A dirty way to measure GFLOPS

• 06 August 2025
• graphics

In Writing an AVX2 DGEMM kernel there was a valuable metric when estimating GFLOPS.

double time_taken = ((double)(end - start)) / CLOCKS_PER_SEC;
double flops = 2.0 * m * n * k;
double gflops = flops / (time_taken * 1e9);

What is actually going on here?

The expression 2.0 * m * n * k counts the base operations in matmul, i.e. one multiplication and one addition per element pair.

For each element in the result matrix $C(i,j)$, you'd compute:

• Multiply $A(i,k)\times B(k,j)$ for each $k$ ($n$ multiplications)
• Add the products together ($n-1$ additions ≈ $n$ additions for large $n$)

This is roughly $2n$ ops per output element. With $m\times n$ output elements, the total is $2\times m\times n\times k$ ops.

Keep in mind this is an estimation that ignores floating-point nuances, e.g. FMA counts as 2 ops, there is also some non-negligible loop overhead.

It's what CPU/GPU manufacturers use when reporting theoretical peak FLOPS, not by definition the standard industry heuristic for reporting D/SGEMM performance which would be profiling the kernels. There are more details in Intel & NVIDIA's developer forums.

• ← Previous
  Writing an AVX2 DGEMM kernel
• Next →
  Linear algebra reference