MaTEx Performance

MaTEx Deep Learning Performance

Experimental Testbed and Software Description:

1. Eight nodes, each of which has an Intel 20-core Haswell, InfiniBand FDR with OpenMPI 1.8.3, NVIDIA K40m GPUs using CUDA 7.5 and cuDNN 4.

2. Twenty nodes, each of which has an Intel 20-core IvyBridge, InfiniBand FDR with OpenMPI 1.8.4

Neural Networks

We test AlexNet, GoogLeNet, InceptionV3 and ResNet50 networks.

Results

Scaling of ImageNet Models relative to themselves on a single node, for testbeds 1 and 2 consisting, respectively, of GPUs and CPUs.