Rust Engine Benchmarks

All measurements via Criterion 0.8, single-threaded, pure CPU. Hardware: 11th Gen Intel Core i5-11600K @ 3.90 GHz (6C/12T), DDR4-2400, Ubuntu 24.04. Verified via lscpu on 2026-05-31.

Last updated: 2026-06-04.

How to Run

Bash

# Full benchmark suite
cargo bench --bench full_bench
cargo bench --bench analysis_bench

# Quick (fewer iterations)
cargo bench --bench full_bench -- --quick
cargo bench --bench analysis_bench -- --quick

---

Core Engine (full_bench)

Bitstream Operations

Benchmark	Median
pack_1m	811 µs
pack_fast_1m	477 µs
pack_dispatch_1m (SIMD)	34.4 µs
popcount_portable_1m	29.6 µs
popcount_simd_1m	6.13 µs
bernoulli_stream_1024	4.25 µs
bernoulli_packed_1024	3.97 µs
encoder_64k_steps	281 µs

Fixed-Point LIF Neuron

Benchmark	Median
lif_10k_steps	56.3 µs
lif_100k_steps	758 µs

Dense Layer / Graph / Attention

Benchmark	Median
dense_forward_64x32	993 µs
mixed_dense_q88_q1616_64x32	2.245 µs
block_floating_dense_q16_64x32	10.056 µs
mixed_dense_q88_q1616_trap_64x32	2.325 µs
block_floating_dense_q16_trap_64x32	8.669 µs
mixed_dense_q88_q1616_envelope_64x32	2.322 µs
block_floating_dense_q16_envelope_64x32	8.748 µs
dcls_q88_tent_kernel_16tap	40.184 ns/sample
ultrascale_plus_target_contract_64x32	130.836 µs/emit
ultrascale_plus_dense_fold_plan_64x32	6.661 ns/plan
attention_10x16_20x32	88.5 µs
gnn_20x8_forward	85.3 µs

The 2026-06-04 dense precision rows were captured on a workstation under concurrent load and without exclusive CPU core isolation. Use the medians as local regression context and parity evidence. Before citing them as production throughput, rerun the benchmark on isolated cores and record CPU affinity, host-load, governor, and frequency evidence in the raw JSON artefact.

mixed_dense_q88_q1616_64x32 was measured on 2026-06-04 with taskset -c 10-11 cargo run --manifest-path engine/Cargo.toml --release --example bench_mixed_dense. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_mixed_dense.json; the matching Python reference artefact is benchmarks/results/local_python_2026-06-04_mixed_dense.json. Both artefacts record safe-workload overflow count 0 and saturating-probe overflow count 32 for parity with the mixed-dense HDL overflow_vector. They also record conservative envelope telemetry for comparison with the HDL abs_bounds_q1616 lanes: Python and Rust safe max absolute bound 531400, and saturating-probe max bound 17454214414336. The Python benchmark uses QFormatMixed(scale_per_tensor=False) here so the comparison is the same raw Q8.8/Q16.16 physical contract implemented by Rust and HDL.

block_floating_dense_q16_64x32 was measured on 2026-06-04 with taskset -c 10-11 cargo run --manifest-path engine/Cargo.toml --release --example bench_block_floating_dense. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_block_floating_dense.json; the matching Python reference artefact is benchmarks/results/local_python_2026-06-04_block_floating_dense.json. Both artefacts record safe-workload overflow count 0 and saturating-probe overflow count 32 for parity with the block-floating HDL overflow_vector. They also record conservative envelope telemetry for comparison with the HDL abs_bounds_q1616 lanes. The Python and Rust block-floating artefacts now use the same physical BFP workload: mantissa checksum -15, exponent checksum 0, safe max absolute bound 610816, and saturating-probe max bound 1125865547104256.

The precision trap rows were measured on 2026-06-04 with taskset -c 10-11 cargo run --manifest-path engine/Cargo.toml --release --example bench_precision_traps. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_precision_traps.json; the matching Python reference artefact is benchmarks/results/local_python_2026-06-04_precision_traps.json. Both Rust trap workloads reported 32 saturated outputs on the deterministic 64×32 overflow workload.

The precision envelope rows were measured on 2026-06-04 with taskset -c 10-11 cargo run --manifest-path engine/Cargo.toml --release --example bench_precision_envelopes. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_precision_envelopes.json; the matching Python reference artefact is benchmarks/results/local_python_2026-06-04_precision_envelopes.json. Both Rust envelope workloads reported conservative_overflow_free=true and matched the Python maximum absolute bounds for the deterministic 64×32 safe workload.

dcls_q88_tent_kernel_16tap was measured on 2026-06-04 with taskset -c 10-11 cargo run --manifest-path engine/Cargo.toml --release --example bench_dcls_q88. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_dcls_q88.json; the matching Python/PyTorch/SystemVerilog artefact is benchmarks/results/local_python_2026-06-04_dcls_q88.json. The Rust path reported overflow_count=0 and active_tap_total=2808700, matching the Python reference workload and DCLS Q8.8 saturation contract.

ultrascale_plus_target_contract_64x32 was measured on 2026-06-04 with the Rust benchmark process pinned to CPUs 10-11 inside a runtime cpuset shield. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_ultrascale_plus_target.json; the matching Python/Vivado-Tcl artefact is benchmarks/results/local_python_2026-06-04_ultrascale_plus_target.json. The Rust target report deliberately records fits_dsp_budget=false: a 64x32 one-DSP-per-MAC dense graph estimates 2048 DSPs against the ZU3EG budget of 360, so ZU3EG deployment needs a folded/time-multiplexed implementation or a larger target before any timing-closure claim is valid.

ultrascale_plus_dense_fold_plan_64x32 was measured on 2026-06-04 with the same runtime cpuset shield. The committed raw artefact is benchmarks/results/local_rust_2026-06-04_ultrascale_dense_folding.json; the matching Python/SystemVerilog artefact is benchmarks/results/local_python_2026-06-04_ultrascale_dense_folding.json. Both surfaces report the same ZU3EG fold plan: 320 DSPs per compute cycle, five output rows per cycle, one input fold, seven output groups, and seven compute cycles for the 64x32 dense contract.

PRNG

Benchmark	Median
prng_chacha_fill_1024	299 ns
prng_xoshiro_fill_1024	194 ns

---

Neuron Models (full_bench)

Legacy Neurons (neuron.rs)

Model	1k steps	10k steps	Per step
Lapicque	2.99 µs	19.5 µs	2.0 ns
ExpIF	25.0 µs	237 µs	24 ns
AdEx	29.1 µs	291 µs	29 ns

Interneurons (neurons/interneurons.rs) — Interneuron model group

Model	1k steps	Per step	Sub-steps	Notes
VIP	365 µs	365 ns	4	HH + A-type K+
Martinotti	530 µs	530 ns	4	Pospischil + M-current
SST+	552 µs	552 ns	4	Pospischil LTS + T-type + Ih
PV+ FS	4.25 ms	4.25 µs	50	Wang-Buzsáki + Kv3.1
Chandelier	4.29 ms	4.29 µs	50	WB + Kv1 + Kv3.1
Basket (cerebellar)	5.60 ms	5.60 µs	50	WB + A-type + KCa

PV+/Chandelier/Basket use 50 sub-steps (dt=0.01 ms, 0.5 ms per call) for Wang-Buzsáki gating stability. SST/VIP/Martinotti use 4 sub-steps (dt=0.025 ms, 0.1 ms per call) with Pospischil-style gating.

Sensory Neurons (neurons/sensory.rs) — Sensory model group

Model	10k steps	Per step	Type	Notes
Retinal ganglion	1.08 ms	108 ns	spiking	Pillow 2005 GLM (stim+history filters)
Inner hair cell	407 µs	40.7 ns	graded	Meddis vesicle pool + CaV1.3
Merkel cell	312.93 µs	31.29 ns	spiking	exact slow adaptation
Rod photoreceptor	663 µs	66.3 ns	graded	cGMP cascade + Ca²⁺-GC feedback
Nociceptor	259.48 µs	25.95 ns	spiking	exact membrane/sensitisation relaxation
Pacinian corpuscle	489.26 µs	48.93 ns	spiking	exact derivative-pressure adaptation
Olfactory receptor	411 µs	41.1 ns	spiking	cAMP + Ca²⁺/CaM + PDE4

Merkel cell remeasured 2026-05-31 after exact membrane/adaptation relaxation and fail-closed state guards. Pacinian corpuscle remeasured 2026-05-31 after exact derivative-pressure membrane/adaptation relaxation and fail-closed state guards. Nociceptor remeasured 2026-05-31 after exact membrane/sensitisation relaxation and fail-closed state guards. Remaining sensory rows retain their previously documented measurement date until each model is remeasured after its own hardening pass.

Motor Neurons (neurons/motor.rs) — Motor model group

Model	1k steps	Per step	Sub-steps	Notes
Alpha motor	6.48 ms	6.48 µs	50	WB + PIC (h_pic) + AHP + Ca²⁺
Gamma motor	161 µs (10k)	16.1 ns	1	LIF + adaptation
Upper motor	601.68 µs	601.68 ns	4	Pospischil RS + Ca²⁺, exact gates + conductance membrane
Renshaw cell	4.92 ms	4.92 µs	50	exact WB gates + adaptation

Renshaw cell remeasured 2026-05-31 after exact gate, adaptation, conductance integration, and fail-closed state guards. | Motor unit | 327 µs (10k) | 32.7 ns | 1 | exact LIF + force model |

Motor unit remeasured 2026-05-31 after exact membrane, adaptation, force-decay, subtype-parity, and fail-closed state guards.

Alpha motor is the most expensive per-step model due to WB gating (50 sub-steps), PIC evaluation, Ca2+ dynamics, and AHP computation at each sub-step.

Cerebellar Neurons (neurons/cerebellar.rs) — Cerebellar model group

Model	Steps	Median	Per step	Sub-steps	Notes
Granule cell (D'Angelo 2001)	10k	7.64 ms	764 ns	4	exact gates + 8 conductances: Na, K_dr, K_A, Ca_T, K_Ca, Ih, leak, tonic GABA
Golgi cell (Solinas 2007)	1k	2.96 ms	2.96 µs	10	exact gates + 11 currents: Na_t, Na_p, K_dr, K_A, K_M, Ca_T, Ca_N, BK, SK, Ih, leak

Granule cell remeasured 2026-05-31 after exact gate, calcium, conductance integration, full polyglot parity, and fail-closed state guards.

Golgi cell remeasured 2026-05-31 after exact gate, calcium, conductance integration, full polyglot parity, and fail-closed state guards. | Stellate cell | 1k | 6.08 ms | 6.08 µs | 50 | exact WB gates + exact Kv3.1 + conductance membrane | | Lugaro cell | 10k | 230.56 µs | 23.06 ns | 1 | exact LIF + adaptation + 5-HT | | Unipolar brush cell | 10k | 211.47 µs | 21.15 ns | 1 | LIF + persistent NMDA-like, exact relaxation | | DCN neuron | 1k | 2.71 ms | 2.71 µs | 20 | exact gates + 7 currents: Na_t, Na_p, K_dr, Ca_T, AHP, Ih, leak |

Lugaro cell remeasured 2026-05-31 after exact membrane/adaptation relaxation, Python/Rust/Go/Julia/Rust-safety parity, Mojo metadata alignment, and fail-closed state guards.

Stellate cell remeasured 2026-05-31 after exact Wang-Buzsáki h/n gates, exact Kv3.1 p-gate relaxation, exact conductance-form membrane integration, Python/Rust/Go/Julia/Rust-safety parity, and fail-closed state guards.

DCN neuron remeasured 2026-05-31 after exact gate, calcium, conductance integration, full polyglot parity, and fail-closed state guards.

Granule cell uses four exact-integration sub-steps for D'Angelo-style conductance dynamics and tonic GABA inhibition.

Ion Channel Variant Neurons (neurons/channels.rs) — Ion-channel variant group

Model	Steps	Median	Per step	Sub-steps	Notes
Persistent Na+	1k	4.61 ms	4.61 µs	50	WB + INaP subthreshold amplification
Ih (HCN)	1k	5.17 ms	5.17 µs	50	WB + Ih sag/rebound
T-type Ca²⁺	1k	9.17 ms	9.17 µs	50	WB + IT rebound bursting
A-type K+	1k	4.92 ms	4.92 µs	50	WB + IA onset delay
BK (Ca²⁺-K+)	1k	5.54 ms	5.54 µs	50	WB + BK fast AHP
SK (Ca²⁺-K+)	1k	4.35 ms	4.35 µs	50	WB + SK medium AHP
NMDA receptor	1k	4.81 ms	4.81 µs	50	WB + NMDA + Mg²⁺ block

Map Neurons (neurons/maps.rs) — Map-neuron model group

Model	Steps	Median	Per step	Notes
Aihara map	100k	3.38 ms	33.8 ns	Chaotic sigmoid map
Kilinc-Bhatt map	100k	2.45 ms	24.5 ns	Adaptive threshold map
Ermentrout-Kopell	100k	2.90 ms	29.0 ns	Canonical Type I (theta)

Population / Mean-Field (neurons/population.rs) — Population and mean-field model group

Model	Steps	Median	Per step	Notes
Montbrio-Pazo-Roxin	100k	1.57 ms	15.7 ns	Exact mean-field of QIF
Brunel balanced	100k	1.62 ms	16.2 ns	E/I balance, 2 rate ODEs
TUM (STP)	100k	3.03 ms	30.3 ns	Rate + depression + facilitation, 3 ODEs
El Boustani (NMDA)	100k	2.74 ms	27.4 ns	E/I + NMDA gating, 3 ODEs

Miscellaneous (neurons/misc.rs) — Miscellaneous model group

Model	Steps	Median	Per step	Notes
Graded synapse	100k	1.23 ms	12.3 ns	Non-spiking, passive RC + release sigmoid
Gap junction	100k	2.96 ms	29.6 ns	LIF + electrical synapse + Cx36 rectification
FH axon (GHK)	1k	5.84 ms	5.84 µs	Myelinated nerve, GHK driving force, 50 sub-steps
Node of Ranvier (MRG)	1k	1.46 ms	1.46 µs	Nav1.6 + INaP + Kv7, 20 sub-steps
Myelinated axon	1k	1.40 ms	1.40 µs	MRG node + internode cable
Cardiac Purkinje	1k	1.05 ms	1.05 µs	DiFrancesco-Noble, 6 currents, 10 sub-steps
Smooth muscle	1k	198 µs	198 ns	CaL + BK + IP3R/SERCA, 4 sub-steps
Beta cell	1k	196 µs	196 ns	CaL + K_dr + K_ATP + K_Ca, 4 sub-steps

Biophysical Models (neurons/biophysical.rs)

Model	Steps	Median	Per step	Sub-steps	Notes
Hodgkin-Huxley 1952	1k	13.3 ms	13.3 µs	100	Full 4-ODE HH with safe_rate kinetics
Wang-Buzsáki 1996	1k	6.94 ms	6.94 µs	50	FS interneuron, m_inf (no m state)
Connor-Stevens 1977	1k	71.13 ms	71.13 µs	100	Candidate-first RK4 A-type K+ model
Traub-Miles 1991	1k	1.80 ms	1.80 µs	10	CA3 pyramidal + M-current
Mainen-Sejnowski 1996	1k	1.86 ms	1.86 µs	20	Two-compartment (soma+axon)
Plant R15 1976	1k	1.11 ms	1.11 µs	5	Aplysia parabolic burster + Ca²⁺/KCa
De Schutter-Bower 1994	1k	775 µs	775 ns	5	Purkinje cell, Ca²⁺-dependent K
Golomb FS 2007	1k	711 µs	711 ns	10	Kv3 fast-spiking interneuron
Pospischil 2008	1k	686 µs	686 ns	4	Minimal HH, 5 cortical cell types
Destexhe 1993	1k	654 µs	654 ns	5	Thalamocortical + T-current rebound
Hill-Tononi 2005	1k	350 µs	350 ns	1	Na-dependent K, sleep/wake
Durstewitz 2000	1k	149 µs	149 ns	1	PFC + D1 dopamine + NMDA Mg²⁺ block
Bertram 2000	1k	132 µs	132 ns	1	Phantom burster, dual slow K
Av-Ron 1991	1k	120 µs	120 ns	1	Cardiac ganglion, Type III burst
Yamada 1989	1k	105 µs	105 ns	1	Subcritical Hopf burster
Huber-Braun 1998	1k	71.4 µs	71.4 ns	1	Temperature-sensitive cold receptor
Prescott 2008	10k	537 µs	53.7 ns	1	Type I/II/III excitability tuning
GLIF (Allen)	10k	363 µs	36.3 ns	1	LIF + threshold adapt + ASC
GIF population	10k	368 µs	36.8 ns	1	Escape-rate generalized IF
Mihalas-Niebur 2009	10k	123 µs	12.3 ns	1	Generalized IF, 20 spike patterns

Models with sub-steps have larger per-step cost due to inner ODE integration loops. HH is most expensive (100 sub-steps at dt=0.01 ms). Measured 2026-04-05 on i5-11600K @ 3.90 GHz.

Simple Spiking (neurons/simple_spiking.rs)

Model	Steps	Median	Per step	Notes
SuperSpike	10k	17.8 µs	1.8 ns	Surrogate gradient LIF
Brunel-Wang (LIF+NMDA)	10k	23.4 µs	2.34 ns	AMPA + NMDA Mg²⁺ block + GABA
e-prop ALIF	10k	28.3 µs	2.8 ns	Adaptive LIF for e-prop learning
Resonate-and-Fire	10k	41.8 µs	4.2 ns	2D subthreshold oscillator
Alpha synapse LIF	10k	53.1 µs	5.3 ns	Excitatory + inhibitory alpha
McKean	10k	320.24 µs	32.0 ns	RK4 piecewise-linear McKean
Hindmarsh-Rose	10k	77.4 µs	7.7 ns	3D burster (x,y,z)
COBA LIF	10k	104 µs	10.4 ns	Conductance-based with g_e, g_i
FitzHugh-Nagumo	10k	508.24 µs	50.824 ns	RK4 two-state qualitative spike model
FitzHugh-Rinzel	10k	572.65 µs	57.265 ns	RK4 three-state burster with slow y
Wilson HR	10k	506.27 µs	51.4 ns	RK4 polynomial cortical
Benda-Herz	10k	218 µs	21.8 ns	Stochastic rate + adaptation
Learnable neuron	10k	224 µs	22.4 ns	Learnable parameters (tau, beta)
Pernarowski	10k	632.87 µs	63.287 ns	RK4 three-state beta-cell burster
Terman-Wang	10k	1.2380 ms	123.80 ns	RK4 LEGION relaxation oscillator
Gutkin-Ermentrout	10k	289 µs	28.9 ns	Type I excitability
Sherman-Rinzel-Keizer	1k	29.0 µs	29.0 ns	Beta cell burster
Chay-Keizer	1k	30.5 µs	30.5 ns	Beta cell with KCa
Chay	1k	32.8 µs	32.8 ns	Beta cell 3-variable
Morris-Lecar	10k	561 µs	56.1 ns	Ca/K 2D model
Butera respiratory	1k	341.62 µs	341.62 ns	bounded RK4 Pre-Bötzinger + INaP

Most simple spiking models in this table retain their previously documented integration contract. Butera respiratory was remeasured on 2026-05-31 after bounded RK4 hardening across Python, Rust, Go, Julia, and Rust safety surfaces. Older rows in this section retain their 2026-04-05 measurement date until their own model-specific hardening pass is completed.

Rate / Mean-Field (neurons/rate.rs)

Model	Steps	Median	Per step	Notes
Threshold linear rate	100k	39.4 µs	0.4 ns	Rectified linear
Sigmoid rate	100k	921 µs	9.2 ns	Sigmoidal firing rate
TsodyksMarkram STP	10k	86.4 µs	8.6 ns	Short-term plasticity
Parallel spiking	10k	60.9 µs	6.1 ns	Multi-subunit LIF
Astrocyte	10k	190 µs	19.0 ns	Ca²⁺/IP3/SERCA dynamics
Compte WM	10k	232 µs	23.2 ns	Working memory NMDA
LTC	10k	409 µs	40.9 ns	Liquid time constant
FractionalLIF	10k	739 µs	73.9 ns	Fractional-order memory kernel
LeakyCompeteFire	10k	972 µs	97.2 ns	WTA lateral inhibition (4 units)
AmariNeuralField	10k	24.2 ms	2.42 µs	32-unit neural field
Siegert	100k	44.4 ms	444 ns	Transfer function with erf()

Hardware Neuromorphic (neurons/hardware.rs)

Model	Steps	Median	Per step	Notes
TrueNorth	100k	94.2 µs	0.9 ns	IBM integer LIF
SpiNNaker2	100k	103 µs	1.0 ns	Fixed-point 3-compartment
Akida	100k	60.1 µs	0.6 ns	Event-driven threshold
Loihi CUBA	100k	342 µs	3.4 ns	Intel 2-variable integer IF
Loihi2	100k	416 µs	4.2 ns	Intel 3-compartment integer
SpiNNaker LIF	10k	44.5 µs	4.5 ns	ARM LIF emulation
DPI	100k	1.27 ms	12.7 ns	Differential pair integrator
BrainScaleS AdEx	1k	31.4 µs	31.4 ns	Analog accelerated AdEx
NeuroGrid	1k	43.6 µs	43.6 ns	Subthreshold analog, 2-comp

AI-Optimised (neurons/ai_optimized.rs)

Model	Steps	Median	Per step	Notes
Meta-plastic	10k	20.9 µs	2.1 ns	Learning rate adaptation
Differentiable surrogate	10k	38.4 µs	3.8 ns	Surrogate gradient SNN
Multi-timescale	10k	99.1 µs	9.9 ns	Fast + slow dynamics
Predictive coding	10k	126 µs	12.6 ns	Prediction error driven
Attention-gated	10k	175 µs	17.5 ns	Gate modulates response
Compositional binding	10k	184 µs	18.4 ns	Phase-based variable binding
Self-referential	10k	404 µs	40.4 ns	Self-modifying tau
Continuous attractor	10k	6.58 ms	658 ns	16-unit bump attractor
Arcane	10k	1.37 ms	137 ns	Deep accumulator + novelty

Multi-Compartment (neurons/multi_compartment.rs)

Model	Steps	Median	Per step	Notes
Dendrify	1k	20.2 µs	20.2 ns	Simplified dendritic
Two-compartment LIF	10k	26.9 µs	2.7 ns	Soma + dendrite LIF
Rall cable	1k	75.6 µs	75.6 ns	5-compartment cable
Pinsky-Rinzel	1k	122 µs	122 ns	2-comp pyramidal
Marder STG	1k	138 µs	138 ns	Stomatogastric, 6 currents
Booth-Rinzel	1k	253 µs	253 ns	Motoneuron soma+dendrite
Hay L5 pyramidal	1k	591 µs	591 ns	3-comp (soma+trunk+apical)

Maps — Additional (neurons/maps.rs)

Model	Steps	Median	Per step	Notes
Ibarz-Tanaka	100k	803 µs	8.0 ns	Chaotic bursting map
Cazelles	100k	955 µs	9.6 ns	Coupled map lattice
Medvedev	100k	1.13 ms	11.3 ns	Reduce-and-fire map
Courage-Nekorkin	100k	1.34 ms	13.4 ns	FHN-like map
Rulkov	100k	1.67 ms	16.7 ns	Slow-fast bursting map
Chialvo	100k	1.75 ms	17.5 ns	2D excitable map

Sensory — Additional (neurons/sensory.rs)

Model	Steps	Median	Per step	Notes
Outer hair cell	10k	106 µs	10.6 ns	Prestin electromotility
Taste receptor	10k	120 µs	12.0 ns	Gustatory transduction
Cone photoreceptor	10k	135 µs	13.5 ns	Colour vision, cGMP cascade

Trivial IF Variants (neurons/trivial.rs)

Model	Steps	Median	Per step	Notes
Perfect integrator	100k	206 µs	2.1 ns	Pure integration, simplest IF
KLIF	100k	224 µs	2.2 ns	Kernel-based LIF
Complementary LIF	100k	265 µs	2.7 ns	Dual-pathway (v_pos + v_neg)
Sigma-delta	100k	311 µs	3.1 ns	Spike-based A/D encoder
Integer QIF	100k	368 µs	3.7 ns	Integer arithmetic QIF
Inhibitory LIF	100k	447 µs	4.5 ns	Self-inhibiting trace
Gated LIF	100k	555 µs	5.6 ns	Input gating mechanism
Parametric LIF	100k	810 µs	8.1 ns	Learnable alpha, beta
Stochastic LIF	10k	85.1 µs	8.5 ns	Gaussian noise injection
Quadratic IF	100k	875 µs	8.7 ns	Quadratic subthreshold
Non-resetting LIF	10k	165 µs	16.5 ns	Threshold adaptation, no reset
Energy LIF	10k	183 µs	18.3 ns	Energy-budget constrained
MAT	10k	187 µs	18.7 ns	Multi-timescale adaptive threshold
SFA	10k	196 µs	19.6 ns	Spike frequency adaptation
NLIF	10k	261 µs	26.1 ns	Nonlinear leak + recovery
Adaptive threshold IF	10k	288 µs	28.8 ns	Dynamic threshold
Escape rate	10k	517 µs	51.7 ns	Stochastic, exp() hazard
Theta neuron	100k	6.97 ms	69.7 ns	Phase model, cos()/sin()
CFC	100k	11.6 ms	116 ns	Closed-form continuous, exp()

Trivial models are the fastest — no sub-stepping, minimal state. Measured 2026-04-05 on i5-11600K @ 3.90 GHz.

---

Analysis Modules (analysis_bench)

22 modules, 84 benchmark points. Full results in rust-analysis-engine.md.

Highlights (fastest per category)

Category	Function	Input	Median
Basic	firing_rate	100 spikes	24 ns
Waveform	waveform_amplitude	64 samples	39 ns
Variability	fano_factor	100 spikes	98 ns
Basic	spike_times	100 spikes	142 ns
Distance	isi_distance	100 spikes	254 ns
Temporal	change_point_detection	1K spikes	347 ns
Decoding	bayesian_decode	20n, 8 stim	982 ns
Patterns	spike_directionality	5K spikes	68 µs
Spectral	power_spectrum	100K samples	5.7 ms
GPFA	gpfa	4n, 500t, 5 iter	5.8 ms

Scaling Characteristics

Function	100 → 100K	Scaling
spike_times	142 ns → 102 µs	O(n)
power_spectrum	4.0 µs → 5.7 ms	O(n log n)
sample_entropy	46 µs → (n/a)	O(n²)
functional_connectivity	—	O(n² × T)

---

Benchmark Files

File	Harness	Content
engine/benches/full_bench.rs	Criterion	Core + neurons (31 benchmarks)
engine/benches/analysis_bench.rs	Criterion	22 analysis modules (84 benchmarks)
engine/benches/bitstream_bench.rs	Criterion	Bitstream-specific deep dive
engine/benches/scaling_bench.rs	Criterion	Network scaling tests

Criterion Output

JSON results stored in engine/target/criterion/*/new/estimates.json after each run. Use cargo-criterion or the HTML reports in engine/target/criterion/report/ for trend analysis.