Platform · Runtime

Edge Runtime

The event-driven execution engine that loads and runs .snn binaries on neuromorphic silicon. Written in C99, zero OS dependency, static memory layout — designed for bare-metal deployment where every microwatt matters.

Runtime State

Spike queue depth4 / 512
Active neurons12 / 4096
Timestep window20 µs
Power stateACTIVE
Heap used0 B (static)

Core Design

Runtime architecture

Event-driven dispatch

The runtime is dormant between spike events. When a spike arrives on the HAL input bus, the dispatcher wakes the affected neuron groups, runs the timestep window, and returns to sleep. Power draw during dormancy: <5 µW.

Static memory layout

All neuron state buffers, weight arrays, and spike queues are allocated at load time from a contiguous memory arena. No malloc after init. Memory footprint is deterministic and declared in the .snn header.

HAL shim interface

The runtime depends only on the HAL shim API: 8 function pointers covering spike I/O, memory access, timer, and power mode. Port to a new chip by implementing these 8 functions — no runtime code changes required.

Fault tolerance

Configurable watchdog hooks: spike-timeout detection, neuron saturation recovery, and checkpointed neuron state saves. Graceful degradation to reduced-accuracy mode on partial silicon failure.

Integration

Loading a .snn binary

No OS dependency
Bare-metal C99. Works in superloop, RTOS task, or interrupt context.
32 KB minimum footprint
Runtime + HAL shim fits in 32 KB SRAM on most targets. Model arena is additive.
Deterministic latency
No dynamic allocation after init. nrm_step() worst-case latency is bounded and measurable.

Deploy your first model

The Quickstart walks through compile → deploy → verify in one session.