Cost-effective high performance distributed GPU cluster for Deep Learning tasks

Abstract

The expenses on computational resources for modern Deep Learning computing can be extremely large. However, most of them are spent on the chassis and not on the GPU units themselves. Since modern mass market graphic cards are usually chipper and have huge performance for video games, it was hypothesized, that a low cost cluster, made of several graphic cards, can reach the same performance for computational tasks as readymade enterprise GPU-server with significantly lower price. The concept of distributed GPU cluster based on mass market GPU units is presented in the article. During the experiments, performance of a cluster with two mass market GPU units was compared with performance of enterprise GPU-server with 8 GPU-units on the Deep Learning bench mark. The results shows benefits and limitations of use proposed distributed cluster. It describes cases, when this solution is up to 7 times more effective than enterprise one in terms of cost savings for chassis itself as well as for, additional equipment and maintenance.