Get Started
Home
Topics
Search
Library
Research questionHow should GPU implementations be compared fairly when their optimization effort is asymmetric?GPU speedups can reflect unequal engineering effort rather than an algorithmic advantage when a tuned kernel is compared with an untuned baseline. This issue is especially consequential for the best-matching-unit search that dominates self-organizing-map training.
Evaluation & Benchmarks
Machine Learning
Neural and Evolutionary Computing
Research Paper
Technology
Latest papersRecent research connected to this question, newest first.From 80x to 385x: A Best-Matching-Unit Search at the L2 Roof, Measured Against a Symmetrically Tuned BaselineThe evidence concerns SparseBin and cuSPARSE GPU implementations of self-organizing-map best-matching-unit search, evaluated across map sizes from 32×32 to 512×512 with tuning applied to both sides. Reported results include per-epoch speedups and an L2-bandwidth roof; broader hardware and workload generality are not established.research paper · Sep 4, 2026
Related questions
How can LLM-generated GPU kernels remain performant across hardware, inputs, and programming models?How can self-organizing maps scale to millions of neurons within GPU memory and bandwidth limits?How can shared GPUs schedule concurrent heterogeneous AI inference without combinatorial profiling as workloads change?How can symbolic regression optimize constants efficiently across heterogeneous expression-tree populations on GPUs?