TOPS

TOPS counts the low-precision arithmetic operations a processor’s AI accelerator can perform each second, in trillions. A chip rated at 40 TOPS claims 40 trillion such operations per second under ideal conditions.

It is a marketing-friendly number and a crude one. Real performance depends on memory bandwidth, the numeric format used, how well the model is compiled for the chip, and thermal limits. Two chips with the same TOPS rating can differ by a factor of several on the same vision model.

For robots it matters because perception and language models have to run onboard: a warehouse robot cannot wait on a cloud round trip to decide whether the thing in front of it is a pallet or a person.