FLOPS
Floating-point operations per second — a measure of raw compute power.
Think of it like
Horsepower for chips — higher number sounds better, but doesn't always translate to real-world speed.
Example
A modern GPU might do 10 trillion FLOPS (10 TFLOPS); a model taking 100 trillion operations would take ~10 seconds.
How it actually works
FLOPS is a useful yardstick but can be misleading. Peak FLOPS (theoretical max) is rarely reached. Actual throughput depends on memory bandwidth, kernel efficiency, and problem size. It's also usually measured as FP32; FP16 or INT8 can be faster or slower depending on the hardware.
For product teams
FLOPS is one number but not the only one that matters — latency and throughput matter more.
For engineers
Use FLOPS for napkin math, but profile your actual code to see real throughput.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome