← The index · Brain models & bodies · by eonfathom

FastFly

CUDA and CuPy simulator targeting real-time or faster execution of the FlyWire v783 graph on a consumer NVIDIA GPU.

FlyWire
Fidelity · reviewed
L1 Dynamics

Neurons are simulated — leaky integrate-and-fire or better — with documented, tuned parameters.

Dispute this rating →

The evidence

No evidence, no level.
Rating note

A CUDA/CuPy simulator of the FlyWire v783 graph (139,255 neurons, 54.5M synapses in CSR with FP16 weights). Neurons are leaky integrate-and-fire with a stated decay constant, threshold and reset (TAU_DECAY 0.9 per 1 ms step, roughly a 10 ms membrane time constant), so membrane state evolves under a documented model even though the units are normalised rather than millivolts: L1. Not L2: the weight array is passed into kernel_propagate_spikes as const half* __restrict__ and only ever read via atomicAdd into the target's input current. The only assignments to weights anywhere in the codebase are on the host, converting float32 to FP16 at load time or generating the synthetic fallback graph, both before the simulation starts.

Record

CategoryBrain models & bodies
DatasetsFlyWire
LevelL1 — reviewed
Listed incobanov/awesome-fly
First indexed2026-09-15

Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.