Reproduce “Exploring Fast Fourier Transforms on the Tenstorrent Wormhole” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nick Brown, Jake Davies, Felix LeClair
The paper evaluates the performance and energy efficiency of the Fast Fourier Transform (FFT) algorithm on the Tenstorrent Wormhole n300 RISC-V accelerator, comparing it to a server-grade Intel Xeon Platinum CPU. While the accelerator was slower for a 1024x1024 2D FFT, it achieved approximately 3.6 times greater energy efficiency due to significantly lower power consumption.
Key insights
- Data reordering within the Tensix core architecture represents a dominant performance bottleneck for FFT, accounting for roughly 94% of initial execution time.
- RISC-V accelerators like the Tenstorrent Wormhole n300 offer substantial energy efficiency benefits for HPC workloads, consuming significantly less power than conventional server CPUs.
- Careful algorithmic adaptation and low-level architectural optimizations, such as leveraging specific scalar units for data movement, are crucial for achieving performance on specialized AI/ML accelerators for HPC tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.