📟OpenIsingHW

This repository aims at analyzing the system-level performance (like energy, latency and area) for Ising accelerators under variable applications.

Instead of assuming perfect hardware utilization and unlimited peripheral memories, the framework considers the under-utilization due to suboptimal mapping and peripheral memory overhead.

Features of the framework

• The first analytical system-level simulation framework for digital Ising accelerators.
• Both serial-updating architecture and parallel-updating architecture are supported.
• Graph sparsity is considered if weight compression is applied.
• Different problem types and sizes can be flexibly configured.

Supported Ising Problem Types and Features

Features	MaxCut	TSP	Sudoku	MIMO
Problem-specific weight	Y	Y	N	N
Graph density	appro. 0.015	0.05-0.25	0.04	1
Weight precision	1b/2b	10b-16b	3b	~16b
With magnetic field (h)	N	Y	Y	Y
Typical problem size (M)	800-4,000 nodes	10-100 cities	81 cells	4-32 users
Required Ising nodes (N)	800-4,000	100-10,000	<729	8-256
Average gragh degree	4-50	2 ($\sqrt{N}-1$)	28	N - 1

Getting Started

Requirements

• Python Version: 3.12
• Python-deps: Automatically installed via pip using the provided setup script.

Setup

cd openisinghw
pip install -r ./requirements.txt
source .setup

How to get results

To simulate, just run:

python main.py

When evaluating a different workload or architecture, just modify the input files before run the command. Input configuration files (YAML) are within the folder ./inputs (please see the readme within the folder for further details):

• hardware: the Ising hardware architecture specification, including parallelism, memories and weight compression method.

• mapping: the mapping constraint specification of the ising architecture.

• workload: the Ising workload specification, like problem size, average degree, variable precision.

If the memory compiler is not available by hand, the repository has incorporated the open-sourced CACTI 6.0 for use. Just run the get_cacti_cost.py after modifying its memory specification.

Note: CACTI is not a commercial memory compiler and its results may differ from real memory compiler.

Model validation

Since cost model validation is important for analytical simulation framework, we have conducted several model validations against state-of-the-art accelerators. Relevant scripts are under the folder ./model_validation (please see the readme within the folder for further details). The validation result is shown below.

Re-generate the results in the paper

To re-generate the figures of the case studies in the paper, you can run the following scripts in python:

• exp_mac.py: the script for the first case study on MAC parallelism analysis.
• exp_encoding.py: the script for the second case study on the compression method analysis.
• exp_iteration.py: the script for the thrid case study on the solution quality analysis (please uncomment the wanted problem in the pb_pool).
• exp_simulation_time.py: the script for the simulation time comparison analysis.