Neural network simulation on shared-memory vector multiprocessors

Real-Time Scalable Cortical Computing at 46 Giga-Synaptic OPS/Watt with ~100× Speedup in Time-to-Solution and ~100,000× Reduction in Energy-to-Solution

“…Neural network simulations on supercomputers have a rich history, from early work simulating artificial neural networks [20], to more recent spiking neural networks projects including the NEST simulator [21], [22] on the K-computer [23], bio-physically detailed cortical microcircuits [24], [25], [26], GPU acceleration [27], and others summarized in [28]. Our simulations, using our optimized function-level kernel, have progressed from 16 racks of Blue Gene/L [4] (30 million neurons, 240 billion synapses) to 36 racks of LLNL Dawn Blue Gene/P [5] (1.6 billion neurons and 8.87 trillion synapses) to 96 racks of LLNL Sequoia Blue Gene/Q [6], [7], culminating in one hundred trillion synapses at "human-scale.…”