Build on Akida.
Understand the edges.
Practical installation guides, reproducible demos and engineering notes for AKD1500 on Raspberry Pi — with vendor facts, our own measurements and experiments kept deliberately separate.
A setup path that exposes failure points.
We do not collapse PCIe, kernel, driver, Python and model mapping into one magic command. Each layer gets its own verification step.
Topology
Know what is hardware, adapter, bus and software before installing anything.
02Driver
Build the official PCIe driver and account for kernel/CMA constraints.
03Runtime
Create a clean Python environment and install the current Akida toolchain.
04Verify
Check module, device nodes and akida.devices().
05Map
Load a .fbz model and map it to a physical or virtual AKD1500 target.
AKD1500 is an Akida 1.0 target.
The current MetaTF hardware guide explicitly groups AKD1000 and AKD1500 under Akida 1.0 capabilities. That matters when choosing examples, quantization paths and mapping expectations.
from akida import devices hw = devices() print(f"devices: {len(hw)}") for idx, dev in enumerate(hw): print(idx, dev.desc, dev.version) # keep vendor facts and lab facts separate # then record kernel + driver + topology
Where the hardware earns its keep.
The site is not only an installation manual. It documents why we are building with the hardware in the first place.
VORSA-M3
Neuromorphic vision for free classes, variable shapes and overlapping objects in a production workcell.
OPEN PROJECT →EXPERIMENT4× AKD1500 on one Pi
Device enumeration, memory pressure, topology and model assignment. Results are published only when measured.
OPEN LAB NOTE →REPRODUCIBLESmall demos
Hardware discovery first, then mapping, inference and finally application-level experiments.
VIEW DEMOS →Failures belong in the documentation.
Short notes become durable engineering knowledge instead of disappearing in chat logs and terminal history.
Four AKD1500 devices on one Raspberry Pi
What is confirmed by vendor material, what is only shown on other hosts, and what we still have to measure ourselves.
CMA, kernel updates and why the driver can break later
The official driver notes that AKD1500 can require CMA for larger models and needs rebuilding after kernel updates.
Neuromorphic edge AI without the marketing layer
A practical engineering view of what we actually need to measure and validate.
MetaTF 2.19.3 checked against current public documentation.
We show the date because installation instructions and model tooling move. Old screenshots are not a versioning strategy.