Competitive Landscape: Neuromorphic Computing Frameworks¶
An honest comparison of SC-NeuroCore with peer frameworks. Every claim is backed by measured data or cited literature. Unverified claims are marked explicitly.
Last updated: 2026-05-14 (v3.14.0 pre-submission claim audit)
Performance and count entries below are retained as release-evidence snapshots. Rerun the cited benchmark or inventory command before reusing exact numbers in the JOSS paper, README headline, or external outreach.
1. Framework Overview¶
| Framework | Primary Focus | Language | License | First Release |
|---|---|---|---|---|
| SC-NeuroCore | Stochastic computing + FPGA co-design | Python + Rust | AGPL-3.0 | 2024 |
| snnTorch | PyTorch-native SNN training | Python | MIT | 2021 |
| Norse | Bio-inspired SNN on PyTorch | Python | LGPL-3.0 | 2020 |
| Lava | Intel Loihi neuromorphic SDK | Python | BSD-3 | 2021 |
| Brian2 | Flexible neuroscience simulator | Python + C++ | CeCILL-2.1 | 2014 |
| Nengo | Large-scale brain modelling | Python | Other | 2013 |
| BindsNET | Biologically plausible SNN | Python | AGPL-3.0 | 2018 |
2. Feature Parity Matrix¶
| Feature | SC-NeuroCore | snnTorch | Norse | Lava | Brian2 |
|---|---|---|---|---|---|
| Stochastic computing (bitstream) | Yes | — | — | — | — |
| Bit-true RTL co-simulation | Yes | — | — | — | — |
| Verilog / FPGA synthesis | Yes | — | — | Loihi only | — |
| IR compiler → SystemVerilog | Yes | — | — | — | — |
| Equation → Verilog compiler | Yes | — | — | — | — |
| IR compiler → MLIR/CIRCT | Yes | — | — | — | — |
| Rust SIMD engine | Yes (historical Criterion pack benchmark; rerun before citation) | — | — | — | — |
| Surrogate gradient training | Yes (6 surrogates, 12 cells) | Yes | Yes | Yes | — |
PyTorch nn.Module SNN |
Yes (+ SC export) | Yes | Yes | — | — |
| GPU acceleration | PyTorch + CuPy | PyTorch | PyTorch | — | — |
| Neuron models | 173 | 11 | 6 | 3 | Arbitrary |
| Rust neuron models (PyO3) | 173 | — | — | — | — |
| NetworkRunner (fused loop) | 160 models | — | — | — | — |
| Network simulation backends | 3 (Python, Rust, MPI) | PyTorch | PyTorch | Lava | C++ codegen |
| MPI distributed simulation | Yes | — | — | — | — |
| Pre-trained model zoo | 10 configs, 3 weights | — | — | — | — |
| Spike train analysis | 127 functions | — | — | — | — |
| Visualization plots | 14 | — | — | — | — |
| Advanced plasticity rules | 13 | — | — | — | — |
| STDP / R-STDP plasticity | Yes | — | Yes | Yes | Yes |
| Quantum hybrid circuits | Yes | — | — | — | — |
| Hyperdimensional computing | Yes | — | — | — | — |
| Formal verification | 7 modules, 68 props | — | — | — | — |
| Sobol low-discrepancy encoding | Yes | — | — | — | — |
| Multi-head attention (SC) | Yes | — | — | — | — |
| Connectome generators | Yes | — | — | — | Yes |
| JAX JIT training | Yes | — | — | — | — |
| CuPy sparse GPU | Yes | — | — | — | — |
| AI-optimized neurons | 9 (ArcaneNeuron + 8) | — | — | — | — |
| Identity substrate | Yes (persistent SNN + checkpoint) | — | — | — | — |
| Neural data compression | 6 codecs (ISI, predictive, delta, streaming, AER, waveform) | — | — | — | — |
| Trainable per-synapse delays | Yes (DelayLinear, differentiable) | — | — | — | — |
| NIR support | Yes (FPGA backend) | Yes | Yes | Yes | — |
| conda-forge recipe draft | Draft only; not published | Yes | — | — | Yes |
| PyPI package | Yes | Yes | Yes | Yes | Yes |
Capability coverage map¶
quadrantChart
title Framework Capabilities (technical breadth vs FPGA depth)
x-axis "Shallow FPGA" --> "Deep FPGA"
y-axis "Narrow Scope" --> "Broad Scope"
quadrant-1 "Full Stack"
quadrant-2 "Broad but no HW"
quadrant-3 "Narrow, no HW"
quadrant-4 "HW-focused"
SC-NeuroCore: [0.85, 0.9]
snnTorch: [0.1, 0.65]
Norse: [0.1, 0.55]
Lava: [0.5, 0.45]
Brian2: [0.05, 0.7]
Nengo: [0.15, 0.6]
Flexi-NeurA: [0.7, 0.3]Selected SC-NeuroCore capabilities¶
- Stochastic computing — Includes bitstream-level simulation, packed AND+popcount operations, and Sobol LDS encoding
- FPGA co-design — IR compiler emits synthesizable SystemVerilog and MLIR/CIRCT, with bit-exact Python↔Verilog co-simulation
- Formal verification — SymbiYosys jobs and formal statements; use the generated README inventory for current counts and scope
- Rust SIMD engine — Architecture-specific dispatch and model wrappers; use the generated README inventory for current counts
- Network simulation — 3 backends (Python, Rust, MPI), 6 topology generators, 10 model zoo configs, 3 pre-trained weight sets
- Analysis toolkit — Spike train analysis modules; inspect the exact release inventory before comparing counts with peer packages
- ArcaneNeuron — research model with five coupled subsystems
- Identity substrate — persistent spiking network with checkpointing, trace encoding/decoding, L16 Director cybernetic closure
- Quantum-SC bridge — IBM Heron r2 noise model, parameter-shift gradients, VQE pipeline
Where others lead¶
- snnTorch — Larger community, more tutorials, established research ecosystem with 40+ citing publications
- Norse — Bio-plausible SNN equations with auto-differentiation, active research community
- Lava — Intel Loihi hardware workflows and event-driven execution
- Brian2 — Arbitrary neuron equations (string-based), 3000+ publications, gold standard for computational neuroscience
- Nengo — Large-scale brain modelling (100K+ neurons), NEF (Neural Engineering Framework), SpiNNaker support
- Flexi-NeurA — Bit-exact Python/RTL co-simulation for FPGA/ASIC SNN deployment (arXiv:2602.18140, Feb 2026)
3. Performance Comparison¶
3.1 Inference Throughput (single-sample, CPU)¶
Measured on Intel i5-11600K (AVX-512), Python 3.12.
| Framework | Operation | Throughput | Source |
|---|---|---|---|
| SC-NeuroCore (Rust) | LIF neuron step | 456 Mstep/s | Criterion bench |
| SC-NeuroCore (Rust) | Pack 1M bits | Historical Criterion result; rerun before citation | Criterion bench |
| SC-NeuroCore (Python) | LIF neuron step | 1.07 Mstep/s | benchmark_suite.py |
| Brian2 | LIF neuron (compiled) | ~10 Mstep/s | Brian2 docs (estimate) |
| snnTorch | LIF neuron (PyTorch) | ~5 Mstep/s | PyTorch CPU baseline |
Note: snnTorch and Norse are designed for GPU batch training, not single-sample CPU inference. Their GPU throughput far exceeds CPU numbers above.
3.2 Brunel Balanced Network (10,000 neurons)¶
SC-NeuroCore Brunel benchmark (20 variants), measured on same hardware:
| Variant | Wall time (s) | Spike rate (Hz) |
|---|---|---|
| V01 baseline (LIF, 1K neurons) | 0.18 | 47.2 |
| V05 Izhikevich (1K neurons) | 0.31 | 52.8 |
| V14 Sobol bitstream (1K) | 0.22 | 45.1 |
| V18 Numba JIT (1K) | 0.019 | 47.2 |
Brian2 comparison (same network, 1K excitatory + 250 inhibitory):
| Metric | SC-NeuroCore | Brian2 2.10.1 |
|---|---|---|
| V01 wall time | 0.18 s | 0.21 s |
| V01 ratio | 1.17× faster | baseline |
Honest framing: The 1.17× figure is for 1K-neuron Python-path simulation.
The committed Brunel balanced-network scaling artefact
benchmarks/results/rust_scaling_benchmark.json records 39-202x wall-clock
ratios against Brian2 for its 10K-100K Rust rows, but those rows measure an
unconnected fixed-point LIF layer under constant external drive rather than
the recurrent Brunel network, so they are not a dynamics-parity result. The
dynamics-parity comparison in the same artefact is the NumPy dense lane:
2.4x faster than Brian2 at 1,000 neurons, 1.3x at 5,000, and 0.75x (slower)
at 10,000. Brian2 is faster at small networks where its C++ code generation
amortises overhead.
3.3 GPU Scaling (NVIDIA RTX A6000)¶
| Neurons | Synapses | Wall (s) | Syn events/s |
|---|---|---|---|
| 1,000 | 100K | 1.55 | 3.2 M |
| 5,000 | 2.5M | 2.74 | 29.0 M |
| 20,000 | 40M | 8.80 | 59.2 M |
| 50,000 | 250M | 35.4 | 51.9 M |
4. FPGA Resource Estimates¶
SC-NeuroCore MNIST classifier (Yosys synthesis, target: iCE40 UP5K):
| Module | LUTs | FFs | BRAMs |
|---|---|---|---|
sc_lif_neuron |
89 | 48 | 0 |
sc_bitstream_encoder |
34 | 17 | 0 |
sc_dense_layer_core |
~2,400 | ~800 | 2 |
| 16→10 classifier | ~56K | ~18K | 16 |
This repository includes a path from selected models to synthesizable RTL. Compare generated modules and target-specific evidence before drawing a hardware capability conclusion across frameworks.
5. Accuracy Benchmarks¶
MNIST Digit Classification¶
| Method | Accuracy | Framework | Status |
|---|---|---|---|
| Float baseline (sklearn) | 94.2% | SC-NeuroCore | Verified (stored artifact) |
| Quantized Q8.8 | 94.2% | SC-NeuroCore | Verified (stored artifact) |
| Stochastic computing (L=1024) | 94.0% | SC-NeuroCore | Verified (stored artifact) |
| ConvSpikingNet (learnable params) | 99.49% | SC-NeuroCore | Verified by benchmarks/results/mnist_conv_accuracy_reproducibility.json |
| Surrogate gradient SNN | ~97% | snnTorch | Published |
| Surrogate gradient SNN | ~96% | Norse | Published |
SC-NeuroCore's ConvSpikingNet achieved 99.49% on MNIST with
learnable beta/threshold, cosine LR, and data augmentation in the committed
training run recorded by benchmarks/results/mnist_conv_accuracy_reproducibility.json.
Use that manifest as the evidence boundary for the public claim.
6. Possible Starting Points¶
| Use Case | Relevant Option | Why Consider It |
|---|---|---|
| FPGA RTL research | SC-NeuroCore | IR→Verilog/MLIR path; board deployment requires separate validation |
| Intel Loihi hardware | Lava | Native Loihi support |
| PyTorch SNN training | snnTorch or SC-NeuroCore | snnTorch has larger community; SC-NeuroCore adds SC export + FPGA path |
| Computational neuroscience | Brian2 | Arbitrary neuron equations |
| Bio-plausible learning | Norse or BindsNET | STDP/bio-learning focus |
| Large-scale brain models | Nengo | NEF, SpiNNaker support |
| Stochastic + quantum hybrid | SC-NeuroCore | Research bridge between these model paths |
| Selected RTL property checks | SC-NeuroCore | SymbiYosys jobs; scope and counts depend on the exact tree |
7. Community and Ecosystem¶
| Metric | SC-NeuroCore | snnTorch | Norse | Lava | Brian2 |
|---|---|---|---|---|---|
| GitHub stars | 4 | ~1.5K | ~500 | ~600 | ~1K |
| PyPI downloads/month | < 50 | ~15K | ~3K | ~2K | ~30K |
| Publications citing | 0 | 40+ | 20+ | 15+ | 3000+ |
| First-party tutorials | 85 | 15 | 8 | 10 | 30+ |
| Active maintainers | 1 | 5+ | 3+ | 10+ | 5+ |
Honest assessment: SC-NeuroCore has 4 GitHub stars and zero citations. The competitive advantage is purely technical (stochastic+FPGA). The adoption gap is the problem, not the engineering. A published paper (JOSS submission planned June 2026), clean MNIST artifact, and external validation are needed to translate engineering quality into credibility.
8. References¶
- Eshraghian et al., "Training Spiking Neural Networks Using Lessons From Deep Learning," Proc. IEEE, 2023 (snnTorch)
- Pehle & Pedersen, "Norse — A Library for Gradient-Based Learning with Spiking Neural Networks," 2021
- Intel Labs, "Lava: An Open-Source Software Framework for Neuromorphic Computing," 2021
- Stimberg et al., "Brian 2: an intuitive and efficient neural simulator," eLife, 2019
- Bekolay et al., "Nengo: a Python tool for building large-scale functional brain models," Front. Neuroinform., 2014
- Alaghi & Hayes, "Survey of Stochastic Computing," ACM TECS, 2013
- NeuroBench Collaboration, "NeuroBench: A Framework for Benchmarking Neuromorphic Computing Algorithms and Systems," 2023