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

Mohammad Rasoul Roshanshah, Mehdi Modarressi, Saeed SafariarXiv 2026 · arXiv

This paper presents ChroNoC, a hardware-software co-design for Spiking Neural Network (SNN) accelerators that tackles the "memory wall" and congestion issues in predictable on-chip communication. By replacing bulky configuration tables with synthesized logic and treating biological delays as scheduling "slack," the authors achieve significant gains in energy efficiency and communication speed.

TL;DR

ChroNoC eliminates the massive memory tables usually required for predictable TDM networks by synthesizing routing logic directly into hardware. Paired with a mapping framework (DARSA) that uses axonal delays to stagger network traffic, it reduces per-hop energy by over 75% and communication cycles by up to 60% compared to traditional baselines.

What Problem Does This Paper Address?

Neuromorphic chips need to move "spikes" (data packets) between artificial neurons with extreme precision. Traditional packet-switched Networks-on-Chip (NoCs) are flexible but suffer from unpredictable delays due to traffic jams (contention). Time-Division Multiplexed (TDM) circuit switching offers a solution by giving every connection a reserved "timeslot," ensuring data arrives exactly when expected.

However, TDM has two major drawbacks:

  1. Memory Overhead: To know which "switch" to flip at which "time," every router needs a large memory table. This table grows linearly with the length of the schedule, consuming precious chip area and power.
  2. Mapping Complexity: Deciding where to place neurons (placement), how to connect them (routing), and when they should talk (scheduling) is a mathematically "intractable" problem. Most tools simplify this by forcing spikes to be sent immediately, which creates massive traffic bursts.

Methods

The authors propose a two-part solution:

1. ChroNoC (The Hardware)

Instead of storing a schedule in a memory table, ChroNoC uses tableless TDM. The researchers precompute a set of "all-to-all" broadcast trees and synthesize the control logic for these trees directly into combinational logic.

  • Fixed Substrate: Every core has a dedicated, hard-wired timeslot to broadcast its spikes to the entire grid.
  • Synthesized Decoding: Because the paths are fixed, the "switches" are controlled by simple decoders rather than memory lookups, drastically reducing power.

2. DARSA (The Software)

To deploy a neural network onto this hardware, the authors developed Delay-Aware Reservations and Spatial Annealing (DARSA).

  • Exploiting Slack: In biology, an "axonal delay" is the time it takes a spike to travel down a nerve. DARSA treats this as a "delivery window." If a neuron has a 10ms delay, the system doesn't have to send the spike immediately; it can stagger the injection anytime within that window to avoid network congestion.
  • Simplified Search: Because the routing is already fixed in the hardware, DARSA only needs to solve the spatial placement of neurons, turning a three-dimensional nightmare (placement + routing + scheduling) into a one-dimensional optimization problem.
Fig. 1. Workload preparation and deployment on ChroNoC using workload-independent broadcast routing and delay-aware periodic reservations.

Datasets

The framework was evaluated using two time-sensitive neuromorphic benchmarks:

  • UCI-HAR: Human Activity Recognition data from smartphone inertial sensors (accelerometers/gyroscopes).
  • N-MNIST: A neuromorphic (event-based) version of the classic MNIST handwritten digit dataset, captured using a DVS camera. Both workloads were mapped to a recurrent "Liquid State Machine" (LSM) reservoir to test complex communication patterns.

Key Insights and Ideas

  • Logic over Memory: The "aha moment" in the hardware design is that for fixed schedules, combinational logic is far more efficient than memory. As shown in Figure 5, increasing the schedule depth (T) has a negligible effect on the total switch power because the design is dominated by registers, not the synthesized decoders.
  • Delay as a Resource: Most mappers view axonal delay as a constraint to be met after delivery. DARSA flips this, viewing delay as a resource that provides "scheduling slack" to smooth out traffic bursts.
  • Native Multicast: By using broadcast trees, the hardware naturally supports one-to-many communication, which is the bread and butter of neural connectivity, without needing complex header-parsing or multiple unicast transmissions.
Fig. 4. Energy per hop across technology nodes (logarithmic axis).

Key Results

  • Energy Efficiency: In a 130nm CMOS process, ChroNoC reduced per-hop energy consumption by 79.25% compared to table-based TDM and 83.14% compared to standard XY wormhole routing.
  • Speedup: On the UCI-HAR workload, DARSA reduced the communication cycles per timestep by 60%. For N-MNIST, the reduction was 45%.
  • Memory Savings: By staggering spikes, the system requires fewer buffers to handle "bursty" traffic, leading to a nearly 50% reduction in provisioned communication memory for the UCI-HAR task.

Limitations and Future Work

  • Broadcast Efficiency: ChroNoC uses full-grid broadcast trees. While this simplifies hardware, it might be overkill for neural networks with very sparse or localized connectivity, where a lot of bandwidth might go to waste.
  • Single-Spike Coding: The current evaluation assumes "single-spike" coding (where each neuron fires at most once per sample). While this is common in low-power edge AI, the authors note that future work could extend this to more complex rate-coding schemes.
  • Fixed Topology: The current study focuses on a 4x4 mesh. Scaling this to much larger many-core systems (e.g., 64x64) would require investigating how the synthesized logic complexity grows at larger scales.

Summary updated

Abstract

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.
BibTeX
@misc{roshanshah2026chronoc,
  author = {Mohammad Rasoul Roshanshah and Mehdi Modarressi and Saeed Safari},
  title = {{ChroNoC: A Tableless TDM Circuit-Switched NoC With Delay-Aware Deployment for Time-Sensitive Neuromorphic Computing}},
  year = {2026},
  eprint = {2610.34083},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34083},
}
Figure