UC Berkeley, in collaboration with MIT, UC Irvine, and UT Austin, has released FlashLib, an open-source GPU library of classic ML operators. It includes 15 high-level operators designed to accelerate machine learning workflows and agent-based scenarios. Built on Triton and CuteDSL, FlashLib on an H200 GPU delivers peak speedups of up to 208x over NVIDIA's cuML 25.10 for operators like KMeans and KNN.
In H200 GPU benchmarks, Flash-KMeans achieves 61% of peak FLOPs efficiency, while Flash-KNN hits 85.2% of HBM memory bandwidth. Compared to NVIDIA's cuML 25.10, FlashLib speeds up KMeans clustering by 26x, KNN search by 19x, and HDBSCAN clustering by 40x. For TruncatedSVD decomposition, with specific tolerances and algorithmic trade-offs, it reaches a peak 208x acceleration. PCA, exact t-SNE, and MultinomialNB see speedups of 47x, 147x, and 49x, respectively.
As AI moves into the agent era, classic ML operators are shifting from offline batch processing to online real-time primitives. In long-chain reasoning tasks like scientific computing agents, models frequently call clustering, search, and dimensionality reduction operators, making traditional offline operators a system latency bottleneck. FlashLib introduces a performance prediction API that, without triggering GPU evaluation, estimates workflow runtime and memory cost in about 5 microseconds of CPU time, providing low-cost analysis support for task planning and agent decision-making. FlashLib is now open source on GitHub.