PaperlayerView exact revision on arXiv

Source-linked AI summary

Mapping Spiking Neural Networks to Neuromorphic Hardware

Adarsha Balaji, Anup Das, Yuefeng Wu, Khanh Huynh, Francesco Dell'Anna, Giacomo Indiveri, Jeffrey L. Krichmar, Nikil Dutt, Siebren Schaafsma, Francky Catthoor

arXiv:1909.01843v1cs.ETcs.LGcs.OS

TL;DR

Large SNNs mapped across crossbars can encounter shared-interconnect congestion and latency, while placement choices also affect energy and latency. SpiNeMap clusters synapses and places them on hardware, reducing energy consumption on the shared interconnect by 45% and spike latency by 21% compared to state-of-the-art techniques. The evaluation also reports lower ISI distortion and improved application accuracy.

  • Problem

    Shared-interconnect congestion and latency can distort ISIs, and placement strategies affect energy consumption and latency in large SNN mappings.

    Source: §I · p. 1

  • Method

    SpiNeMap partitions SNN synapses into local and global groups, then places clusters on physical crossbars using particle swarm optimization.

    Source: §VII · p. 10

  • Results

    Energy consumption on the shared interconnect falls by 45% and spike latency by 21%, while ISI distortion falls by 36% and application accuracy improves by 12% over state-of-the-art approaches.

    Source: §VII · p. 10

  • Takeaways & Limitations

    The authors demonstrate SpiNeMap on DynapSE, and report that its optimization strategies improve application accuracy.

    Source: §VII · p. 10, 2) Using SpiNeMap for other… · p. 11

Abstract

from arXiv · show

Neuromorphic hardware platforms implement biological neurons and synapses to execute spiking neural networks (SNNs) in an energy-efficient manner. We present SpiNeMap, a design methodology to map SNNs to crossbar-based neuromorphic hardware, minimizing spike latency and energy consumption. SpiNeMap operates in two steps: SpiNeCluster and SpiNePlacer. SpiNeCluster is a heuristic-based clustering technique to partition SNNs into clusters of synapses, where intracluster local synapses are mapped within crossbars of the hardware and inter-cluster global synapses are mapped to the shared interconnect. SpiNeCluster minimizes the number of spikes on global synapses, which reduces spike congestion on the shared interconnect, improving application performance. SpiNePlacer then finds the best placement of local and global synapses on the hardware using a meta-heuristic-based approach to minimize energy consumption and spike latency. We evaluate SpiNeMap using synthetic and realistic SNNs on the DynapSE neuromorphic hardware. We show that SpiNeMap reduces average energy consumption by 45% and average spike latency by 21%, compared to state-of-the-art techniques.

I. INTRODUCTION

SpiNeMap addresses the energy and latency costs of mapping large SNNs across crossbars and a shared interconnect. Its partitioning and placement steps target global-spike congestion and hardware placement, while the background explains how interconnect latency can distort spike timing.

I. INTRODUCTION

  • For large SNNs, shared-interconnect congestion adds spike latency and can distort inter-spike intervals, affecting application performance.The introduction identifies SNNs with more than a million synapses as a scale at which a shared bus becomes a latency and energy bottleneck.Source: §I · p. 1
  • SpiNeMap combines heuristic clustering to reduce global spikes with meta-heuristic placement of local and global synapses to reduce energy consumption and spike latency.Local synapses map within crossbars; global synapses use the shared interconnect.Source: §I · p. 1
  • SpiNeMap reduces energy consumption by 45% and spike latency by 21% compared to state-of-the-art techniques.The evaluation uses SNN-based applications on DynapSE.Source: §I · p. 1

II. BACKGROUND

  • The background describes supervised SNN inference, with performance measured using accuracy assessed through inter-spike intervals (ISIs).For feedforward networks, accuracy is assessed from ISIs in the output decision layer; for other architectures, critical-neuron ISIs contribute.Source: §II · p. 2
  • CARLsim supplies application-level spike data, but hardware latency must be incorporated because synapses mapped locally and globally can have different latencies.The described sources are fixed propagation latency within a crossbar and variable time-multiplexing latency on the shared interconnect.Source: §II · p. 2
  • Local synapses have uniform spike latency and zero ISI distortion, whereas varying congestion and routing paths make global synapses contribute to distortion.SpiNeMap targets average ISI distortion on global synapses and can also optimize maximum distortion.Source: §II · p. 2
  • Reducing spikes on global synapses reduces shared-interconnect congestion and ISI distortion, which can improve application performance; SpiNeMap's partitioning is designed to scale beyond PSOPART.The passage describes this as the intuition behind PSOPART and SpiNeMap, while noting SpiNeMap's improved scalability for larger problem sizes.Source: §II · p. 3

A. High-Level overview and difference with state-of-the-art

The overview contrasts prior mapping approaches with SpiNeMap's explicit placement step. SpiNeMap uses spike-trace timing to evaluate placement alternatives for energy and latency, alongside partitioning that reduces global spikes.

A. High-Level overview and difference with state-of-the-art

  • NEUTRAMS and PACMAN use a three-step deployment flow: train and validate, pack neurons and synapses to minimize resources, then deploy for inference.These approaches minimize hardware resource requirements during packing.Source: A. High-Level overview and difference… · p. 3
  • PSOPART minimizes spikes on global synapses to reduce ISI distortion, using validation-data spike counts for partitioning.The passage describes PSOPART as a three-step inference deployment approach.Source: A. High-Level overview and difference… · p. 3
  • The illustrated example partitions an eight-neuron SNN into three clusters, distinguishing local from global synapses.The example reports eight total spikes on global synapses.Source: Figure 3 · p. 3
  • SpiNeMap adds an explicit placement step that uses precise CARLsim spike timing to simulate interconnect latency and energy for placement alternatives.Its spike traces include timing information, in addition to spike counts, and its partitioning and placement steps jointly target performance, energy, and latency.Source: A. High-Level overview and difference… · p. 3

B. Detailed design of SNN Partitioning via SpiNeCluster

SpiNeCluster partitions an SNN into clusters to reduce spikes on global synapses, using a greedy approach designed to scale to large SNNs.

B. Detailed design of SNN Partitioning via SpiNeCluster

1) Time complexity:

SpiNePlacer maps clustered SNNs onto crossbars, where placement affects interconnect distance, energy, and latency; Noxim++ models these effects for placement evaluation.

1) Time complexity:

  • Algorithm 1's complexity analysis counts repeated passes over cluster pairs and neuron sweeps within a cluster.Source: 1) Time complexity · p. 4

C. Detailed design of SNN Placement via SpiNePlacer

  • Different placements of partitioned clusters traverse different interconnect lengths, affecting energy consumption and latency.Source: C. Detailed design of SNN Placement… · p. 4
  • Noxim++ extends cycle-accurate Noxim to simulate CARLsim spike traces, neuromorphic interconnect topologies and routing, and technology-specific wire and switch energy and latency.It is integrated into SpiNePlacer and configured to model DynapSE.Source: C. Detailed design of SNN Placement… · p. 4
  • The placement problem maps clustered SNN H(C, E) onto architecture A(V, I), whose V set contains crossbars and I set contains their interconnect connections.Source: C. Detailed design of SNN Placement… · p. 4
  • The mapping is a binary cluster-by-crossbar matrix, with each entry indicating whether a cluster maps to a crossbar.The stated constraints include assigning a cluster to only one crossbar.Source: C. Detailed design of SNN Placement… · p. 4

2. A crossbar can accommodate at most one cluster, i.e.,

SpiNePlacer evaluates cluster mappings with Noxim++ and uses binary PSO to search for a mapping that minimizes average spike hop count.

2. A crossbar can accommodate at most one cluster, i.e.,

  • SpiNePlacer seeks the mapping with minimum average hop count, which prior work equates with minimizing energy consumption and spike latency on the interconnect.Source: §2 · pp. 4–5
  • Each PSO particle represents a cluster mapping, with search dimension D equal to |C| × |V|.The mapping is represented as a binary logical matrix.Source: §2 · p. 5
  • PSO updates particle positions and velocities using their own and neighbors’ best experience, then binarizes the updates to preserve the mapping formulation.The algorithm updates personal and global best positions after evaluating each particle’s fitness.Source: §2 · p. 5

1) PSO Algorithm:

SpiNeMap minimizes global-interconnect spikes because their increase is associated with higher latency, greater ISI distortion, and lower application accuracy.

1) PSO Algorithm:

  • The PSO placement algorithm initializes feasible particles, evaluates fitness, updates positions using local and global bests, and binarizes the updates.Its stated complexity is O(nISO × np × |C| × |V|).Source: 1) PSO Algorithm · p. 5

1) Minimize spike count at the partitioning stage:

  • Increasing shared-interconnect spikes increases latency and average ISI distortion and lowers application accuracy; the authors report similar behavior across evaluated applications.Source: 1) Minimize spike count at the… · p. 6

2) Integration of Noxim++ within PSO:

  • Noxim++ is integrated within PSO because hop count depends on both cluster mapping and routing, requiring cycle-accurate simulation of each mapping with CARLsim spike traces.Source: 2) Integration of Noxim++ within PSO · p. 6

3) Using PSO only for SpiNePlacer:

  • Binary PSO is described as less computationally expensive and faster-converging than genetic algorithms or simulated annealing.Source: 3) Using PSO only for SpiNePlacer · p. 6
  • SpiNeMap uses PSO for placement but a greedy method for partitioning because the partitioning search becomes intractable for modest SNN sizes.The compared PSO dimensions are D = |N| × |C| for partitioning and D = |C| × |V| for placement.Source: 3) Using PSO only for SpiNePlacer · p. 6

IV. EVALUATION METHODOLOGY

SpiNeMap is evaluated using simulated spike traces, hardware and interconnect models, and a suite of synthetic and realistic SNN applications. Its results are compared with three existing mapping approaches.

IV. EVALUATION METHODOLOGY

  • CARLsim supplies spike times, while extended Noxim++ simulates spike communication and estimates interconnect energy, latency, and ISI distortion.Spike timing from CARLsim is passed as a trace to Noxim++.Source: §IV · p. 6
  • The DynapSE model uses 256-neuron crossbars and a multi-stage network-on-chip, with hardware energy and latency values and technology-scaled parameters.The model is configured at 65 nm and uses predictive technology mapping to scale parameters to 28 nm.Source: §IV · p. 6

B. Evaluated applications

  • The test set contains seven synthetic and eight realistic SNN applications, including image processing, ECG analysis, and neural-network-based classification.The synthetic networks range from 1,000 to 4,000 neurons; realistic applications include converted CNN models.Source: B. Evaluated applications · pp. 6–7, Table III · p. 7

C. Evaluated state-of-the-art techniques

  • The evaluation compares SpiNeMap with the Baseline, SCO, and PSOPART; SpiNeMap combines spike-minimizing SpiNeCluster with SpiNePlacer, which optimizes cluster placement for energy and latency.The Baseline uses NEUTRAMS to minimize crossbar use, SCO balances crossbar utilization, and PSOPART minimizes shared-interconnect spikes.Source: C. Evaluated state-of-the-art techniques · p. 7

D. Evaluated metrics

The evaluation measures shared-interconnect traffic and its performance effects using four metrics. It also summarizes average improvements against three comparison approaches.

D. Evaluated metrics

  • The four evaluated metrics are shared-interconnect spike count, average spike latency, interconnect energy consumption, and average ISI distortion.The improvement summary compares SpiNeMap against the Baseline, SCO, and PSOPART.Source: D. Evaluated metrics · p. 7, A. Improvement summary · p. 7
  • Spike count is the number of spikes on the shared interconnect after synapse clusters are mapped to hardware crossbars.Spike latency is the average delay spikes experience before reaching their destinations; the metric accounts for hop and wire delays.Source: D. Evaluated metrics · p. 7
  • Interconnect energy totals the energy consumed by shared-interconnect spikes, while ISI distortion is averaged across spikes using Equation 7.The energy metric accounts for wire and hop energy.Source: D. Evaluated metrics · p. 7

B. Energy consumption on the shared interconnect

Across the evaluated applications, SpiNeMap reports lower shared-interconnect energy, latency, ISI distortion, and spike count, alongside higher application accuracy. Its results improve on the Baseline, SCO, and PSOPART across these measures.

B. Energy consumption on the shared interconnect

  • 45% lower average energy consumption than the Baseline, 40% lower than SCO, and 20% lower than PSOPART; SpiNeMap has the lowest energy consumption among evaluated systems.The authors attribute the result to reducing shared-interconnect spikes and placing clusters to minimize interconnect energy and latency.Source: B. Energy consumption on the shared… · p. 8

C. Spike latency on the shared interconnect

  • 21% lower average spike latency than the Baseline, 27% lower than SCO, and 13% lower than PSOPART; SpiNeMap has the lowest latency among evaluated systems.The authors cite fewer shared-interconnect spikes and fewer average hops to destinations as contributing policies.Source: Figure 9 · p. 8, C. Spike latency on the shared… · p. 8

D. Average ISI distortion of spikes on the shared interconnect

  • 36% lower average ISI distortion than the Baseline, 39% lower than SCO, and 23% lower than PSOPART; SpiNeMap has the lowest distortion among evaluated systems.The authors attribute the improvement over PSOPART to SpiNePlacer, which further reduces ISI distortion while reducing latency.Source: Figure 10 · p. 8, D. Average ISI distortion of spikes on… · p. 8

E. Application accuracy

  • 12% higher average application accuracy than the Baseline, 20% higher than SCO, and 5% higher than PSOPART; SpiNeMap achieves the highest average accuracy.The accuracy results are reported as correlating with ISI distortion results.Source: E. Application accuracy · p. 8, Figure 11 · p. 9

F. Evaluation of SpiNeCluster in terms of spike count

  • 26% fewer shared-interconnect spikes than the Baseline, 24% fewer than SCO, and 9% fewer than PSOPART; SpiNeMap generates the fewest spikes.The authors attribute the improvement over PSOPART to SpiNeCluster’s greedy algorithm, particularly for large applications.Source: Figure 12 · p. 9, F. Evaluation of SpiNeCluster in terms… · p. 9

G. Evaluation of SpiNeCluster in terms of optimization time

SpiNeCluster runs faster and produces fewer shared-interconnect spikes than the PSOPART clustering approach. The paper also evaluates routing and interconnect alternatives and situates SpiNeMap among related work and tools.

G. Evaluation of SpiNeCluster in terms of optimization time

H. Interconnect design explorations

  • NorthLast and WestFirst routing have 7% and 4% higher average latency than XY, while the segmented bus has the lowest latency among the evaluated routing alternatives.The segmented bus averages 54% lower spike latency than the three evaluated routing algorithms; the methodology can simulate these and other interconnects.Source: H. Interconnect design explorations · p. 9, Figure 14 · p. 9

VI. RELATED WORKS

  • The paper describes SpiNeMap as the first work to jointly address SNN partitioning and placement on crossbar-based neuromorphic hardware while targeting energy, latency, ISI distortion, and application accuracy.It contrasts the combined approach with PSOPART, which partitions synapses but does not address their hardware placement.Source: §VI · pp. 9–10

A. SNN-based machine learning

  • The evaluation includes SNN applications from the MLPerf benchmark suite and applications converted from analog computation models to spiking networks.The last three realistic applications are identified as MLPerf applications and converted using N2D2.Source: A. SNN-based machine learning · p. 10

B. Neuromorphic hardware

  • SpiNeMap focuses on chip architecture integrating multiple crossbars, unlike cited work focused on crossbar design or other mapping platforms.The related-work discussion compares SpiNeMap with mapping work including NEUTRAMS.Source: B. Neuromorphic hardware · p. 10

C. SNN and neuromorphic simulators

  • CARLsim is selected for its STDP and homeostasis models, parameter tuning, and multi-GPU support, but SpiNeMap can also be combined with other SNN simulators.The discussion names PyNN, Brian, GeNN, and NEST as other simulators.Source: C. SNN and neuromorphic simulators · p. 10

D. Related concepts in similar domains

  • The authors select SpiNeCluster because it is scalable and provides a good starting solution for SpiNePlacer; graph-partitioning techniques are not directly transferable because ISI distortion is SNN-specific.Source: D. Related concepts in similar domains · p. 10

VII. CONCLUSION AND FUTURE OUTLOOK

SpiNeMap combines clustering and placement to map SNNs to neuromorphic hardware, with reported reductions in interconnect energy, spike latency, and ISI distortion alongside improved application accuracy. The outlook describes extending the methodology to emerging SNN approaches, other simulators, and additional hardware platforms.

VII. CONCLUSION AND FUTURE OUTLOOK

  • SpiNeMap clusters synapses into local and global groups, then uses particle swarm optimization to place clusters on physical crossbars.Clustering seeks to reduce shared-interconnect spikes and ISI distortion; placement optimizes energy consumption and spike latency.Source: §VII · p. 10
  • 45% lower energy consumption on the shared interconnect and 21% lower spike latency than state-of-the-art techniques.The authors also report 36% lower ISI distortion and 12% higher application accuracy over state-of-the-art approaches.Source: §VII · p. 10

1) Mapping new machine learning approaches to hardware:

  • Emerging spiking approaches can be mapped by simulating application behavior in CARLsim and using spike traces to partition and place clusters.The paper demonstrates this workflow with a liquid state machine in HeartEstm; SpiNeCluster can also be extended to minimize spike disorder for rate models.Source: 1) Mapping new machine learning… · p. 11
  • Analog CNN and MLP applications can enter the methodology after conversion to spike-based models; three CNN applications were converted using N2D2.The authors describe this route for algorithms designed with the analog model.Source: 1) Mapping new machine learning… · p. 11

2) Using SpiNeMap for other neuromorphic platforms:

  • SpiNeMap is demonstrated on DynapSE and integrated with Brian and XNet; other interconnect simulators must support cycle-accurate and trace-driven simulation.Support for TrueNorth and Loihi is described as continuing work, and the framework is open-sourced.Source: §VII · p. 10, 2) Using SpiNeMap for other… · p. 11
Loading 1909.01843v1…