The Rapid Python Deep Learning Infrastructure (RaPyDLI) project delivers productivity and performance to the Deep Learning community by combining high level Python, C/C++, and Java environments with carefully designed libraries supporting GPU accelerators and MIC coprocessors (Intel Xeon Phi). Deep Learning (DL) has made major impacts in areas like speech recognition, drug discovery and computer vision. This success relies on training large neural nets—currently, up to 10 billion connections trained on 10 million images—using either large scale commodity clusters or smaller HPC systems where accelerators perform with high efficiency. This approach is of prime importance as the hardware accelerators enable much more sophisticated neural networks by increasing the available computational power by more than an order of magnitude.
RaPyDLI is a collaboration between ICL, Indiana University, and Stanford University, with each institution contributing their long standing expertise in the field. Currently, ICL’s focus for the RaPyDLI project is on efficient GPU kernel execution and optimization of scheduling strategies to reduce inefficiencies in the current code base in terms of performance and idle time.
Find out more at http://salsaproj.indiana.edu/RaPyDLI/