Fast, faithful and bit-reproducible GPU simulation of the Drosophila whole-brain leaky integrate-and-fire model

Abstract

Shiu et al. built a leaky integrate-and-fire (LIF) model of the whole Drosophila central brain from the FlyWire connectome. The model has 138,639 neurons and 15.1 million connections and is used for in silico silencing experiments. The reference implementation in Brian2 runs at only 0.2–0.5 simulated seconds per wall-clock second, and the many GPU re-implementations published since then were not compared on the same hardware. We present flysim, an open-source simulator for Apple (Metal) and NVIDIA (CUDA) graphics processing units (GPUs). flysim uses delay-window fusion, skipping of inactive 32-neuron words, ordering of neurons by cell type and batched trials. Synaptic input is accumulated as integer synapse counts, so the spike output is bit-identical on Apple and NVIDIA GPUs. In comparison with 40 Brian2 runs, flysim shows no significant difference in total or per-neuron firing, and reproduces the published rates of 15 named neurons (r = 0.9998). On an NVIDIA T4 it simulates 88–108 brain-seconds per second for one trial. Under sensory stimulation this is 3–24 times faster than the fastest other faithful implementation on the same GPU, but with activity spread randomly over half of the brain it is up to 30% slower. In a benchmark of twelve other implementations on the same machines, four of them accumulate synaptic input during the refractory period, unlike Brian2, and fire 18–26% more spikes. With common random numbers, a whole-brain silencing screen is exact with 341 times fewer simulated trials and takes less than four minutes on an Apple M5 GPU. Preprint. Supplementary Tables S1 and S2 are provided as a separate file. Code, benchmark adapters and raw measurements: doi:10.5281/zenodo.23298512 (flysim v1.0.0); https://gitlab.com/DustinZhu/flysim.