Hardware & topology
Understand which parts are vendor capabilities and which parts are our lab architecture.
AKD1500 and Akida 1.0
Current MetaTF documentation identifies AKD1500 as an Akida 1.0 reference SoC. The runtime exposes akida.AKD1500() as a virtual device so mapping can be checked without physical hardware.
Raspberry Pi is our host target
Our project uses Raspberry Pi as the orchestration host. That is an engineering choice for this lab, not a claim that every Raspberry Pi + adapter + kernel combination is an official reference configuration. We therefore record the exact Pi model, PCIe path, adapter/switch, OS image and kernel for every reproduced result.
Target topology
| Layer | Target | Status |
|---|---|---|
| Host | Raspberry Pi | Our lab target |
| Accelerator | 4 × AKD1500 M.2 B+M | Our lab target |
| Interface | PCIe via suitable adapter/switch | Must be documented per build |
| Vision input | 2 × Raspberry Pi Camera Module 3 | VORSA-M3 target |
| Workloads | Detection + segmentation experiments | Experimental |
What multi-device evidence actually proves
BrainChip's public symphony-akida repository demonstrates multiple AKD1500 devices as independent PCIe endpoints on a documented x86 reference host. That is useful evidence for the software/device model, but it does not establish that our Raspberry Pi topology will behave identically. We measure that separately.