PaperlayerView exact revision on arXiv

Source-linked AI summary

ChroNoC: A Tableless TDM Circuit-Switched NoC With Delay-Aware Deployment for Time-Sensitive Neuromorphic Computing

Mohammad Rasoul Roshanshah, Mehdi Modarressi, Saeed Safari

arXiv:2610.34083v1cs.ARcs.NE

TL;DR

Time-sensitive neuromorphic systems need predictable communication, but joint placement, routing, and scheduling are complex, and immediate spike injection can ignore available axonal-delay slack. ChroNoC combines a tableless fixed broadcast fabric with DARSA, which uses that slack to schedule periodic injections. The evaluations report lower per-hop energy and reduced communication cycles and provisioned memory on the evaluated workloads.

  • Problem

    Joint placement, routing, and scheduling create a combinatorial search space, while existing deployment tools force immediate injection instead of exploiting axonal-delay slack.

    Source: §I · p. 1, §II · p. 2

  • Method

    ChroNoC uses synthesized fixed broadcast trees, and DARSA searches neuron placements while deriving periodic injection calendars from prescribed axonal delays.

    Source: §I · pp. 1–2

  • Results

    Periodic reservations cut communication frame length by 60.0% on UCI-HAR and 46.2% on N-MNIST versus DARSA immediate, while total provision drops from 94,368 to 48,396 bits and from 83,668 to 53,046 bits, respectively.

    Source: B. Comparison with State-of-the-Art… · p. 6

  • Takeaways & Limitations

    The evaluations report reduced per-hop energy, frame duration, and provisioned communication memory while preserving exact logical consumption timing.

    Source: §V · p. 6

Abstract

from arXiv · show

Time-sensitive neuromorphic architectures require predictable interconnects to preserve precise event timing. Statically scheduled time-division-multiplexed (TDM) circuit-switched networks-on-chip (NoCs) provide deterministic latency at the cost of linear configuration-memory scaling with schedule depth. Furthermore, jointly optimizing task placement, routing, and scheduling creates an intractable combinatorial search space. Existing deployment frameworks also ignore algorithmic axonal delays by forcing immediate spike injections that increase network contention. This paper proposes a hardware-software co-design to address these challenges. ChroNoC is a tableless TDM NoC that embeds precomputed full-grid broadcast trees into combinational logic as a fixed routing substrate. This fixed substrate reduces deployment to a spatial search via Delay-Aware Reservations and Spatial Annealing (DARSA). DARSA converts axonal delays into periodic scheduling slack to stagger network injections and reduce communication overhead. In a 130nm CMOS process, ChroNoC reduces per-hop energy consumption by more than 75% relative to the evaluated circuit- and packet-switched baselines. On UCI-HAR and N-MNIST workloads, DARSA reduces per-timestep communication cycles by approximately 60% and 45%, respectively, compared with immediate transmission, while also reducing provisioned communication memory. These results demonstrate how combining lightweight TDM switch design with delay-aware deployment improves both hardware efficiency and system-level communication performance while preserving neural event timing.

I. INTRODUCTION

ChroNoC combines a fixed, tableless broadcast substrate with DARSA to reduce deployment to capacity-constrained neuron placement. DARSA uses prescribed axonal delays as scheduling slack, while the fixed substrate excludes unrestricted joint routing and scheduling choices.

I. INTRODUCTION

  • Existing mappers prioritize spike volume, hop count, congestion, and locality but do not exploit prescribed delays as injection slack.Immediate injection can concentrate spikes into narrow windows, while early delivery requires destination buffering.Source: §I · p. 1
  • Jointly assigning logical tasks to cores, candidate routes, and TDM injection slots creates a combinatorial search space that the fixed substrate constrains to spatial placement.This restriction sacrifices unrestricted joint optimality; placement is addressed using simulated annealing.Source: §I · pp. 1–2
  • DARSA searches capacity-constrained neuron placements and derives periodic calendars from prescribed axonal delays, balancing frame duration and provisioned communication memory.The fixed substrate removes independent route and within-frame injection-slot choices from the deployment search.Source: §I · p. 2
  • ChroNoC synthesizes fixed, conflict-free broadcast trees into crossbar-control logic, eliminating programmable per-timeslot switch memories and supporting multicast with one injection per source event.Assigning a neuron to a core fixes its broadcast tree and within-frame injection slot.Source: §I · p. 2
  • Under single-spike execution assumptions, delay-aware periodic reservations stagger injections while preserving logical event-consumption times.The prescribed delays define permissible delivery intervals once trained and fixed for inference.Source: §I · pp. 1–2

II. BACKGROUND AND MOTIVATION

Distributed spiking networks need predictable spike delivery, but simultaneous events can create injection contention. The section motivates using fixed axonal delays as scheduling slack to spread transmissions while preserving logical consumption timing.

II. BACKGROUND AND MOTIVATION

  • Even sparse spiking activity can produce severe injection contention when neurons firing in the same timestep generate multicast traffic.Hardware multicast replicates an injected packet across network branches.Source: §II · p. 2
  • Single-spike coding bounds each neuron's events per input sample, while fixed axonal delays define permissible delivery windows despite unknown firing timesteps.The delay applies to all outgoing connections of a source neuron.Source: §II · p. 2
  • A spike fired at timestep gs has a logical consumption deadline set by its axonal delay, distinct from its physical network arrival.Early arrivals are held in destination buffers until their consumption timestep.Source: §II · p. 2
  • Existing tools inject every spike immediately, requiring worst-case bandwidth for simultaneous firing and buffering long-delay spikes across timesteps.The section instead identifies axonal delay as slack for spreading network injection.Source: §II · p. 2
  • Staggering injections across frames lowers peak per-frame injection capacity, while placement determines local reservation periods, frame duration, and buffer storage.Neurons' source counts and minimum axonal delays on each core couple spatial placement to temporal reservations.Source: §II · p. 2

III. ARCHITECTURE AND DEPLOYMENT

The deployment flow separates workload analysis, offline routing, and placement-specific calendar evaluation. ChroNoC uses a reusable mesh-wide broadcast substrate with counter-driven synthesized switch control.

III. ARCHITECTURE AND DEPLOYMENT

  • The flow extracts workload connectivity and source-specific delays, independently generates reusable broadcast routes and injection slots, then evaluates a deterministic NI calendar for each candidate placement.After placement and calendar configuration, NIs inject pending runtime events during reserved opportunities.Source: §III · pp. 2–3

A. Fixed Broadcast Substrate

  • ChroNoC connects cores through a 2D mesh with one full-grid broadcast tree per source, independent of neural connectivity; receiving cores filter datagrams for local targets.Offline routing fixes each core's injection timeslot and the switch configurations for reuse across placements.Source: A. Fixed Broadcast Substrate · p. 3
  • A synchronized timeslot counter drives synthesized combinational crossbar decoding, eliminating per-timeslot configuration memories and runtime header routing.The switch retains the counter, decoder logic, and crossbar wiring.Source: A. Fixed Broadcast Substrate · p. 3

B. Offline Space-Time Broadcast Routing

Offline space-time routing unfolds mesh links over a search horizon and reserves link-timeslot pairs to construct conflict-free broadcast trees. Randomized attempts are validated, and the smallest-depth schedule is retained for decoder synthesis.

B. Offline Space-Time Broadcast Routing

  • The offline router unfolds directed core-to-switch and inter-switch links across a finite logical-timeslot horizon to form an acyclic graph with time-indexed vertices.The stated horizon is H = 4N^2 + 4N timeslots.Source: B. Offline Space-Time Broadcast Routing · p. 3
  • The router retains only validated conflict-free schedules, with reservations on link-timeslot pairs preventing conflicts across repeated link traversals.Failed attempts restart from an empty reservation set.Source: B. Offline Space-Time Broadcast Routing · pp. 3–4
  • Each randomized routing attempt shuffles core order, samples injection slots, and grows each source tree by multi-source BFS while excluding reserved link-timeslots.Uniform edge costs favor shortest available branches, and previously reserved link-timeslots are excluded.Source: B. Offline Space-Time Broadcast Routing · p. 3
  • Among successful attempts, the schedule with the smallest depth T = 1 + maxv,u av,u is retained and synthesized into tableless crossbar decoders.The algorithm outputs broadcast trees and injection slots alongside the schedule depth.Source: B. Offline Space-Time Broadcast Routing · p. 4

C. Periodic Injection Reservations

Periodic reservations assign each transmitting source a recurring phase and injection subslot, staggering injections while preserving each event’s consumption deadline.

C. Periodic Injection Reservations

  • Each nonempty core assigns transmitting sources a recurring frame phase and injection subslot, with a shared reservation period Pv and maximum per-phase count cv.Source: C. Periodic Injection Reservations · p. 4
  • Ordering sources by index j distributes them across phases, and sources sharing a phase occupy consecutive injection subslots starting from zero.Source: C. Periodic Injection Reservations · p. 4
  • Waiting is at most Pv −1 frames, so delivery completes by hs regardless of firing time and preserves the deadline hs = gs + δax.The core period Pv governs transmission repetition without altering individual consumption deadlines.Source: C. Periodic Injection Reservations · p. 4

D. Communication Memory Accounting

Communication-memory accounting covers buffering for pending events and early arrivals, then sums source, destination, calendar, NI, and network provisions.

D. Communication Memory Accounting

  • Sources buffer pending events until injection, while destinations retain early arrivals for the logical waiting interval.Source: D. Communication Memory Accounting · p. 4
  • For unknown firing times, provisioning allocates one pending entry per transmitting source and one shared destination entry per source–target core pair before local synaptic expansion.Event entries include a source ID, timing tag, and valid bit.Source: D. Communication Memory Accounting · p. 4
  • Total provisioned communication memory sums source, destination, calendar, NI, and network memory, excluding neural states, weights, and trace-specific runtime queue peaks.Calendar entries and counters, NI datapath registers, and two ω-word banks per receiving channel are included.Source: D. Communication Memory Accounting · p. 4

E. Spatial Placement

With routing and substrate fixed, DARSA searches spatial neuron placements, rebuilding reservations for candidates and optimizing frame duration and provisioned memory.

E. Spatial Placement

  • DARSA maps logical neurons A onto physical cores V over the fixed substrate, subject to unique assignment and each core’s task capacity Cv.Source: E. Spatial Placement · pp. 4–5
  • Each placement deterministically defines its source sets and calendar without firing traces, and DARSA equally weights normalized frame duration and memory against the cyclic Round-Robin baseline.Source: E. Spatial Placement · p. 5
  • DARSA explores placements with single-neuron moves and two-neuron swaps, accepting non-increasing-cost candidates or uphill moves probabilistically under geometric cooling.The search runs for K attempts or a time budget Btime.Source: E. Spatial Placement · p. 5

IV. EXPERIMENTAL RESULTS

The evaluation examines the proposed architecture and deployment framework through hardware implementation and deployment experiments.

IV. EXPERIMENTAL RESULTS

  • The evaluation examines both the proposed architecture and deployment framework.Source: §IV · p. 5
  • Hardware implementation is one evaluation component for the proposed architecture and deployment framework.Source: §IV · p. 5
  • Deployment experiments are the other stated evaluation component for the proposed architecture and deployment framework.Source: §IV · p. 5

A. Hardware Implementation and Comparative Efficiency

ChroNoC reduces hop energy against evaluated circuit- and packet-switched baselines, while its complete-switch power changes little as schedule depth increases. Its fixed multicast substrate also supports in-network replication without programmable per-timeslot switching tables.

A. Hardware Implementation and Comparative Efficiency

  • The implemented prototype is a 4 × 4 mesh of 16 switches with 32-bit unidirectional datapaths and schedule depth T = 64, using SkyWater 130 nm CMOS.Physical implementation used the OpenLane RTL-to-GDSII flow.Source: A. Hardware Implementation and… · p. 5
  • 79.25% and 83.14% lower hop energy than a table-based TDM switch and four-stage XY wormhole router, respectively, at 1.8 V and 100 MHz.The comparison is shown across technology nodes in Fig. 4.Source: Figure 4 · p. 5, A. Hardware Implementation and… · p. 5
  • ChroNoC supports native in-network multicast without route fields or programmable per-timeslot tables, whereas Argo requires separate unicast transfers for multiple destinations.The separate transfers add traffic and energy for spike fan-out.Source: A. Hardware Implementation and… · p. 5
  • 3.76%: complete-switch power rises from 1.344 to 1.394 mW as schedule depth increases 16-fold from T = 16 to 256.The isolated route-decision unit rises 479.17%, from 5.15 to 29.82 µW, over the same range.Source: A. Hardware Implementation and… · p. 5

B. Comparison with State-of-the-Art Deployment Frameworks

On UCI-HAR and N-MNIST, periodic reservations reduce mean inference communication cycles relative to immediate reservations and also lower provisioned communication memory. The evaluation uses shared workload models and a common 4 × 4 substrate to compare DARSA with adapted deployment frameworks.

B. Comparison with State-of-the-Art Deployment Frameworks

  • UCI-HAR signals are encoded into 48 spike channels, while N-MNIST polarity events are aggregated into binary input frames at 1 ms resolution.Each workload uses a 128-neuron recurrent LIF reservoir with fixed random weights and delays; only the logistic-regression readout is trained.Source: B. Comparison with State-of-the-Art… · p. 6
  • Deployment methods use the same trained models, delays, and input samples, and are checked against reference execution for preserved spike trains, neural states, and predictions.Samples are processed sequentially, with delayed spikes consumed before the reservoir resets.Source: B. Comparison with State-of-the-Art… · p. 6
  • DARSA is compared with adapted SNEAP, SpiNeMap, NeuMap, Xiao, and Jin methods on a shared 4 × 4 substrate at T = 59.The counterparts use immediate reservations (Pv = 1); DARSA immediate also uses Pv = 1 to isolate periodic-reservation gains.Source: B. Comparison with State-of-the-Art… · p. 6
  • Mean inference communication cycles are evaluated over test examples, with each communication frame requiring Tω clock cycles excluding neuron computation and external I/O.The reported total includes idle frames and the delayed-event tail.Source: B. Comparison with State-of-the-Art… · p. 6
  • 60.0% and 46.2% fewer mean inference communication cycles than DARSA immediate on UCI-HAR and N-MNIST, respectively.Total provisioned storage falls from 94,368 to 48,396 bits on UCI-HAR and from 83,668 to 53,046 bits on N-MNIST.Source: B. Comparison with State-of-the-Art… · p. 6

V. CONCLUSION

ChroNoC and DARSA combine a tableless fixed routing substrate with delay-aware deployment for predictable communication in single-spike SNNs with prescribed axonal delays.

V. CONCLUSION

  • The co-design replaces per-timeslot switch memories with synthesized fixed-tree decoding and uses axonal delays to stagger periodic injections without advance firing-time knowledge.The evaluations report reductions in per-hop energy, frame duration, and provisioned communication memory while preserving exact logical consumption timing.Source: §V · p. 6
Loading 2610.34083v1…