Install

A layered installation route for Linux hosts, with the AKD1500 memory and kernel caveats visible instead of buried.

0. Capture the host before changing it

Record the environment first. Debugging is dramatically easier when the kernel and OS are not a mystery.

uname -a
cat /etc/os-release
python3 --version
lspci -nn

1. Build the PCIe driver

The public BrainChip driver repository requires build tools and matching kernel headers. Its install script builds, loads and configures the driver to load at boot.

sudo apt update
sudo apt install build-essential linux-headers-$(uname -r)

git clone https://github.com/Brainchip-Inc/akida_dw_edma
cd akida_dw_edma
sudo ./install.sh

2. Know the CMA constraint

The driver repository notes that AKD1500 over PCIe can require CMA (Contiguous Memory Allocator) support for larger models or fuller pipeline usage. Do not rebuild a kernel just because a guide says so: first establish whether your current kernel has the required support and whether your workload actually needs it.

grep -E 'CONFIG_CMA|CONFIG_DMA_CMA' /boot/config-$(uname -r) 2>/dev/null || true
grep -i cma /proc/meminfo /proc/cmdline 2>/dev/null || true

3. Use a clean Python environment

The current public MetaTF installation page reports version 2.19.3. We keep the site date-stamped and will pin the exact version only after reproducing the Pi build.

python3 -m venv ~/venvs/akida
source ~/venvs/akida/bin/activate
python -m pip install --upgrade pip
pip install metatf==2.19.3

4. Verify the package before hardware

python -c "import akida; print(akida.__version__); print(akida.devices())"

On a machine without Akida hardware, the official installation page explicitly shows an empty device list as valid. On the Pi with a working device/driver stack, we expect actual hardware entries instead.