SCATTER

By Ziang Yin, Nicholas Gangi, Meng Zhang, Jeff Zhang, Rena Huang and Jiaqi Gu^†.

This repo is the official implementation of "SCATTER: Algorithm-Circuit Co-Sparse Photonic Accelerator with Thermal-Tolerant, Power-Efficient In-situ Light Redistribution"

Introduction

SCATTER is a novel algorithm-circuit co-sparse photonic accelerator featuring dynamically reconfigurable signal path via thermal-tolerant, power-efficient in-situ light redistribution and power gating.

SCATTER is designed to overcome the limitations of traditional photonic accelerators with these key features: ➊ phase-agnostic incoherent photonic tensor cores for robust tensor computing; ➋ shared input modulation modules and readout circuitry to balance area, power, and control flexibility;

➌ in-situ tunable rerouter for light redistribution;

➍ hardware gating to support structured row-column sparsity with enhanced power efficiency and thermal crosstalk robustness;

➎ co-optimized devices, circuits, and architecture configurations with maximum efficiency and thermal variation tolerance.

SCATTER achieves a 511× area reduction and 12.4× power saving with superior crosstalk tolerance that enables unprecedented circuit layout compactness and on-chip power efficiency.

A power-optimized, crosstalk-aware dynamic sparse training framework is introduced to explore row-column structured sparsity and ensure marginal accuracy loss and maximum power efficiency.

Dependencies

• Python >= 3.8
• pyutils >= 0.0.2. See pyutils for installation.
• pytorch-onn >= 0.0.5. See pytorch-onn for installation.
• Python libraries listed in requirements.txt
• NVIDIA GPUs and CUDA >= 10.2

Structures

• configs/: configuration files
• core/
  • models/
    • layers/
      • base_layer.py: base layer definition
      • tempo_conv2d.py: TeMPO convolution 2D block definition
      • tempo_linear.py: TeMPO linear block definition
      • utils.py: utility functions layers
    • dst2.py: pruning scheduler definition
    • tempo_base.py: TeMPO base model definition
    • tempo_cnn.py: TeMPO CNN model definition
    • tempo_resnet.py: TeMPO resnet model definition
    • tempo_vgg.py: TeMPO vgg model definition
    • utils.py: utility functions for models
  • builder.py: build training utilities
  • utils.py: customized loss function definition
• hardware/
  • photonic_core_base.py: architecture simulator base definition
  • photonic_crossbar.py: architecture simulator for TeMPO definition
• MZIdata/
  • MZIPower.csv: real MZI power data
• scripts/: contains experiment scripts
• crosstalk_spacing.py: inference logic
• sparse_train.py: sparsity train logic
• train.py: train logic

Usage

• Train benchmark MODEL using SCATTER's algorithm, replace MODEL with one option from (cifar10/vgg8, cifar100/resnet18, fmnist/cnn/train),
  > python3 scripts/MODEL/train_origin.py
• Or, train sparse MODEL using SCATTER's algorithm, replace MODEL with one option from (cifar10/vgg8, cifar100/resnet18, fmnist/cnn/train),
  > python3 scripts/MODEL/train_sparse.py
• Test the trained model with different hardware restriction. (Need to modify the path to trained checkpoint in the script). python3 scripts/MODEL/test_main_results.py

Citing SCATTER

@inproceedings{yin2024scatter,
      title={SCATTER: Algorithm-Circuit Co-Sparse Photonic Accelerator with Thermal-Tolerant, Power-Efficient In-situ Light Redistribution}, 
      author={Ziang Yin and Nicholas Gangi and Meng Zhang and Jeff Zhang and Rena Huang and Jiaqi Gu},
      year={2024},
      booktitle={International Conference on Computer-Aided Design (ICCAD)},
      url={https://arxiv.org/abs/2407.05510}, 
}