Installation

Prerequisites

• Linux or WSL2

• A CUDA-enabled GPU

• MiniConda or Anaconda (required to install the CUDA Toolkit)

Tested environments

All experiments are run on Linux. Windows is not tested.

• NVIDIA A6000 48 GB × 8, Ubuntu 24.04, CUDA 12.4

• NVIDIA H100 96 GB × 2, Red Hat Enterprise Linux 9.5, CUDA 12.6

• NVIDIA H100 80 GB × 8, Ubuntu 24.04, CUDA 12.4

• NVIDIA H200 141 GB × 8, Ubuntu 24.04, CUDA 12.4

Step-by-Step Setup

1. Configure SSH access to GitHub. One dependency — MASE — requires SSH to clone. Follow GitHub’s guide on Connecting to GitHub with SSH and ensure ~/.ssh/config is set up correctly.

2. Clone the repository.

   git clone https://github.com/AICrossSim/NewComputeBench.git
   cd NewComputeBench
   # Download the MASE dependency, which is bundled as a git submodule
   git submodule update --init
   

3. Activate the environment and install dependencies.

   Install the required packages. Choose one option:

   Option 1 — uv (recommended, assumes CUDA is pre-installed on the system):

   uv sync
   uv pip install ./submodules/mase
   

   Note

   uv sync reads pyproject.toml and uv.lock to reproduce the exact environment. Run git submodule update --init before this step to ensure the MASE submodule is available.

   Option 2 — conda + pip (use this if CUDA is not pre-installed):

   conda env create -f environment.yaml
   conda activate new-compute
   pip install -r requirements.txt
   pip install ./submodules/mase
   

   Note

   The MASE submodule provides the quantization backend used by PIM and other hardware simulation passes.

   Note

   uv can be installed with pip install uv or via the standalone installer:

   curl -LsSf https://astral.sh/uv/install.sh | sh
   

   Option 3 — Docker (fully isolated, reproducible environment):

   git clone https://github.com/AICrossSim/NewComputeBench.git
   cd NewComputeBench
   # Download the MASE dependency, which is bundled as a git submodule
   git submodule update --init
   docker build -t newcomputebench .
   

   The image is based on nvidia/cuda:12.4.1-devel-ubuntu22.04 and uses uv to install all dependencies from the lockfile. All source code and dependencies are baked into the image at /app, so the container is fully self-contained and reproducible.

   For development, mount your local project directory so that edits on the host are reflected inside the container:

   docker run --gpus all -v $(pwd):/workspace:z -it newcomputebench bash
   

   Note

   The NVIDIA Container Toolkit must be installed for --gpus to work. Submodules must be checked out before docker build because the build context does not include .git.

4. (Optional) Log in to Weights & Biases to track experiment metrics.

   wandb login