GPU Acceleration in Python Using Elementwise Kernels

, NVIDIA
This presentation will demonstrate how Python users can parallelize operations on the GPU by using Elementwise Kernel functions from the CuPy API. Definitions of Elementwise Kernels follow a distinct template with four essential parameters and their goal is to provide elementwise operations that target the NVIDIA GPU without the need of CUDA code. The benefit of Elementwise Kernel functions will be shown through the comparison of serial Python, Elementwise Kernel functions, and CuPy’s Raw Kernels performing the same functionality. Raw Kernels are a user-defined kernel using raw CUDA source. Performance analysis will be done using NVIDIA’s Nsight Systems system-wide profiler. Overall, this talk will use code samples as the main mode of explaining implementation techniques along with step-by-step performance analysis.
活动: GTC Digital April
日期: April 2021
级别: 中级技术
话题: Libraries / Runtimes
行业: Public Sector
语言: 英语
所在地: