Benchmark tests on three RISC-V systems found that juFFTe was faster than FFTW at every tested transform size on the Sophon SG2044, where it averaged a 3x speedup in the multi-core results. On the SiFive X280, its edge lasted through a transform size of about 16,000, while the SpacemiT X100 produced a mixed picture that changed with cache and transform size. AMD Zen 5 was faster across the full range in the cross-architecture comparison.
This is an arXiv version 1 preprint dated 28 Aug 2026. It evaluates juFFTe against FFTW in single- and multi-core experiments on three RVV 1.0-capable CPU configurations: SiFive X280 cores on Tenstorrent Blackhole devices, SpacemiT K3 X100 cores, and Sophon SG2044 C920v2 cores. AMD Zen 5 served as the cross-architecture reference.
The implementation combines SPIRAL-generated vectorized DFT kernels with explicit RVV C-intrinsics in selected situations. The tests measured one-dimensional complex-to-complex discrete Fourier transforms, or DFTs, across transform sizes, with timings averaged over multiple runs. Both libraries used double precision and OpenMP, and FFTW_MEASURE was the default FFTW planning setting unless otherwise stated. The paper also converted runtime into an upper-bound estimate of floating-point operations per second using the transform size, its base-2 logarithm and the measured runtime.
Different chips, different results
On the X280, juFFTe consistently outperformed FFTW through a transform size of about 16,000. The paper places that threshold near the 128 KiB L2-cache boundary for a complex-double vector sized at 16 times the transform size in bytes. For larger transform sizes, the multi-core results were similar for the two libraries. Only four of the Blackhole device's 16 X280 cores were accessible because of software limitations.
The X100 results were more conditional. FFTW was faster at very small sizes and within the L2 cache, while juFFTe was more performant near the L1-cache limit. At large transform sizes, either library could be up to 80% faster, depending on the exact size.
In an eight-core X100 comparison, juFFTe was almost 50% faster than FFTW in most larger-size cases within L2. FFTW used FFTW_ESTIMATE, a heuristic planning setting, for that test. This differs from FFTW_MEASURE, which was the default setting unless otherwise stated, so the multi-core comparisons did not all use the same FFTW planning choice.
In the reported SG2044 multi-core tests, juFFTe was faster than FFTW for every benchmarked transform size. FFTW_ESTIMATE heuristic planning was used, and the average speedup was 3x.
The reported averages were 1.53x for X280 in single-core tests and 1.09x in multi-core tests; 0.90x and 1.91x for X100; and 0.73x and 3.03x for SG2044. These averages varied by processor and core count.
Zen 5 kept the lead
For each transform size in the Zen 5 comparison, the researchers selected whichever of juFFTe and FFTW was faster before making the cross-architecture comparison. Zen 5 was still faster across the full range. On average, the X100 and C920v2 reached about 15% of Zen 5 performance, while the X280 struggled at small and large sizes. The Zen 5 frequency was capped for the comparison, so the result does not represent its maximum reported frequency.
A narrow benchmark
The figures are benchmark measurements, not uncertainty estimates. The study reports averaged timings but no confidence intervals or run-to-run variability estimates. Its evidence is limited to one-dimensional complex-to-complex double-precision DFTs on the named hardware, core configurations and planner settings.
The hardware sample also came with an availability caveat. The SpacemiT K3 was not commercially available when the paper was written, and its data came from a cloud beta-test program. A100 results were omitted after early double-precision testing showed inferior performance relative to the X100.
Parts of the work received funding from EuroHPC project DARE SGA 1 under Grant Agreement No. 101202459. The document is an arXiv version 1 preprint dated 28 Aug 2026.
Paper data and sources
Original title: Performance Evaluation of Fast Fourier Transforms on Emerging RISC-V Hardware with Vector Extension Support
Authors: Daniel Seibel, Kaveh Haghighi Mood, Jayesh Badwaik et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text