Self-contained HTML training material covering BrainChip's Akida 2 neuromorphic IP — from core architecture through quantization, model conversion, device mapping, on-hardware inference and performance measurement, and a reference on supported layers and operations.
Each module is a single, self-contained HTML file (styles, scripts, and images
all inlined) — no build step or server required. Open
Akida/akida_module1.html in a browser and
navigate forward using the sidebar or the bottom-of-page controls.
| File | Module | Covers |
|---|---|---|
akida_module1.html |
1 — What is Akida 2? | Event-based processing, sparsity, at-memory compute, product family |
akida_module2.html |
2 — Hardware Architecture | Node mesh, NPEs, component types (HRC/CNP/FNP/TNP), memory hierarchy, DMA |
akida_module3.html |
3 — Quantization | Scale/zero-point mechanics, calibration, PTQ vs QAT, bitwidth selection |
akida_module4.html |
4 — MetaTF Flow | Keras → quantizeml → cnn2snn → Akida pipeline |
akida_module5.html |
5 — MetaONNX Flow | ONNX → onnx2akida.convert() → HybridModel pipeline |
akida_module6.html |
6 — Devices & Mapping | Virtual/real devices, create_device(), MapMode, model.summary() |
akida_module7.html |
7 — Inference & Performance | forward()/predict(), FPS/power statistics, CLI, Akida Cloud, optimization |
akida_module8.html |
8 — Supported Layers & Operations | Layer/activation/bitwidth reference, structural constraints, HybridModel CPU fallback |
Supplementary / not part of the main sequence:
qat_explainer.html— standalone deep-dive on how QAT works, not linked from the module sidebar/nav.Tenns/— reserved for upcoming TENNs-focused modules (not yet added).
Modules 1–7 have been through a technical accuracy pass reconciling course
content against BrainChip SME feedback and internal consistency checks
(hardware capability claims, quantization scheme details, dangling
cross-references, etc.). A handful of points are intentionally worded as
pending rather than asserted as fact, and are called out inline with
"— confirming with the HW/SW team" (or similar) where the exact behavior
wasn't certain at review time — e.g. the precise CNP1 vs CNP2 selection
criteria, and the exact effect of onnx2akida.convert()'s enable_hwpr
argument. Search the modules for "HW/SW team" to find all such flags.
Module 8 is a new consolidated reference built entirely from facts already stated in Modules 2–6 — it hasn't been through a separate SME pass and carries forward the same pending flags (e.g. the CNP1/CNP2 boundary) rather than resolving them. Treat it as reviewed to the same extent as the modules it draws from, not as independently re-verified.
These pages carry Marketing-approved formatting and branding — edits should be content-only: fix or update text, tables, and inline diagram data without changing HTML structure, CSS, or layout. When a technical claim can't be confirmed with confidence, soften the wording and flag it inline rather than asserting or deleting it outright.