StreamBrain: An HPC DSL for Brain-Like Neural Networks on Heterogeneous Systems
TimeThursday, 19 November 20208:30am - 5pm EDT
DescriptionWe introduce StreamBrain: a high-performance DSL for brain-like neural networks. StreamBrain supports multiple backends such as FPGAs, GPUs and CPUs on heterogeneous HPC systems while providing a convenient Keras-like interface to users. We show that training an MNIST dataset on the BCPNN model only takes 15 seconds. We empirically show that batching is critical for the BCPNN model as it allows the computational intensity to be controlled. Finally, we explored the resilience of the BCPNN model to reduced width of the numerical representation and showed that the mantissa of the double-precision (DP) computation could be reduced to a 9-bit representation, yielding nearly twice the performance of the original DP implementation.