MaTEx Performance (Old)

Charles Siegel edited this page Mar 28, 2017 · 1 revision

MaTEx Deep Learning Performance

Experimental Testbed and Software Description:

Intel 20-core Haswell, InfiniBand FDR, and OpenMPI 1.8.1

Neural Networks

MLP consists of two layers of 800 features each.

CNN consists of two convolution layers of 16 and 32 features, with 5x5 windows, stride of 2, and after each a 2x2 max pooling layer, followed by a fully connected layer of 500 features.

MNIST Results

MNIST CNN Strong Scaling Performance

CIFAR10 Results