MLCAD26_Contest_Scripts_Benchmarks

Disclaimer: Please read the MLCAD Contest description file available in this repository, which describes the contest, before looking into the benchmarks and scripts available in this repository. The README file only helps navigate and understand the benchmarks and scripts; it does not serve as the contest description.

This GitHub repository has the public benchmarks for the 2026 MLCAD Contest to work on Agentic Algorithm Discovery for Timing Optimization. Hidden benchmarks will be released as per the schedule mentioned [here]

Prior timing-optimization contests have largely relied on fixed heuristics applied uniformly across designs. While such approaches provide strong baselines, they do not adapt to design-specific characteristics or evolving tool behaviors. This contest's primary goal is to challenge participants to move beyond static heuristics by introducing adaptive, learning-based, or agentic approaches to timing optimization — framing the problem as one of algorithm discovery, where participants may leverage LLMs to automatically identify design-specific heuristics that improve power, performance, and area (PPA) relative to a one-size-fits-all baseline. The notion of algorithm discovery is inspired by recent advances in automated reasoning and optimization research for EDA tooling.The secondary goal is to stimulate the use of LLMs to design adaptive, tool-evolving methodologies and to discover algorithms for joint timing optimization and detailed placement.

Table of Content

• asap7: ASAP7 cell library for the designs.
• benchmarks: Contains benchmark designs against which contestants have to improve their resizer.
• evaluation: Scripts to check netlist equivalence and perform evaluation flow
• OpenROAD: Version of OpenROAD that will be used to compare contestants solutions.
• sample_algorithm_discovery_flow_and_resources: Folder containing sample Algorithm-discovery flow and relevant research papers
• MLCAD2026-Contest-Proposal-Algorithm-Discovery.pdf: Contest description.

Install and Build OpenROAD

Option 1: Build locally

Please use the copy of OpenROAD provided in this directory. We will use this same copy of OpenROAD to evaluate participant's solutions. We have slightly changed the source code in this version of OpenROAD to check for logic equivalence, and participants are recommened to use this copy of OpenROAD to avoid any inconsistencies in their own evaluation.

git clone https://github.com/ASU-VDA-Lab/MLCAD26-Contest-Scripts-Benchmarks.git
cd MLCAD26-Contest-Scripts-Benchmarks
git submodule update --init --recursive
cd OpenROAD 
git checkout MLCAD26
sudo ./etc/DependencyInstaller.sh -all
./etc/Build.sh

Option 2: Docker Image

To develop solution using a pre-built docker image that has this repo pre-installed along with OpenROAD pre-built, please use the following commands

docker pull tsjafri/mlcad2026:2026-03-24
docker run -it --name mlcad2026 --hostname MLCAD2026 tsjafri/mlcad2026:2026-03-24 /bin/bash -c "cd && cd MLCAD26-Contest-Scripts-Benchmarks/ && git pull && exec bash"

The above commands will setup a Ubuntu-24.04 docker container with this repo and OpenROAD pre-installed and pre-built.

Benchmark Statistics

Design Name	Clock(ps)	Num. of instances	WNS(ns)	TNS(ns)	Leakage power(uW)	Total power(uW)	Slew violation count	Capacitance violation count	Fanout violation count
aes_cipher_top	200	20097	-0.0853	-12.44	96.30	432000.0	0	0	0
jpeg_encoder	350	64602	-0.0908	-48.77	174	294000.0	0	0	0
ariane	900	210455	-0.9413	-7568.43	17900	658000	1508	0	0
nvdla_p	400	106994	-0.2973	-198.18	306	40000	0	0	0
nvdla_c	650	312696	-0.202	-55.58	17300	601000	437	0	0

Reproducing baseline metrics and computing score

• eval.sh - Script to run the evaluation flow for contestant-optimized designs. This script will run the evaluation flow and generate the output files in evaluation/<design_name> folder.
• evaluation.tcl - Evaluation flow script that will be sourced by eval.sh
• evaluation_baseline.tcl- script to reproduce the baseline metrics. Replace the evaluation.tcl with evaluation_baseline.tcl in eval.sh to reproduce the baseline metrics. This script uses OpenROAD's built-in resizer and placer(repair_timing) for optimization.
• parse_log.py - Script to parse the OpenROAD log file generated by evaluation.tcl and extract the relevant metrics for score computation.
• compute_score.py - Script to compute the final score for contestant-optimized designs using the metrics generated by evaluation.tcl.

Questions

Questions related to the contest can be emailed to mlcad2026-contest@googlegroups.com or can be filed as issues in this GitHub repo here