet and label five fields with their units and boundaries.
Check your answer
For bandwidth, identify both endpoints. For memory, distinguish per-device and aggregate. For power, distinguish device and system. Mark anything unclear rather than building a calculation on it.