GPU vs TPU Architecture Workbench

Select a workload to see how an NVIDIA GPU SM and a Google TPU systolic array light up. Metrics are derived from public architecture guides and MLPerf submissions so you can compare utilization, memory pressure, and FLOPs/Watt trade-offs.

GPU SM activity Hopper H100 (128 SMs)

TPU systolic array v5e MXU (32K MACs)