FLOPS Calculator

How it works

peak FLOPS = devices × cores × clock (Hz) × FLOPs per cycle per core · sustained = peak × efficiency · time = workload ÷ sustained FLOPS

FLOPS (floating-point operations per second) is the standard measure of compute throughput. The theoretical peak multiplies the number of cores by the clock rate and by how many floating-point operations each core can retire per cycle: a fused multiply-add counts as two, and a 256-bit AVX2 unit handles four FP64 or eight FP32 lanes. Real applications reach only part of the peak — dense matrix multiply (LINPACK) often 60–90%, memory-bound code far less — which is what the efficiency input is for.

Worked example

One 8-core CPU at 3.5 GHz with AVX2 and two FMA units (16 FP64 FLOPs per cycle) peaks at 8 × 3.5 × 10^9 × 16 = 448 GFLOPS (56 GFLOPS per core). At 70% efficiency it sustains 313.6 GFLOPS, so a 1,000 TFLOP (10^15 FLOP) job takes about 53.1 min.

Sources: Wikipedia — FLOPS (floating-point operations per cycle per core) · NVIDIA — Ada GPU architecture whitepaper (RTX 4090 peak FP32)

Frequently asked questions