POSTER: Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU
We introduce Stream-K, a work-centric parallelization of matrix multiplication (GEMM) and related computations in dense linear algebra. Whereas contemporary decompositions are primarily tile-based, our method operates by partitioning an even share of the aggregate inner loop iterations among physical processing elements. This provides a near-perfect utilization of computing resources, regardless of how efficiently the output tiling for any given problem quantizes across the underlying processing elements.
On GPU processors, our Stream-K parallelization of GEMM produces a peak speedup of up to 14x and 6.7x, and an average performance response that is both higher and more consistent across 32,824 GEMM problem geometries than state-of-the-art math libraries such as CUTLASS and cuBLAS. Furthermore, we achieve this performance from a single tile size configuration per floating-point precision, whereas today’s math libraries employ complex kernel-selection heuristics to select from a large ensemble of kernel variants.
Sun 26 FebDisplayed time zone: Eastern Time (US & Canada) change
18:00 - 20:00 | |||
18:00 2hPoster | POSTER: Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU Main Conference Muhammad Osama University of California, Davis, Duane Merrill NVIDIA Corporation, Cris Cecka NVIDIA Corporation, Michael Garland NVIDIA, John D. Owens University of California, Davis Pre-print | ||
18:00 2hPoster | POSTER: Unexpected Scaling in Path Copying Trees Main Conference Vitaly Aksenov Inria & ITMO University, Trevor Brown University of Toronto, Alexander Fedorov IST Austria, Ilya Kokorin ITMO University | ||
18:00 2hPoster | POSTER: Transactional Composition of Nonblocking Data Structures Main Conference Wentao Cai University of Rochester, Haosen Wen University of Rochester, Michael L. Scott University of Rochester | ||
18:00 2hPoster | POSTER: The ERA Theorem for Safe Memory Reclamation Main Conference | ||
18:00 2hPoster | POSTER: AArch64 Atomics: Might they be harming your performance? Main Conference | ||
18:00 2hPoster | POSTER: Fast Parallel Exact Inference on Bayesian Networks Main Conference Jiantong Jiang The University of Western Australia, Zeyi Wen The Hong Kong University of Science and Technology (Guangzhou), Atif Mansoor The University of Western Australia, Ajmal Mian The University of Western Australia | ||
18:00 2hPoster | POSTER: High-Throughput GPU Random Walk with Fine-tuned Concurrent Query Processing Main Conference Cheng Xu Shanghai Jiao Tong University, Chao Li Shanghai Jiao Tong University, Pengyu Wang Shanghai Jiao Tong University, Xiaofeng Hou Hong Kong University of Science and Technology, Jing Wang Shanghai Jiao Tong University, Shixuan Sun National University of Singapore, Minyi Guo Shanghai Jiao Tong University, Hanqing Wu Alibaba Inc, Dongbai Chen Alibaba Inc, Xiangwen Liu Alibaba Inc | ||
18:00 2hPoster | POSTER: Efficient All-reduce for Distributed DNN Training in Optical Interconnect Systems Main Conference Fei Dai University of Otago, Yawen Chen University of Otago, Zhiyi Huang University of Otago, Haibo Zhang University of Otago, Fangfang Zhang Qilu University of Technology | ||
18:00 2hPoster | POSTER: CuPBoP: A framework to make CUDA portable Main Conference Ruobing Han Georgia Institute of Technology, Jun Chen Georgia Institute of Technology, Bhanu Garg Georgia Institute of Technology, Jeffrey Young Georgia Institute of Technology, Jaewoong Sim Seoul National University, Hyesoon Kim Georgia Tech | ||
18:00 2hPoster | POSTER: Generating Fast FFT Kernels on CPUs via FFT-Specific Intrinsics Main Conference Zhihao Li SKLP, Institute of Computing Technology, CAS, Haipeng Jia SKLP, Institute of Computing Technology, CAS, Yunquan Zhang SKLP, Institute of Computing Technology, CAS, Yuyan Sun Huawei Technologies Co., Ltd, Yiwei Zhang SKLP, Institute of Computing Technology, CAS, Tun Chen SKLP, Institute of Computing Technology, CAS | ||
18:00 2hPoster | POSTER: Learning to Parallelize in a Shared-Memory Environment with Transformers Main Conference Pre-print | ||