Lit Silicon Artifact

Setup

GitHub

Clone the repository with submodules

git clone https://github.com/UnaryLab/lit_silicon_tuning_amd.git --recursive
cd lit_silicon_tuning_amd

Set up a python virtual environment

Python Environment

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python 3.12.0
.venv/bin/activate

Install chopper

uv pip install third_party/chopper

Huggingface

Request access to Llama 3.1 8B. Make sure to export your huggingface key.

export HF_TOKEN=<key>

Run

This script will submit all jobs needed to build the container, collect the data, and plot figures.

./submit_all.sh account_name

Validation

Once all jobs are completed, match the generated plots with their corresponding figures in the paper:

• Figure 16

  • a) deepseek_llama_straggler_per_gpu.png
  • b) deepseek_llama_lead_and_throughput.png
  • c) deepseek_llama_avg_pow_freq.png

• Figure 9

  • a) red_realloc_slosh_lead_and_throughput.png
  • b) red_realloc_slosh_freq_pow.png

Citation

@inproceedings{2026isca_litsilicon,
  author    = {Marco Kurzynski and Shaizeen Aga and Di Wu},
  title     = {{Lit Silicon: A Case Where Thermal Imbalance Couples Concurrent Execution in Multiple GPUs}},
  booktitle = {International Symposium on Computer Architecture},
  year      = {2026},
}