Tuning Both Sides of a GPU Comparison Made the Baseline 2-3x Faster and the Speedup Claim Honest

arXiv 2609.05138·low signalXBlueskyLinkedInCopy link

arXiv 2609.05138 reports a rare symmetric benchmarking exercise: the author tuned both a novel SOM best-matching-unit kernel (SparseBin) and the cuSPARSE baseline it was compared against, giving every tuning lever an analogue on the baseline's side. SparseBin reached 5.6-10.1x per epoch over its previously published configuration at map sizes from 32x32 to 512x512, lifting the margin over the prior CUDA implementation from ~80x to ~385x — while cuSPARSE itself became 2-3x faster in the process. The tuned kernel presses the L2 bandwidth roof at 77% of peak, bounding any further lever at ~1.3x, which the author calls a terminal result rather than a waypoint.