Reproduce “Fast and Scalable Mixed Precision Euclidean Distance Calculations Using GPU Tensor Cores” 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
Brian Curless, Michael Gowanlock
FaSTED, a new algorithm from Northern Arizona University, efficiently computes Euclidean distances on GPUs by fully leveraging mixed-precision Tensor Cores, achieving up to 150 TFLOPS. It delivers 2.5x to 51x speedups over existing GPU methods on real-world datasets with negligible accuracy loss.
Key insights
- Meticulous, low-level GPU architectural optimizations, such as custom PTX instructions and shared memory XOR swizzling, are critical to fully unlock the extreme throughput of mixed-precision Tensor Cores and eliminate shared memory bank conflicts.
- Aggressively optimized brute-force algorithms on powerful hardware can significantly outperform state-of-the-art index-supported methods for high-dimensional similarity search, especially at higher selectivities.
- Mixed-precision FP16-32 arithmetic is sufficiently accurate for practical Euclidean distance calculations in similarity searches, validating its use for substantial performance gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.