FLOPS / TFLOPS (Floating-Point Operations per Second)

How many floating-point calculations a chip can do per second; a TFLOPS is a trillion per second.

FLOPS stands for floating-point operations per second, a measure of how much compute power a GPU or main compute chip has; a TFLOPS is 10¹² operations per second, and a PFLOPS is 10¹⁵. It's easy to confuse with the lowercase-s FLOPs, which is the total number of operations a single forward pass requires — a measure of a model's computational cost; dividing FLOPs by FLOPS gives a rough estimate of inference time. Reading this number requires checking the numeric precision: the same chip's rating can differ several-fold between FP32, FP16, FP8, and FP4, and manufacturers often advertise the peak figure at the lowest precision or under sparse conditions; integer throughput is usually reported separately as TOPS. What a chip can actually sustain is further limited by memory bandwidth and how efficient the operators are. This number comes up when choosing a robot's onboard compute or estimating whether a VLA model can run in real time on the device.

ExampleNVIDIA advertises Jetson Thor's compute figure at FP4 precision, which can't be directly compared to the older Orin's number, which is quoted at FP16.

Also called

FLOPS, TFLOPS

Related

Tera Operations Per Second · Floating-Point Operations (FLOPs) · Numerical Precision Formats (FP32 / BF16 / FP16 / FP8 / INT8) · NVIDIA Jetson Thor · Onboard Compute Platform · Inference Latency

Sources

FLOPS - Wikipedia