Product Overview¶
SC-NeuroCore is a stochastic-computing and neuromorphic hardware co-design toolkit. It helps teams move from numerical neuron and network experiments to auditable implementation artefacts: stochastic bitstreams, fixed-point representations, optional accelerated simulation, generated RTL, synthesis reports, and readiness evidence.
Plain-language summary¶
Most neural-network software optimises floating-point training and inference. SC-NeuroCore focuses on a different question: when can neural computation be represented as probability, spike timing, or stochastic bitstreams, and when can that representation be moved toward energy-aware hardware?
The project gives researchers and engineers one place to explore:
- stochastic encoders and bitstream arithmetic;
- spiking neuron and synapse models;
- network simulation and training paths;
- hardware export and co-simulation workflows;
- benchmark and parity evidence;
- evidence categories for industrial readiness.
Core workflow¶
- Model a neuron, synapse, network, or stochastic operator in Python.
- Validate numerical behaviour with module-specific tests, parity artefacts, or benchmark harnesses.
- Select an execution path: pure Python, optional Rust engine, optional training backend, or hardware-generation route.
- Export evidence: JSON benchmark results, generated RTL, synthesis reports, notebook evidence, or readiness assessments.
- Promote only claims that are backed by committed artefacts.
What a completed workflow produces¶
SC-NeuroCore is most useful when the output is not just a plot or a trained model, but an evidence packet that another engineer can inspect:
- a deterministic Python or Rust reference result;
- a quantisation or precision manifest that states the numerical contract;
- generated RTL or deployment metadata when the hardware path is used;
- benchmark or synthesis artefacts with command and environment context;
- a documented evidence boundary that states what the artefact does not prove.
This is the practical bridge from research code to industrial diligence. A lab can use the same workflow to explore a biological model, an FPGA engineer can audit the emitted arithmetic, and a commercial reviewer can see which claims are ready, which are local-only, and which remain open evidence gaps.
Why stochastic computing matters¶
Stochastic computing represents values as probabilities over bitstreams. That can trade exact arithmetic for compact logic, fault tolerance, and hardware-friendly approximate computation. In neuromorphic systems, this aligns naturally with spikes, rates, stochastic synapses, and noisy physical substrates.
SC-NeuroCore treats stochasticity as a measurable engineering parameter rather than a demo effect. The documentation and benchmark policy require error bounds, parity checks, or committed artefacts before a result becomes a public claim.
What the repository contains¶
| Layer | Purpose |
|---|---|
| Public Python package | Stable entry points for simulation, encoding, layers, selected hardware helpers, and documented APIs. |
| Optional Rust engine | Acceleration for selected hot paths when the local environment provides the engine. |
| Training and interop extras | PyTorch, NIR, JAX, CuPy, quantum, and Studio paths installed only when requested. |
| HDL and hardware evidence | Verilog/SystemVerilog generation, primitive RTL, synthesis reports, and deployment guides. |
| Research notebooks | Executable walkthroughs and evidence notebooks for selected workflows. |
| Polyglot research mirrors | Julia, Go, Mojo, and Rust counterparts for selected kernels and comparison benchmarks. |
Evidence boundary¶
SC-NeuroCore is a broad research and engineering stack, not a blanket certification claim. The public documentation distinguishes:
- stable package surfaces from source-only research modules;
- measured benchmark artefacts from planned comparison gaps;
- synthesis/resource evidence from power or energy evidence;
- readiness assessment from regulated deployment approval;
- optional dependency paths from default installation guarantees.
Where to start¶
- New users: Getting Started.
- Structured learning: Learning Path.
- Notebook users: Notebook Guide.
- API consumers: API Reference Index.
- Benchmark readers: Benchmarks and Cross-Framework Benchmark Evidence.
- Commercial or industrial evaluators: Applications and Market and Industrial Applications.