New SOM algorithm SparseBin achieves 385x speedup over CUDA baseline

A new paper details a novel approach to optimizing GPU implementations for self-organizing map (SOM) algorithms. The research introduces SparseBin, a new SOM algorithm, and implements a rigorous tuning process for both SparseBin and its baseline comparison, cuSPARSE. This meticulous tuning, involving four key levers, resulted in significant speedups, with SparseBin achieving 5.6-10.1x faster performance per epoch across various map sizes and increasing the performance margin over a CUDA implementation from approximately 80x to 385x. The cuSPARSE baseline also saw a 2-3x speed improvement due to analogous tuning. AI

IMPACT This research demonstrates significant optimization techniques for GPU-based machine learning computations, potentially influencing future hardware-software co-design for AI workloads.

RANK_REASON The item is an academic paper detailing a novel algorithm and performance benchmarks. [lever_c_demoted from research: ic=1 ai=0.7]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• CUDA
• cuSPARSE
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Influence Flower
• L2
• Litmaps
• MEDLINE
• ScienceCast
• scite Smart Citations
• SparseBin

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →