Kubism: Disassembling and Reassembling K-Means Clustering for Mobile Heterogeneous(…)
summary
Kubism disassembles K‑means by first scanning which point–centroid distances are skippable, building a bitmap metadata to reorder work so threads only fetch necessary centroids, then applies warp balancing to pack useful work into every GPU warp and a runtime task distribution that assigns about 20% of distance work to CPU on an NVIDIA Jetson Orin AGX and retunes the CPU/GPU split each iteration. Results show up to 2.65× per‑iteration speedup, 1.23× end‑to‑end, 3.3× more active threads and CPU use rising to 77% from 10%.
AI-written summary · about
chapters (8)
- 0:00Kubism introduction
- 1:27K-means on embedded systems
- 5:43Kubism solution overview
- 8:23Warp balancing and skip ratio
- 10:30CPU-GPU task distribution
- 11:33Evaluation results
- 15:11Conclusion and future
- 15:53Q&A
related talks
32mMachine Learning Using Various GPU Technology With Kubeflow.
35mClustering: A Guide for the Perplexed
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
18mCompiling Strassen-like Matrix Multiplication Algorithms to Fast CUDA Kernels
33mAI workloads on Kubernetes - How to maximize GPU utilization and cut costs
32mKoba: Disaggregated Data Loading for Multimodal Training at Luma AI
01h 00mStreaming Euclidean k-median and k-means with o(log n) Space
15mScaling the Next Paradigm of Heterogeneous Intelligence