Skip to content

neuromorphic.md

Latest commit

 

History

History
123 lines (93 loc) · 4.51 KB

neuromorphic.md

File metadata and controls

123 lines (93 loc) · 4.51 KB
Raw

Neuromorphic Compiler (NIR)

Compiles HoloScript to Neuromorphic Intermediate Representation (NIR) — enabling HoloScript spatial scenes to run on neuromorphic hardware chips and WebGPU-based Spiking Neural Network simulations.

Overview

Neuromorphic computing processes information using networks of artificial neurons that fire spikes — like biological brains — instead of traditional floating-point operations. It offers extreme energy efficiency for pattern recognition, sensory processing, and autonomous decision-making.

HoloScript has two neuromorphic output targets:

Compiler Target Output
NIRCompiler Physical chips NIR JSON + bitfile
NIRToWGSL Browser (WebGPU) WGSL compute shader 
# Compile to neuromorphic chip (Intel Loihi2)
holoscript compile snn.holo --target nir --nir-backend loihi2 --output ./chip/

# Compile to WebGPU SNN simulation (browser)
holoscript compile snn.holo --target nir-wgsl --output ./web/

Supported Hardware

Backend Chip Organization
loihi2 Intel Loihi 2 Intel Labs
spinnaker SpiNNaker 2 University of Manchester
synsense Xylo / Speck SynSense AG

The nir-wgsl backend runs the same SNN model in a WebGPU compute shader — no hardware required, runs in any WebGPU-capable browser.

LIF Neuron Model

HoloScript uses the Leaky Integrate-and-Fire (LIF) neuron as the default model:

composition "SpatialSNN" {
  template "Neuron" {
    @physics         // membrane dynamics
    state {
      membrane_potential: -70.0   // mV
      threshold:         -55.0    // mV
      resting_potential: -70.0    // mV
      time_constant:      20.0    // ms
    }
  }

  spatial_group "SensoryLayer" {
    object "N1" using "Neuron" { position: [0, 0, 0] }
    object "N2" using "Neuron" { position: [1, 0, 0] }
    object "N3" using "Neuron" { position: [2, 0, 0] }
  }

  logic {
    // Spike propagation: N1 fires → N2 → N3
    on_spike("N1") {
      N2.membrane_potential += 15.0
    }
    on_spike("N2") {
      N3.membrane_potential += 15.0
    }
  }
}

NIR JSON Output

The NIR compiler produces a JSON graph consumable by chip SDKs:

{
  "neurons": [
    { "id": "N1", "type": "LIF", "tau_m": 20, "v_thresh": -55, "v_reset": -70 },
    { "id": "N2", "type": "LIF", "tau_m": 20, "v_thresh": -55, "v_reset": -70 }
  ],
  "synapses": [{ "pre": "N1", "post": "N2", "weight": 1.0, "delay": 1 }],
  "inputs": ["N1"],
  "outputs": ["N3"]
}

WebGPU SNN (Browser)

The nir-wgsl pipeline simulates the SNN in a WebGPU compute shader — enabling real-time neuromorphic inference in the browser alongside the spatial scene:

holoscript compile snn.holo --target nir-wgsl
# → snn.compute.wgsl (runs on GPU as compute shader)
# → snn-bridge.js   (connects spike outputs to scene objects)

This is backed by @holoscript/snn-webgpu (successor to the retired @holoscript/snn-poc PoC).

Use Cases

  • Energy-efficient NPCs — AI agents that process sensory input on Loihi2 at milliwatt power
  • Spatial pattern recognition — neuromorphic vision on AR glasses
  • Human-in-the-loop — @biofeedback trait connects biological signals to SNN inputs
  • Research — visual scene editor for neuroscience SNN experiments
  • Edge AI — robotics + IoT devices using SpiNNaker

Related Packages

Package Role
@holoscript/snn-webgpu GPU-accelerated SNN compute shaders + reference LIF implementation

Compiler Options

Option Default Description
--nir-backend wgsl loihi2, spinnaker, synsense, wgsl
--nir-timestep 1.0 Simulation timestep in ms
--nir-duration 100 Default sim duration in ms
--nir-dt 0.1 Integration step for continuous model

See Also