ELSA: An ELastic SNN Inference Architecture for Efficient Neuromorphic Computing

Official implementation of ELSA (ISCA 2026)

This repository contains the artifact submission for the paper ELSA accepted at ISCA 2026 (Paper ID: 124). The artifact supports reproducing the results presented in the paper, including evaluations of Spiking Neural Networks (SNNs) at each time-step and performance assessment of the ELSA accelerator.

ELSA: An ELastic SNN Inference Architecture for Efficient Neuromorphic Computing
Kang You, Chen Nie, Lee Jun Yan, Ziling Wei, Cheng Zou, Zekai Xu, Yu Feng, Honglan Jiang, Zhezhi He
Proceedings of the 53rd Annual International Symposium on Computer Architecture (ISCA), 2026

Abstract

Spiking neural networks (SNNs) exploit event-driven and addition-only computation to substantially improve efficiency for intelligent computation. A key temporal property of SNNs, elastic inference, allows outputs to emerge progressively, enabling responses to salient inputs much earlier than full evaluation. However, existing SNN-specific accelerators cannot capitalize on this property. Layer-by-layer designs emit outputs only after all layers are complete, while time-step-by-time-step designs rely on coarse-grained, layer-wise pipelines that require synchronizing all spines/tokens within a layer. This barrier prevents results from being forwarded immediately, delaying the earliest possible response and forfeiting the benefits of elastic inference.

To address these challenges, we propose ELSA, a near-SRAM dataflow architecture that realizes true elastic inference through a fine-grained spine/token-wise pipeline and hardware optimizations tailored to SNNs. ELSA forwards each spine/token immediately upon production, forming a continuous streaming pipeline that substantially reduces the latency to the first response. To enhance this lightweight execution, ELSA introduces a bundled address event representation protocol to lower communication traffic of network-on-chip (NoC), and leverages mini-batch spiking Gustavson-product to cut memory access and exploit inherent sparsity. Combined with mapping and scheduling optimizations, ELSA achieves efficient, event-driven computation without compromising accuracy. Experiments show that SNNs can outperform quantized artificial neural networks (QANNs) while maintaining on-par accuracy. For a 4-bit ResNet-50, ELSA achieves $3.4\times$ speedup and $13.6\times$ higher energy efficiency over the SOTA QANN accelerator (ANT), and $2.9\times$ speedup and $22.1\times$ energy efficiency gains over the SOTA SNN accelerator (PAICORE).

ELSA Artifact Evaluation with Docker image

This document provides step-by-step instructions for reviewers to reproduce the main results using the provided Docker image.

1. Download the Docker Image

Please download the Docker image archive from the following link:

SwissTransfer Download Link
https://www.swisstransfer.com/d/78528f19-a5a5-4b18-bb7a-d8078ba8724e

• Filename: ELSA.tar
• Available until: May 21
• CUDA Version: 12.1

2. Verify File Integrity

To ensure the downloaded file is complete and uncorrupted, please verify its MD5 checksum.

Expected MD5

117e4ce424c954c4fced931a8f256130

Run Verification

md5sum ELSA.tar

Please confirm that the output matches the expected checksum above.

3. Load and Launch the Docker Container

After uploading ELSA.tar to your server, run the following commands.

3.1 Load the Docker Image

docker load -i ELSA.tar

3.2 Start the Container

docker run -it --gpus all --ipc=host --name ELSA elsa_image:1 /bin/bash

4. Activate the Conda Environment

Inside the container, activate the prepared environment:

conda activate ELSA

5. Reproduce Paper Results

5.1 Accuracy Results

Run:

cd ~/ELSA/ELSA_Algorithm
bash run_all.sh

Reported Accuracy

Model	Dataset	ANN Accuracy	SNN Accuracy	Elastic SNN Accuracy	Latency Reduction
VGG16	CIFAR-10	91.57	91.48	N/A	N/A
VGG16	CIFAR-100	73.94	73.85	N/A	N/A
ResNet18	ImageNet	67.611	67.537	67.27 (confidence threshold=0.3)	14.76%
ResNet34	ImageNet	71.512	71.558	68.36 (confidence threshold=0.3)	23.31%
ResNet50	ImageNet	74.842	74.776	73.40 (confidence threshold=0.3)	19.90%
ViT	ImageNet	78.398	78.66	78.39 (confidence threshold=0.8)	11.50%

5.2 Figure 16 and Figure 17

cd ~/ELSA/ELSA_Simluator

python3 run_figure16.py --cache-dir tracer_files --skip-sim

python3 run_figure17.py --cache-dir tracer_files --skip-sim

Figure 16

Figure 17

Manually Environment Setup from Scratch

1. Install Anaconda

This project requires Anaconda 3 (version 24.5.0). You can install it using the following commands:

wget https://repo.anaconda.com/archive/Anaconda3-2023.03-Linux-x86_64.sh
bash Anaconda3-2023.03-Linux-x86_64.sh

Follow the on-screen instructions to complete the installation.

2. Set up the ELSA Conda Environment

Download the ELSA Conda environment from the provided link ELSA.tar.gz and extract it:

mkdir -p ~/anaconda3/envs/ELSA
tar -xzf ELSA.tar.gz -C ~/anaconda3/envs/ELSA
source ~/anaconda3/envs/ELSA/bin/activate

This will activate the ELSA environment containing all dependencies required for running the artifact.

3. Download the checkpoints for algorithm and hardware evaluation

3.1 Download the model checkpoint for SNN accuracy and ELSA performance evaluation with url:

Checkpoints.zip

3.2 Unzip the Checkpoints.zip and get two directories:

Algorithm_Evaluator

Hardware_Simulator

3.3 Put the files in Algorithm_Evaluator to ~/ELSA_Algorithm/model_pool.

3.4 Put the files in Hardware_Simulator to ~/ELSA_Simluator/model_pool, and then run ~/ELSA_Simluator/model_pool/extract_all.sh.

After these steps, we have prepared the checkpoints for ELSA evaluation.

ELSA Algorithm

This module evaluates the accuracy of SNNs at each time-step, as well as the accuracy and latency of elastic inference.

ELSA Algorithm Artifact

Introduction

This project evaluates the accuracy of the SNN models used in ELSA, as well as the latency of elastic inference.

We evaluate multiple SNN models on ImageNet, CIFAR-10, and CIFAR-100.
The corresponding checkpoints are provided at the following link:

Checkpoint Download Link: https://zenodo.org/records/19446695

Reported Accuracy

Model	Dataset	ANN Accuracy	SNN Accuracy	Elastic SNN Accuracy	Latency Reduction
VGG16	CIFAR-10	91.57	91.48	N/A	N/A
VGG16	CIFAR-100	73.94	73.85	N/A	N/A
ResNet18	ImageNet	67.611	67.537	67.27 (confidence threshold=0.3)	14.76%
ResNet34	ImageNet	71.512	71.558	68.36 (confidence threshold=0.3)	23.31%
ResNet50	ImageNet	74.842	74.776	73.40 (confidence threshold=0.3)	19.90%
ViT	ImageNet	78.398	78.66	78.39 (confidence threshold=0.8)	11.50%

1. VGG16

This part mainly evaluates the accuracy of the VGG16 model on CIFAR-10 and CIFAR-100.

VGG16 on CIFAR-10

Go to the directory ~/ELSA_Algorithm/VGG and Run the following script:

bash run_snn_bn_vgg16_cifar10.sh

VGG16 on CIFAR-100

Go to the directory ~/ELSA_Algorithm/VGG and run the following script:

bash run_snn_bn_vgg16_cifar100.sh

2. ResNet

This part mainly evaluates the accuracy of ResNet18, ResNet34, and ResNet50 on ImageNet.

Full Inference

Go to the directory ~/ELSA_Algorithm/ResNet and run the following scripts:

bash run_snn_bn_resnet18.sh
bash run_snn_bn_resnet34.sh
bash run_snn_bn_resnet50.sh

Elastic Inference

Go to the directory ~/ELSA_Algorithm/ResNet and run the following scripts:

bash run_snn_bn_resnet18_elastic.sh
bash run_snn_bn_resnet34_elastic.sh
bash run_snn_bn_resnet50_elastic.sh

3. Vision Transformer

Full Inference

Go to the directory ~/ELSA_Algorithm/ViT and run the following script:

bash ./scripts/vit-small_SNN_16.sh

Elastic Inference

Run the following script:

bash ./scripts/vit-small_SNN_16_elastic.sh

ELSA Simulator

This module evaluates the performance of the ELSA accelerator, including energy consumption, latency, and area. It reproduces the results shown in Figure 16 and Figure 17 of the paper.

ELSA Simulator

Introduction

This project provides an end-to-end, cycle-level simulator for ELSA. The simulator supports two execution modes:

Slow Path

1. Obtain the Quantized Model Checkpoint

Noticing: You should finish the step-3 in experiment setup.

After downloading, the model path are specified in:

~\ELSA_Simulator\convolution\configs\elsa_models.yaml

You may also modify this path yourself if needed.

2. Reproduce Figure 16

Run the following command:

python3 run_figure16.py

3. Reproduce Figure 17

Run the following command:

python3 run_figure17.py

Fast Path

We provide pre-generated tracer files in:

/home/kang_you/ELSA_Simulator/tracer_files

Users can use these tracer files to directly reproduce the results of Figure 16 and Figure 17 without running full simulations.

1. Reproduce Figure 16

python3 run_figure16.py --cache-dir tracer_files --skip-sim

2. Reproduce Figure 17

python3 run_figure17.py --cache-dir tracer_files --skip-sim

Figure 16

Figure 17