E0402: synaptic delays cannot be mapped to akida-akd1500
What happened
A .thx artifact was validated against the akida-akd1500 backend and one of its layers contains a synaptic delay encoding. The AKD1500 capability descriptor declares native_delay_max_steps = 0 and delay_fallback = Reject, meaning delays are neither supported natively nor emulated.
Why
AKD1500 is an Akida 1.0 device. Temporal Neural Processors (TNP) that provide buffered or recurrent delay processing are an Akida 2.0 feature not present on AKD1500. The hardware executes a single feedforward pass per inference call with no per-synapse timing buffers.
There is no mechanism to emulate per-synapse delays in a single-frame feedforward model. Silently dropping delays would produce incorrect output, so the backend rejects the artifact instead.
How to fix
Option A — Use the simulator. The sim backend supports delays up to 65535 timesteps natively (native_delay_max_steps = 65535, delay_fallback = Emulate):
Option B — Retrain without delays. Remove the delay encoding from the model before compiling for akida-akd1500. This changes the model’s temporal behaviour and will require retraining:
Example
