All numbers from reproducible benchmark runs. Baseline: Brian2 on identical hardware.
| Scale | SC-NeuroCore (Rust) | Brian2 | Speedup |
|---|---|---|---|
| 1K neurons | 0.029 s | 2.689 s | 93× |
| 5K neurons | 0.285 s | 4.681 s | 16× |
| 10K neurons | 0.172 s | 6.754 s | 39× |
| 50K neurons | 0.582 s | 31.03 s | 53× |
| 100K neurons | 1.153 s | 232.3 s | 202× |
Throughput at 100K: 27.7 billion synaptic events/second
| Framework | Backend | Execution Time | Peak Memory | Speedup vs Brian2 |
|---|---|---|---|---|
| SC-NeuroCore | Rust | 0.081 s | 0.3 MB | 832× |
| SC-NeuroCore | NumPy | 12.898 s | 26.4 MB | 5.2× |
| Norse | PyTorch | 41.740 s | 18.9 MB | 1.6× |
| snnTorch | PyTorch | 49.149 s | 48.5 MB | 1.4× |
| Brian2 | NumPy | 67.686 s | 87.9 MB | 1.0× |
A heavier full-network reference workload than the 300 ms Brunel scaling above, so the absolute speedup vs Brian2 differs — both runs are committed benchmark artefacts. This comparison also benchmarks the spiking-network frameworks Norse and snnTorch on identical hardware (validated 2026-06-06).
SIMD dispatch: AVX-512 → AVX2 → NEON → SVE → RVV → portable fallback. Auto-detected at runtime.
| Layer | Neurons | Synapses | Latency/step | Throughput |
|---|---|---|---|---|
| SCDenseLayer(8×4, L=256) | 4 | 32 | 1,293 µs | 6.3 MOps |
| SCDenseLayer(16×8, L=512) | 8 | 128 | 2,446 µs | 26.8 MOps |
| VectorizedSCLayer(16×8, L=512) | 8 | 128 | 348 µs | 188 MOps |
| VectorizedSCLayer(64×32, L=1024) | 32 | 2048 | 2,476 µs | 847 MOps |
| Architecture | Method | Accuracy |
|---|---|---|
| FC-SNN (784→128→128→10) | Surrogate gradient, 10 epochs | 95.5% |
| FC-SNN + learnable membrane | Learnable beta/threshold | 97.7% |
| ConvSpikingNet | Learnable beta/threshold + cosine LR | 99.49% |
Parity check: snnTorch identical architecture → 95.8%. SC-NeuroCore → 95.5% (within noise). Learnable parameters push to 99.49%.
ISI+Huffman, Predictive (4 learnable predictors), Delta, Streaming, AER, WaveformCodec. Unified API: get_codec(name), recommend_codec(). Rust backend (780×).