Hardware Targets

Overview

THRINDEX separates compilation from execution. A model is compiled once to a .thx artifact that carries the target it was built for. On a machine with the appropriate hardware and driver, the artifact is loaded and run by the corresponding backend.

Developer machine           Device (e.g. Raspberry Pi + AKD1500)
──────────────────          ───────────────────────────────────────
thx.compile(model,          akida-runtime model.fbz
  "model.thx",                  ← reads JSON batch from stdin
  target="akida-akd1500")       → writes JSON output to stdout
        ↓
akida_compile.py
  model.thx → model.fbz

BrainChip AKD1500

The akida-akd1500 target maps THRINDEX models to BrainChip’s Akida 1.0 neuromorphic processor.

Capability

Field	Value
Target name	`akida-akd1500`
Native timestep	1.0 ms
Native delay support	None (`native_delay_max_steps = 0`)
Delay fallback	Reject
Precision	4-bit integer per tensor
Weight range	`[-7, 7]` (int8 container, 4-bit values)

What the AKD1500 supports

AKD1500 executes feedforward integer networks. Supported layer types:

Dense (snn.Dense) — fully connected, no delays
Conv2d (snn.Conv2d) — 2D convolution, no delays

What the AKD1500 rejects

Condition	Error
LIF layer in artifact	E0401
Synaptic delays in artifact	E0402
Non-finite (NaN / Inf) weights	E0403
Artifact compiled for a different target	E0404
Input with more than one timestep (`T > 1`)	E0407

AKD1500 is not an SNN backend. It does not implement leaky integrate-and-fire dynamics, temporal state, or spike-triggered reset. If your model uses LIF neurons, use target="sim" instead. See E0401 for a full explanation.

Weight quantization

Weights are quantized from f32 to 4-bit signed integers during the Tier-1 compile step:

scale = max(|W|) / 7
W_int = round(W / scale)  — clamped to [-7, 7]

The quantization error is at most scale / 2 per weight. For models with large weight ranges, this error can be significant. Training with weight regularization that keeps max(|W|) small reduces quantization error.

Deployment: two-tier setup

AKD1500 deployment uses two separate machines: Tier 1 — Developer machine (Python + MetaTF) Requires: thrindex + akida Python package from BrainChip.

import thrindex as thx

# Compile to .thx — validates the model for akida-akd1500
thx.compile(model, "model.thx", target="akida-akd1500")

Then convert to the BrainChip .fbz runtime format:

python akida_compile.py model.thx model.fbz

akida_compile.py is included in crates/thrindex-backends/akida/python/ in the THRINDEX repository. Tier 2 — Device (Raspberry Pi + AKD1500 via M.2) Requires: THRINDEX built with --features hardware and the BrainChip Engine Library.

# Set path to Engine Library source (obtained from BrainChip)
export THRINDEX_AKIDA_ENGINE_PATH=/path/to/engine/engine

# Build the runtime binary
cargo build -p thrindex-backend-akida --features hardware --release

# Run inference
echo '{"batch": [[[0.5, 0.2, 0.8, ...]]]}' | \
  akida-runtime model.fbz --pcie-addr 0001:01:00.0

The runtime reads a JSON batch on stdin and writes a JSON result on stdout:

// stdin — shape [N_samples, T=1, n_features]
{"batch": [[[0.5, 0.2, 0.8, 0.1, ...]]]}

// stdout on success — shape [N_samples, n_outputs]
{"outputs": [[0.03, 0.91, 0.02, 0.04]]}

// stdout on error
{"error": "E0407: ..."}

Environment variables

Variable	Required	Description
`THRINDEX_AKIDA_ENGINE_PATH`	At build time	Path to the BrainChip Engine Library source drop
`THRINDEX_AKIDA_DEVICE`	At runtime	PCIe address of the AKD1500 (default: `0001:01:00.0`)

Getting the Engine Library

The Engine Library is proprietary BrainChip software distributed to registered AKD1500 customers. It is not included in the THRINDEX repository. Contact BrainChip to obtain it.

Conformance

akida-akd1500 is permanently excluded from the THRINDEX Certified [v0] conformance suite. The Certified badge measures spike-raster equivalence between a hardware backend and the reference simulator. AKD1500 does not implement SNN dynamics — it performs integer feedforward inference — and cannot produce a spike raster that is equivalent to the LIF simulator output. When the conformance harness runs, it prints:

EXCLUDED akida-akd1500: reason=non-snn-backend; metric=top-1-accuracy-frozen-brainchip-dataset

The alternative metric (top-1 accuracy on the frozen BrainChip dataset) is the appropriate correctness measure for AKD1500 inference. See Conformance Overview for details.

Getting Started

Tutorials

Concepts

Systems

API Reference

Errors

Benchmarks

Conformance

Overview

BrainChip AKD1500

Capability

What the AKD1500 supports

What the AKD1500 rejects

Weight quantization

Deployment: two-tier setup

Environment variables

Getting the Engine Library

Conformance

​Overview

​BrainChip AKD1500

​Capability

​What the AKD1500 supports

​What the AKD1500 rejects

​Weight quantization

​Deployment: two-tier setup

​Environment variables

​Getting the Engine Library

​Conformance

Overview

BrainChip AKD1500

Capability

What the AKD1500 supports

What the AKD1500 rejects

Weight quantization

Deployment: two-tier setup

Environment variables

Getting the Engine Library

Conformance