KAMI: Communication-Avoiding General Matrix Multiplication within a Single GPU

This repository contains the artifact for the SC '25 paper submission "KAMI: Communication-Avoiding General Matrix Multiplication within a Single GPU."

Artifact Setup

Hardware

KAMI is evaluated on four GPUs: NVIDIA GH200, NVIDIA RTX 5090, AMD 7900 XTX and Intel Max 1100.

Software

• The NVIDIA GH200 is installed with Ubuntu 22.04 using GCC v11.4 and NVCC v12.8.

• The NVIDIA RTX 5090 is installed with Ubuntu 24.04 using GCC v11.4 and NVCC v12.8.

• The AMD 7900 XTX is installed with Ubuntu 24.04 using GCC v11.4 and ROCm 6.10.

• The Intel Max 1100 is using intel® Tiber™ AI Cloud.

Steps to Reproduce

1. Clone the repository:

   git clone https://github.com/ForADAE/SC25-pap926.git

2. Navigate into the repository:

   cd SC25-pap926

3. Run evaluation scripts (Optional):

   Navigate to the scripts/ directory:

   cd scripts

   Then, depending on your hardware:

   • For NVIDIA GH200:

     bash all_GH200.sh

   • For NVIDIA RTX 5090:

     bash all_5090.sh

   • For AMD 7900 XTX:

     bash all_AMD.sh

   • For Intel Max 1100:

     bash all_intel.sh

     The output logs will be stored in the logs/ directory. Each full benchmark run typically takes 100 minutes.

4. Reproduce paper plots:

   Navigate to the plots/ directory:

   cd plots

   Install necessary Python packages (if not already installed):

   pip3 install numpy pandas matplotlib seaborn

   Then run:

   bash plots_all.sh

Expected Results

The artifact includes performance results for KAMI in both double-precision (FP64) and half-precision (FP16) floating-point formats, along with baseline implementations including cuBLAS, cuBLASDx, CUTLASS, MAGMA, and SYCL-Bench.

Upon successful execution:

• Performance: In all tested configurations (matrix sizes, precisions, and hardware platforms), KAMI is expected to outperform cuBLAS, cuBLASDx, CUTLASS, MAGMA, and SYCL-Bench in terms of achieved TFLOPS.
• Output Files: The logs containing timing results will be saved under the logs/ directory, with one file per benchmark run.
• Plots: The plots/ directory will contain regenerated figures identical to those in the SC '25 paper submission.