Performance Benchmarks

All numbers from reproducible benchmark runs. Baseline: Brian2 on identical hardware.

Rust Engine vs Brian2 (Brunel Balanced Network, 300 ms)
1K neurons
93×
5K neurons
16×
10K neurons
39×
50K neurons
53×
100K neurons
202×
ScaleSC-NeuroCore (Rust)Brian2Speedup
1K neurons0.029 s2.689 s93×
5K neurons0.285 s4.681 s16×
10K neurons0.172 s6.754 s39×
50K neurons0.582 s31.03 s53×
100K neurons1.153 s232.3 s202×

Throughput at 100K: 27.7 billion synaptic events/second

SIMD Bitstream Operations (1M bits)
190
Gbit/s SIMD Popcount
108.4
Gbit/s Portable
0.8
Gbit/s Pack

SIMD dispatch: AVX-512 → AVX2 → NEON → SVE → RVV → portable fallback. Auto-detected at runtime.

GPU Acceleration (wgpu)
21×
Speedup (GTX 1060)
Cross-platform
Vulkan / Metal / DX12
Layer Performance (NeuroBench Microbench)
LayerNeuronsSynapsesLatency/stepThroughput
SCDenseLayer(8×4, L=256)4321,293 µs6.3 MOps
SCDenseLayer(16×8, L=512)81282,446 µs26.8 MOps
VectorizedSCLayer(16×8, L=512)8128348 µs188 MOps
VectorizedSCLayer(64×32, L=1024)3220482,476 µs847 MOps
MNIST SNN Training
ArchitectureMethodAccuracy
FC-SNN (784→128→128→10)Surrogate gradient, 10 epochs95.5%
FC-SNN + learnable membraneLearnable beta/threshold97.7%
ConvSpikingNetLearnable beta/threshold + cosine LR99.49%

Parity check: snnTorch identical architecture → 95.8%. SC-NeuroCore → 95.5% (within noise). Learnable parameters push to 99.49%.

Neural Data Compression
24×
Waveform (1024-ch)
50–750×
Spike Raster
6
Codecs

ISI+Huffman, Predictive (4 learnable predictors), Delta, Streaming, AER, WaveformCodec. Unified API: get_codec(name), recommend_codec(). Rust backend (780×).