A whole nematode nervous system runs 1211x faster on Loihi 2

What the work claims

This is a systems-and-benchmarking paper from Virginia Tech's Institute for Advanced Computing. Its central claim is double. First, a tooling claim: NeuroML2Loihi (N2L) can take a model written in NeuroML, the XML standard the OpenWorm ecosystem uses for the worm, compile it to Lava primitives, and execute it on real Loihi 2 silicon, to the authors' knowledge the first time a full 302-neuron C. elegans model has run on a single neuromorphic chip¹. Second, a measurement claim: against the conventional reference workflow, NEURON running on a Jetson, the Loihi 2 execution achieves 514 to 1211 times lower latency and consumes between 0.0458 and 0.105 percent of the energy, depending on model scale¹.

The work is deliberately infrastructural. The authors are not claiming the worm is simulated well in any deep biological sense; they are claiming that standards-based cross-substrate simulation is now feasible, measurable, and honest enough to benchmark. That framing, unusual in neuromorphic papers, is exactly what makes it useful to read against living-tissue computing.

How it works

The pipeline parses NeuroML and LEMS model files, builds an intermediate representation, and maps it onto Loihi 2's Lava programming model, using leaky integrate-and-fire neurons and dense synaptic processes. The reference model is the c302 C. elegans network from OpenWorm: 302 neurons and 3119 synapses, with connection weights set to anatomical connection counts rather than measured synaptic strengths. Three scales are benchmarked: the full connectome; an Active subset of 18 neurons and 125 synapses that actually deviate from rest during simulation; and a Minimal model of just the two posterior touch receptor neurons, PLML and PLMR, which receive the external stimulus¹.

Fidelity is measured two ways. For spikes, Victor-Purpura and van Rossum distances against the NEURON reference are normalized by the same distances against a rate-matched Poisson baseline; a ratio below one means the Loihi 2 traces are closer to the reference than random spike trains are. The ratios are 0.286 at VP cost q=1, 0.456 at q=50, and 0.725 for van Rossum, with mean absolute spike-count deviation of exactly one spike per neuron. The authors summarize this as Loihi 2 reproducing spike trains roughly twice as close to reference as the random baseline at the strictest cost¹. For voltage, the picture weakens: Pearson correlation between affine-mapped Loihi 2 and NEURON membrane traces is negative for the minimal model (r = -0.144, a phase-offset artifact of oscillatory traces) and only 0.433 for the active and full models, with several-millivolt RMSE. The paper is direct about the cause: Lava's LIF and Dense primitives approximate, but do not equal, the IAF neurons and ExpTwoSynapse models of the NEURON reference, and matching behavior across substrates required hand-tuning of input scaling, recurrent scaling and decay terms for this specific model¹.

The benchmark numbers, verified against the paper's tables. On the full 302-neuron model, Loihi 2 completes the simulation in 0.0258 seconds versus 31.19 seconds for NEURON on the Jetson, a 1211-fold speedup, at 0.0131 joules versus 28.52 joules, 0.0458 percent of the reference energy. Loihi 2 latency grows less than 6 percent from the minimal to the full model, while the conventional cost scales with network size, so the advantage widens with scale. Instantaneous power, by contrast, differs only modestly: Loihi 2 draws about 0.505 W and the Jetson about 0.92 to 0.98 W, a 1.8 to 1.9-fold ratio. The energy win is therefore mostly a time win, a distinction the authors state explicitly¹.

Where a skeptic should push

The most load-bearing assumption is that spiking fidelity on two neurons certifies fidelity of a 302-neuron nervous system. Read the tables carefully: action potentials occur only in the PLM touch neurons across all models, so the spike-train metrics that look excellent are computed on exactly two spiking cells, with the other 300 neurons contributing subthreshold voltages at best, quiescence at worst. The honest summary is that this paper simulates a small active circuit embedded in a large silent connectome. That is not a flaw in the engineering; the full model was run and its energy measured. But claims about whole-animal fidelity should be scaled accordingly, and the Active-subset construction, 18 of 302 neurons, quietly concedes the point¹.

Second, the energy comparison is asymmetric in a way that matters. The Loihi 2 figure integrates total chip power including static draw; the Jetson figure is idle-subtracted dynamic power only, and both are measured on embedded-class hardware rather than the HPC clusters where serious c302 simulations usually run. The authors flag this openly and argue it makes their energy advantage a lower bound¹. Third, the mapping is not automatic: reproducing reference behavior required manual per-model hyperparameter tuning, and the authors note that deriving those parameters directly from the NeuroML description is future work. Cross-substrate fidelity was achieved by a human in the loop, which is precisely the expensive part at larger scales¹. Finally, connection weights are anatomical counts, not physiology; the model's link to the actual worm is structural, and the stimulation protocol probes only posterior touch.

Cross-substrate fidelity rules organoids must adopt

For organoid intelligence, the non-obvious implication is that fidelity is a budgeted quantity, not a property, and this paper shows what an honest budget looks like. Every serious biological-computing claim eventually confronts the same question this paper confronts: how do you know the computation you wanted is the computation the substrate performed? The answer demonstrated here is a three-part protocol: pick standardized distance metrics against both a reference simulator and a random baseline, report where they fail (subthreshold dynamics, phase drift), and publish the asymmetric energy accounting with its conservative direction stated. An organoid program needs the equivalent, and it is harder to build because there is no reference simulator. Tissue is the ground truth and the unknown at once.

The opportunity is methodological import. The Poisson-normalized fidelity ratio is a template for electrode-array readouts: report how much closer your decoded organoid state is to a trained model's prediction than to a matched shuffled ensemble, and treat subthreshold disagreement as a first-class result rather than noise to average away. The quiescent-connectome observation is equally transferable. If 284 of 302 neurons sit silent under a standard stimulus, then whole-substrate claims built on the 18 active neurons are inflated by construction, and the same inflation risk applies to every paper that benchmarks organoid computation on the small fraction of units its electrodes can see. Benchmark the silent majority honestly, or state its exclusion as a limitation.

The threat is speed asymmetry, and it is structural. Loihi 2 executes this biological network more than a thousand times faster than biological time, on a chip that costs less than a month of organoid culture, with deterministic repeatability and a published toolchain. A 1211-fold acceleration means a silicon copy of a biological circuit can run a year's worth of stimulation protocols in seven hours, with identical seeds. Living tissue cannot compete on emulation throughput, reproducibility, or parameter sweeps, and papers like this keep moving those numbers in silicon's favor. What tissue can still compete on is what this paper had to hand-tune into existence: the substrate participating in its own calibration, plasticity that makes a network better matched to its task rather than worse, and dynamics that no one scaled or decayed into place. The 0.433 voltage correlation is the quiet proof that silicon cross-substrate mapping buys speed at the price of never quite reproducing the biological detail; an organoid, for all its messiness, never has that particular problem. Its problem is that nobody can write down its transfer function at all, which is why the measurement discipline this paper demonstrates matters more than its speedup.

The bottom line

Established: a standards-based path from NeuroML to Loihi 2 silicon exists, runs a full 302-neuron C. elegans model on one chip, reproduces reference spike timing to well within random-baseline distance, and delivers 514 to 1211 times lower latency than a Jetson NEURON reference at a fraction of a percent of its energy. Established with caveats: fidelity rests on two spiking neurons, subthreshold dynamics do not transfer, the mapping required manual calibration, and the energy win is driven by latency rather than a large power ratio. For organoid intelligence the durable lessons are about measurement: normalize fidelity against chance, report failures, benchmark the inactive substrate, and publish conservative asymmetric energy accounting. What would strengthen the silicon side further: automatic parameter derivation from model descriptions, muscle and body loops closing the circuit to behavior, and application of the same honest protocol to connectomes with richer spontaneous activity. What would strengthen the biological side: organoid studies adopting this paper's reporting discipline as a prerequisite for cross-lab comparability.