Abstract

Simulating the activity of the brain requires enormous computational resources: a recent simulation of a human brain scale neural network (around 10^11 neurons) for just one second of simulated time took 50 days on a cluster of 27 machines. Considerable effort is being put into innovative applications of diverse, modern computational hardware to this problem: general purpose graphics processing units; the SpiNNaker supercomputer specialised for simulating neural networks that will use 1M ARM cores; field programmable gate arrays; etc. Unfortunately, in all these cases the use of this cutting edge hardware comes at a very high cost of technical expertise and the majority of neuroscientists are simply not equipped to make use of them. I will discuss some of the challenges in making these techniques accessible to a wider non-specialist user base. In particular, I will describe two important problems that need to be solved to make this possible: automatic runtime code generation targeted at different hardware platforms; and automatic expression rewriting to maximise accuracy with low precision numerics.