Research analysis · Neural simulation

The complete fruit-fly central nervous system now runs in real time on one mid-range FPGA

The synapse-resolution male Drosophila CNS connectome, 166.7k neurons and 25.6 million connections, previously needed a workstation cluster running 4.4 to 14 times slower than real time or 12 Loihi 2 chips. A single Gowin GW5AST-LV138 FPGA now executes it at 0.1 millisecond steps, finishing each 100 ms window in 86.83 ms on average and 93.05 ms at worst across 2000 tests, with bit-identical results to the uncompressed design. Two engineering ideas do the work: deferring per-synapse spike expansion until delivery, and packing neuron states into shared 24-bit codewords.

Source: FPGAFly: A 166.7k-Neuron Fruit-Fly Connectome on One FPGA via Deferred Spike Expansion and Shared-Codeword States, arXiv:2610.38118, 28 October 2026. Primary source. Read: full PDF, verified inline against every number quoted here.

What the work claims

Ye Ziyang demonstrates the first single-device real-time execution of an entire measured connectome, the MaleCNS v1.0 male fly central nervous system including the ventral nerve cord, on one FPGA costing a fraction of a neuromorphic multi-chip setup. Two constraints made this impossible before. Queue traffic: expanding each spike at emission into one record per outgoing synapse scales memory traffic with fan-out, which in fly connectomes is heavy-tailed and reaches 11,200 synapses for the most connected neuron; early expansion would need up to 195.2 MiB of in-flight queue data. State storage: two 18-bit neuron states plus a refractory counter need 41 bits per neuron, 6.83 Mbit total, exceeding the chip's 6.27 Mbit of block RAM; off-chip state is hopeless because updating every neuron each 0.1 ms step would demand about 17 GB/s against a 2.56 GB/s DDR3 controller ceiling1.

The solution is deferred spike expansion: when a neuron spikes, the system queues a single 64-bit descriptor in ascending source-ID order rather than one record per synapse. Because all synapses share one 18-step (1.8 ms) delay and the refractory period is 22 steps, each neuron can hold at most one descriptor in flight, bounding valid queue data to 1.27 MiB. The trick cuts external memory traffic by 54.5% to 57.9% and peak queue data by 95.6% to 96.5% versus early expansion, while delivering the exact same event sequence, because per-event saturation makes update order matter and the bitmap-ordered scan makes it deterministic. For state storage, the design exploits the fact that a refractory neuron holds both state variables at zero: refractory and out-of-range states are encoded in code space the normal state encoding cannot reach, compressing 41 bits to 24 and cutting block RAM from 438 to 270 units at a cost of 1245 extra LUTs, with zero changed MNIST predictions across all 10,000 test images1.

This is a primary engineering result with a working board implementation, not a proposal: the design met timing at 80 MHz after place-and-route with 270 block RAMs, 39.2k LUTs and 136 DSPs, and every one of 2000 measured windows matched the software model's spike counts, event counts and classification scores exactly.

How it works

The network is a frozen fixed-point leaky integrate-and-fire model derived from MaleCNS v1.02, with 18-bit signed weights stored in a compressed sparse row table and uniform 1.8 ms synaptic delay. Seven compute units each time-multiplex six neurons across local block RAM, while external DDR3 holds the 25.6M-entry connection table and the delay queue. Each 0.1 ms step has three phases: integrate all neurons and read out class scores, scan a spike bitmap in source order and enqueue descriptors, then deliver due events, expanding each descriptor into target-weight updates through a shared dispatcher1.

The order-preserving property is the subtle part. Updates saturate and quantize on each delivered event, so the sequence of updates, not just their set, determines the trajectory. Naive parallel-lane completion order would be nondeterministic; scanning the spike bitmap in ascending source-ID order makes delivery deterministic regardless of which lane finished first, which is what keeps deferred expansion bit-identical to early expansion at every step. The state encoding pulls the same move at the bit level: because firing resets both state variables to zero, the encoding can spend the unreachable normal-state codes on refractory counters and on out-of-range states quantized by 8, preserving in-domain precision exactly while gracefully degrading outside it1.

The benchmark is modest: 16 by 16 MNIST images pulse-encoded onto 3335 visual-input neurons over a 100 ms window, classified by a linear readout trained offline on 50,000 images over 554 selected neurons. Its purpose is throughput and fidelity measurement, not fly behaviour.

Where a skeptic should push

The headline fidelity caveat is stated in the paper's own table and easy to miss: exact spike match against the 41-bit reference is only about 42%. The compressed encoding preserves population statistics beautifully, 0.9904 to 0.9995 correlation in 5 ms bins and a firing-rate error of 0.0709 to 0.231 Hz, and it changes not a single MNIST prediction out of 10,000, but individual spike timings diverge from the uncompressed trajectory for most events. For a closed-loop robot or sensorimotor application, where downstream circuits are spike-triggered, that distinction may matter. Also note what the model is not: all synapses share one 1.8 ms delay, the neuron model is a fixed-point LIF with hard resets, and the connectome is frozen. This is a faithful event router for one specific dynamical model, not a physiological emulation of the fly nervous system.

Real-time is also conditional. All 2000 MNIST windows finished within budget, but sustained all-white input averaged 108.12 ms per 100 ms window, 5 of 28 synthetic stress conditions overran, and on a representative MNIST image 7.93% of steps exceeded the 8000-cycle budget, meaning the tail is being absorbed somewhere else in the schedule. The comparison table leans on literature numbers for Loihi 2 and Brian 2 rather than same-box measurements, and the whole result is a single-author, single-device implementation. None of this undercuts the engineering; all of it bounds the claim.

Connectome emulation sets the organoid bar

For biological computing this paper is best read as a price signal. Every serious argument for computing on living neural tissue leans, at some point, on a comparison with simulation: tissue does something that models do not, or does it cheaper, or does it with less machinery. FPGAFly compresses that escape route. A complete, synapse-explicit wiring diagram of 166.7k neurons now runs in real time on one commodity FPGA with deterministic, bit-auditable dynamics, for a hardware budget that does not appear in any organoid facility's top ten costs. If an organoid experiment claims an advantage over simulation, this is the baseline class it must now beat: cheap, reproducible, connectome-scale, and honest about its own fidelity limits in a way the organoid literature often is not.

The non-obvious lesson runs the other direction too. Both of FPGAFly's tricks are instances of a design principle that living tissue enforces for free: defer expansion of events until the structure that justifies them is known, and exploit natural invariants, here reset states, to compress bookkeeping. A cultured network has no 41-bit state budget and no DDR3 wall; its bottleneck is measurement and interpretation, not state storage. The right conclusion is not that organoids lose, but that the comparison is migrating from capacity, where silicon just won decisively, toward adaptation: the fly model is frozen, and nothing in this architecture learns. A connectome that could not rewire gave the FPGA its biggest simplifications, bounded in-flight work and unreachable reset codes. Take away the frozen structure and deferred expansion has to fight plasticity instead of exploiting it.

The near-term opportunity is tactical: a silicon twin. A real-time, whole-connectome emulator is the obvious control condition for claims that organoid dynamics compute rather than merely resemble computation. Run the same stimulus protocol, the same readout class, the same closed-loop schedule against the emulated connectome and the culture, and measure what the tissue adds. The threat is that the answer, in several published organoid computing claims, would be embarrassingly small. That is precisely why the experiment has value.

The bottom line

Established: a single low-cost FPGA can execute the complete male Drosophila CNS connectome in real time at biological 0.1 ms resolution, with bit-identical event delivery to the naive design and zero change in 10,000 classification predictions. Established with caveats: fidelity is population-level, exact spike match is roughly 42%, and real-time performance degrades under sustained high-rate input. For organoid intelligence, the durable consequence is competitive and methodological: connectome-scale emulation just became cheap, which raises the evidential bar for every claim that living tissue computes better than a model, and simultaneously offers the field its most rigorous new control condition.

Frequently asked questions

What is the MaleCNS connectome and how big is it?

MaleCNS v1.0 is a synapse-resolution reconstruction of the complete male fruit-fly central nervous system, including the brain and ventral nerve cord: 166.7k neurons connected by 25.6 million synapses.

What problem does deferred spike expansion solve?

Expanding a spike into one queue record per synapse scales memory traffic with fan-out, which reaches 11,200 for the best-connected neuron. Queuing a single 64-bit descriptor per spike instead cuts external traffic by 54.5% to 57.9% and peak queue data by 95.6% to 96.5%.

How can 24 bits replace a 41-bit neuron state?

A refractory neuron holds both state variables at zero, so those codes are unreachable during normal states. The design spends them on refractory counters and quantized out-of-range values, preserving exact in-domain arithmetic and cutting block RAM from 438 to 270 units.

Is the simulation faithful down to individual spikes?

Only partly. Population activity correlates with the reference at 0.9904 to 0.9995 and all 10,000 MNIST predictions are unchanged, but exact spike-level match is about 42% because out-of-range states are quantized.

Why does a frozen connectome matter for comparing with organoids?

Because the frozen structure is what makes the compression possible. Living tissue rewrites its own connectivity, which is both its claimed computational advantage and the reason the cheap-trick playbook demonstrated here does not transfer to it.

References

  1. Ye Z. FPGAFly: A 166.7k-Neuron Fruit-Fly Connectome on One FPGA via Deferred Spike Expansion and Shared-Codeword States. arXiv:2610.38118. 2026. https://arxiv.org/abs/2610.38118. Accessed 2026-11-02.
  2. Berg S et al. Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell 189(18):5504-5526.e15. 2026. (Connectome source, as cited in reference 1.)