Skip to content

Architecture

Jump to bottom
Raul Montoya Cardenas edited this page Jul 29, 2026 · 2 revisions

Architecture

Owns

  • In-memory NIR graph model and closed node type set
  • Wire-accurate type strings for HDF5 interoperability
  • Reading/writing .nir (HDF5) with graph-level round-trip fidelity
  • Structured errors and basic structural validation

Does not own

  • Training or simulation of SNNs
  • Hardware mapping (silicon-bridge)
  • Framework converters (producer/consumer crates)
  • Custom NIR nodes until upstream adopts them

Layers

┌─────────────────────────────────────┐
│  NirGraph + NirNode + Tensor        │  always available
├─────────────────────────────────────┤
│  io::wire (keys, WIRE_TYPES)        │  always available
├─────────────────────────────────────┤
│  io::read / write (hdf5 feature)    │  native libhdf5
└─────────────────────────────────────┘

Design rules (AGENTS)

  • Wire type names match neuromorphs/NIR exactly (CubaLIF, not CurrLIF)
  • Exhaustive NirNode enum for consumers (e.g. silicon-bridge)
  • No Python in repo/CI
  • Consumers depend on this crate for I/O — do not reimplement HDF5 decode elsewhere
  • Default branch may be Main (capital M) on GitHub

Upstream

Last updated: July 29, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: df72fc9 (main)

Clone this wiki locally