NIFTY routing with SNNs

Most of the routing strategies explored elsewhere on this site — greedy shortest-path, congestion avoidance, token passing — are hand-written, deterministic algorithms. NIFTY is also investigating a more ambitious alternative: using a Spiking Neural Network (SNN) as the decision-making core of the router itself, so that each node’s routing choices come from a small trained model rather than a fixed set of rules. This is meant to make individual chips genuinely intelligent — able to recognize and adapt to traffic and failure patterns that a linear algorithm wasn’t explicitly designed to handle — without the power cost that normally comes with running machine learning inference on-chip.

The key is the SNN computing model itself. Unlike conventional artificial neural networks, spiking neurons communicate with sparse, asynchronous pulses rather than continuously-valued activations, and only consume energy when a spike actually occurs. That event-driven behavior is a natural fit for a router: routing decisions are only needed when a packet arrives, and network traffic in detector readout systems is typically sparse and bursty rather than constant. An SNN-based router can sit idle at near-zero power between events and “fire” only to make a decision, giving it a plausible path to running real machine learning inference within the tight power and area budget of a chip’s routing logic — something a conventional neural network running continuously could not do.

This work is complementary to the adaptive routing goals described on the LArPix and 4D Pixel pages: those focus on the network architecture and buffering that adaptive routing needs, while the SNN effort asks how intelligent the per-node decision itself can be made without paying for it in power.