Eventra · Neuromorphic systems agency
NEXUS-12 is a memory-light inference appliance for always-on vision, audio, and sensor banks. Eventra maps practical edge models across 12 independent Akida processors—without pretending this is a training server or an LLM box.
12×
event processors
96 MB
partitioned SRAM
60–120 W
loaded estimate
What to expect
Every pilot receives per-chip throughput, mJ per inference, system power, and model-fit results.
We document unsupported layers, quantization effects, PCIe lane constraints, host bottlenecks, and thermal conditions.
The engagement ends with a compiled model, deployment evidence, and a one-chip or four-chip reference BOM where appropriate.
Where NEXUS-12 fits
The system is strongest when twelve independent streams can each own a compact quantized model.
01 / Multi-stream inference
Route visual wake, keyword spotting, vibration anomaly, or sensor classification independently. The host preprocesses; each AKD1000 runs its resident event model; scores return over PCIe.
02 / Model mapping
Eventra handles quantization-aware preparation, MetaTF compilation, layer compatibility, accuracy checks, and the delivered .fbz.
03 / Honest fit
Good for spikes, embeddings, and occasional frames. Poor for raw HD video fanned to all 12 chips, large-model training, LLM serving, or HBM-bound workloads.
Industrial lines
Visual wake, PPE events, line counting, acoustic and vibration monitoring.
Tactical edge
SWaP-constrained classification and local operation where connectivity is limited.
Smart infrastructure
Person, vehicle, and keyword event banks that wake heavier compute only when needed.
OEM evaluation
MetaTF bring-up, model-zoo regression, and a path toward AKD1500-based production designs.
Start with your workload
Bring one model family and a representative stream. Eventra will assess fit, define the benchmark, and scope a paid pilot with explicit success and stop criteria.