Fourier Neural Operators Just Got a Turbo Boost: Researchers from UC Riverside Introduce TurboFNO, a Fully Fused FFT-GEMM-iFFT Kernel Achieving Up to 150% Speedup over PyTorch

Fourier Neural Operators Just Got a Turbo Boost: Researchers from UC Riverside Introduce TurboFNO, a Fully Fused FFT-GEMM-iFFT Kernel Achieving Up to 150% Speedup over PyTorch

Fourier Neural Operators (FNO) are powerful tools for learning partial differential equation solution operators, but lack architecture-aware optimizations, with their Fourier layer executing FFT, filtering, GEMM, zero padding, and iFFT as separate stages, resulting in multiple kernel launches and excessive global memory traffic. The FFT -> GEMM -> iFFT computational pattern has received inadequate attention regarding GPU kernel fusion and memory…