TOPS
Trillions of Operations Per Second — AI inference metric.
TOPS measures peak integer-operations throughput on a neural-network accelerator (NPU, GPU tensor unit, etc.). Apple's A19 Pro Neural Engine: ~38 TOPS. Qualcomm's Snapdragon 8 Elite 2 Hexagon NPU: ~50 TOPS. NVIDIA RTX 5090: 3,400+ TOPS in INT8. Caveat: the number depends on data type (INT8 / INT4 / FP16) — quoting "X TOPS" without the precision is half a benchmark, like advertising "MPH" without saying for what vehicle.