FFT & Long Convolutions

July 28, 2026

In the previous post on SSMs, we saw that models like S4 unroll into a convolution. That leaves a practical question: how do you actually run long convolutions fast on modern GPUs?

This post is about the FFT, moving between coefficient and value form. The key idea is that a degree-(n) polynomial is uniquely determined by (n+1) points. Treat two coefficient sequences as polynomials of degrees (n) and (m); their product is then uniquely determined by (n+m+1) points, and that product is exactly a 1-D convolution of the coefficients.

See below for this really cool connection!

Convolutions + the FFT:

PDF

Open in new tab