Research analysis · Hardware

ChroNoC: a tableless network-on-chip that spends axonal delays as scheduling slack

Deterministic interconnects for neuromorphic hardware have always paid for predictability in configuration memory and intractable deployment search. ChroNoC's answer is to delete the tables entirely - hard-wire the broadcast trees - and to treat the axonal delays that schedulers usually ignore as free slack that staggers traffic.

Source: ChroNoC: A Tableless TDM Circuit-Switched NoC With Delay-Aware Deployment for Time-Sensitive Neuromorphic Computing, Roshanshah, Modarressi, and Safari, arXiv:2610.34083, 2026-10-25. Primary source. Read the full arXiv HTML including implementation tables and the deployment comparison.

What the work claims

This is a hardware-software co-design paper from the University of Tehran, evaluated in silicon-grade synthesis rather than simulation alone1. The problem it attacks: time-sensitive neuromorphic architectures - SNNs whose axonal delays are part of the algorithm, not noise to be hidden - need interconnects with deterministic latency. Statically scheduled time-division-multiplexed (TDM) circuit-switched networks-on-chip provide that determinism, but configuration memory grows linearly with schedule depth, jointly optimizing placement, routing, and scheduling is a combinatorial swamp, and existing deployment frameworks force every spike to inject immediately, ignoring the algorithmic delay that would have absorbed it anyway.

ChroNoC's claims: a switch design whose configuration memory cost disappears, and a deployment method whose search collapses to a spatial problem. In a 130 nm CMOS implementation the switch cuts per-hop energy by 79.25 percent against a table-based TDM switch and 83.14 percent against a four-stage XY wormhole router. On two benchmarks, the delay-aware scheduler cuts per-timestep communication cycles by 60.0 percent (UCI-HAR) and 46.2 percent (N-MNIST) relative to immediate injection, while also shrinking provisioned communication memory, and it preserves exact spike timing against reference execution.

How it works

The hardware move is subtraction. Each switch in the 2D mesh precomputes, offline, one full-grid broadcast tree per source core, and burns the whole set into combinational decoding logic driven by a synchronized timeslot counter. There are no per-timeslot configuration memories, no routing headers, no runtime arbitration, and no deep buffers: a transmitted spike word is a bare datagram that replicates along its fixed tree branches, and receiving cores filter for their local targets. Everything the schedule needs to know is encoded once in wiring.

The routing layer constructs those trees on a time-expanded graph - links unfolded over a search horizon of H = 4N2 + 4N timeslots - with randomized attempts until a conflict-free schedule validates. The deployment layer, DARSA (Delay-Aware Reservations and Spatial Annealing), is where the biological metaphor pays. Each transmitting source has an algorithmic axonal delay, the number of timesteps its spike is supposed to travel. Existing frameworks inject immediately and hold the spike in buffers, burning contention on traffic the algorithm would have tolerated later. DARSA instead gives each source a recurring frame phase derived from the smallest axonal delay on its core, staggering injections across phases so that delayed traffic arrives exactly when needed, then anneals task placement spatially on the fixed routing substrate. The delay that used to be buffer occupancy becomes scheduling slack.

The numbers are internally consistent and the comparisons are fair as far as they go. The implementation is a 4×4 mesh of 16 switches, 32-bit links at 100 MHz, synthesized through OpenLane on the open SkyWater 130 nm PDK: 1.6 × 1.6 mm die, 2.5066 mm2 core area, 26.11 mW total power. Scaling schedule depth from 16 to 256 multiplies the isolated route-decision logic power by 5.8 (5.15 to 29.82 µW) but the complete switch by only 3.76 percent (1.344 to 1.394 mW), because sequential cells carry 81 to 83 percent of switch power - the tableless design converts what would have been memory scaling into a minor combinational term. On the workloads - 128-neuron LIF reservoirs with fixed random delays of 1 to 20 timesteps, single-spike latched execution, logistic-regression readouts reaching 75.94 percent (UCI-HAR) and 86.55 percent (N-MNIST) test accuracy - periodic reservations cut the expansion factor from 15 to 6 and from 13 to 7, shrinking the communication frame from 885 to 354 cycles and from 767 to 413 cycles, and provisioned storage from 94,368 to 48,396 bits and from 83,668 to 53,046 bits. Compilation costs about 8 seconds, offline.

Where a skeptic should push

The load-bearing assumption is that axonal delays are known, static, and yours to schedule. Every gain DARSA reports comes from spending delay as slack, which presumes the delay is a design parameter handed to the compiler. That is true for emulated SNNs with assigned delays. It is not true for systems where the delay-bearing medium is physical and adaptive - including, pointedly, the hybrid ones our field cares about. The paper is upfront that this is a co-design for "single-spike SNNs with prescribed axonal delays"; the skeptic's note is simply that "prescribed" is doing all the work in the title.

Second, broadcast-everything has a hidden tax. Every datagram traverses its full-grid tree and every core filters, so link energy is spent on copies most receivers discard. The 79 to 83 percent per-hop energy figures are per hop, not per usefully delivered spike; a workload with localized connectivity could look very different on delivered-spike energy, and the benchmarks here - 128 neurons, 795 and 778 recurrent connections, at most 128 spikes per sample - are small enough that the broadcast substrate never sweats. The 4×4 mesh is also the only scale with end-to-end results; 16×16 schedule scaling is shown for switch power synthesis, not for routing feasibility, and the H = 4N2 + 4N horizon grows quadratically.

Third, the baseline set is deployment frameworks adapted to the same substrate (SNEAP, SpiNeMap, NeuMap, and two others), all forced to P = 1 immediate reservations. That isolates the periodic-reservation gain honestly, but the packet-switched energy comparison leans on published routers at other technology nodes normalized onto one plot; treat the 83 percent as an order-of-magnitude statement, not a measurement.

The timing lesson for hybrid wetware systems

A network-on-chip paper has no obvious business on an organoid-computing site, but this one smuggles in the exact question hybrid systems live or die on: whose clock does the interconnect obey?

The non-obvious implication is that delay-aware scheduling is a two-way street. ChroNoC proves, in silicon, that known axonal delays are worth 46 to 60 percent of communication cycles - that timing slack is a schedulable resource, not idle time. In a hybrid MEA-to-neuromorphic pipeline, the medium has delays too: axonal conduction in the culture, latency in the digitization chain, jitter in the stimulation loop. Today's default is ChroNoC's P = 1 regime - sample immediately, timestamp, absorb the biological delay in buffers and decoder training. This paper's result is a standing invitation to invert that: if you can measure the biological delays online, and they are stable enough over an experiment's duration to be predictable, you can schedule acquisition and stimulation against them the way DARSA schedules spike injection, recovering throughput without touching the tissue. The deterministic-interconnect discipline - counter-driven slots, no arbitration, exact preservation of event timing - is also precisely what a decoding frontend needs when spike timing is the signal, which it always is in electrophysiology.

The genuine threat is the mirror image, and it is severe. Living conduction delays are not prescribed: they drift with development, activity, and electrode interface quality, and no compiler gets to anneal them. A static TDM schedule compiled against biological delays is stale by the time the culture finishes its first night. Worse, the paper's own efficiency mechanism becomes a fragility - deterministic schedules have no arbitration to fall back on when reality deviates from the compiled schedule, and a hybrid system that trusts its timing model too literally will drop or mis-time events rather than queue them. The lesson is not "adopt ChroNoC for organoid rigs"; it is "adopt its accounting" - measure your delays, quantify their stability over the experiment, and only then decide how much determinism you can afford to promise the rest of the stack.

There is also an obsolescence angle worth stating plainly. Tableless determinism buys its efficiency from workload regularity: recurring frames, periodic phases, single-spike latching. Closed-loop biological computing is the opposite workload - nonstationary, adaptive, bursty in ways that track the task, not the calendar. If the long-term substrate is living, the winning interconnect will look less like a compiled TDM schedule and more like the statistically scheduled, drift-tolerant thing a neural decoder already is. ChroNoC is a superb design for the deterministic half of the field's future, and a useful foil for the other half.

The bottom line

Established: in synthesis at 130 nm on a 4×4 mesh, eliminating per-timeslot routing tables cuts per-hop energy by roughly 80 percent against a table-based TDM switch, schedule-depth scaling is nearly flat in total switch power, and delay-aware periodic reservations cut communication cycles by 46 to 60 percent and provisioned memory by 35 to 49 percent on two small benchmarks while preserving exact spike timing. Not established: behavior beyond 4×4 with end-to-end schedules, energy per delivered spike under localized traffic, or any regime where delays are not prescribed constants. What would extend the claim: a deployment at 16×16 or larger with feasible offline schedules; what would limit it: a workload whose delay distribution is nonstationary, which is precisely the regime of the biological substrates this stream covers.

Frequently asked questions

What makes ChroNoC tableless?

Instead of storing per-timeslot routing configurations in memory, the switch synthesizes its crossbar connections from combinational logic driven by a timeslot counter, with precomputed broadcast trees fixed at design time.

What is DARSA?

The deployment method: Delay-Aware Reservations and Spatial Annealing. It assigns each source a recurring injection phase based on its axonal delay, then anneals task placement onto the fixed routing substrate.

How much energy does it save?

Post-layout at 130 nm, per-hop energy drops 79.25 percent versus a table-based TDM switch and 83.14 percent versus a four-stage XY wormhole router, at 1.8 V and 100 MHz.

Why do axonal delays help scheduling?

A spike with a 10-timestep algorithmic delay need not be injected now and buffered; injecting it later, at its reserved phase, frees earlier slots for other traffic, cutting frame length by roughly half.

What are the main limitations?

Only a 4×4 mesh is shown end to end, benchmarks are small 128-neuron reservoirs, per-hop energy ignores the cost of broadcast copies, and the method requires axonal delays to be known and static at compile time.

Why should organoid computing care?

It quantifies how much predictable timing is worth, and shows both how to exploit measured delays as scheduling slack and how dangerous static schedules are when delays drift, as they do in living tissue.

References

  1. Roshanshah, M. R., Modarressi, M., and Safari, S. ChroNoC: A Tableless TDM Circuit-Switched NoC With Delay-Aware Deployment for Time-Sensitive Neuromorphic Computing. arXiv:2610.34083. 2026. https://arxiv.org/abs/2610.34083. Accessed 2026-11-12.