The service started as a Symphony-managed BrainChip Akida inference component serving a single team. Today the same service operates as a multi-tenant neuromorphic inference fabric supporting twelve domains with room for more, multiple sensor modalities, a hundred-plus model files, three backend types, and as of this week, two distinct model architectures sharing the same dispatcher on each Akida chip in the fabric. No new services, no new containers, no new Symphony deployment cycle. The property worth highlighting is not any particular model’s accuracy but the structural capacity of the service to absorb each of those additions as configuration rather than as a new operational surface.
Four axes carry the capacity.
Multi-modal at the input layer
Each tenant declares a source_id_allowlist that gates inputs by sensor type, so visual, acoustic, RF, BLE, satellite, and software-defined event streams all share the same chip-side service without coupling to each other.
Multi-domain at the policy layer
Twelve teams hold their own observation rings, GPFS spike archives, crypto signers, and threat-gating thresholds, all deny-by-default so that an unknown app tag returns 403.
Multi-model at the runtime layer
The registry indexes more than a hundred .fbz files on each node, hot-swappable per request, with backend routing per call across real AKD1000 silicon, software v1 simulation, or v2 simulation. The same service simultaneously holds different models on different backends without restart. Models swap in and out wholesale or one at a time to enable cross-model multi-domain setups running concurrently. The distillery example showed the cadence flexibility clearly, one domain and its model running for two minutes, another for an hour, with swaps timed to whatever the workload demands.
Multi-model-type at the dispatcher layer
The newest axis was opened by the eventcam vehicle classifier. The original single-frame /predict endpoint handles the CNN tenants, everything from vehicle classification to RF mel-spectrum encoding. The new /predict_sequence endpoint handles buffered TENNs Pleiades, with per-sequence state reset and majority-vote aggregation across the temporal window. The two coexist. The next architecture requiring different inference semantics would land as one more dispatch action rather than as a new service.
The payoff
The consequence is that team velocity decouples from deployment cadence. The visual team, the RF team, the event-camera team, and the trust pipeline ship independently without coordinating Symphony cycles or chip-side restarts. The platform absorbs new work as configuration rather than as a new operational surface. Such absorption is the version of “generic” in AkidaGenericService that actually repays the original engineering investment in writing one multi-tenant service in place of N separate ones.
Additional chips on additional nodes extend the fabric horizontally under Symphony, with every axis of multiplicity available to any chip or simulation in the cluster. The result, truly, is a neuromorphic hive mind.