Multifold Acceleration of Neural Network Computations Using GPU

被引：0

作者：

Guzhva, Alexander ^{[1
]}

Dolenko, Sergey ^{[1
]}

Persiantsev, Igor ^{[1
]}

机构：

[1] Moscow MV Lomonosov State Univ, DV Skobeltsyn Inst Nucl Phys, Moscow 119991, Russia

来源：

ARTIFICIAL NEURAL NETWORKS - ICANN 2009, PT I | 2009年 / 5768卷

关键词：

GPGPU; neural networks; perceptron; NVIDIA CUDA; parallel computations;

D O I：

暂无

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

With emergence of graphics processing units (CPU) of the latest generation, it became possible to undertake neural network based computations using CPU on serially produced video display adapters. In this study, NVIDIA CUDA technology has been used to implement standard back-propagation algorithm for training multiple perceptrons simultaneously on CPU. For the problem considered, CPU-based implementation (on NVIDIA CTX 260 CPU) has lead to a 50x speed increase compared to a highly optimized CPU-based computer program, and more than 150x compared to a commercially available CPU-based software (Neuro Shell 2) (AMD Athlon 64 Dual core 6000+ processor).

引用

页码：373 / 380

页数：8