tt-matmul

Benchmarking optimised matrix multiplication kernels on the Tenstorrent Wormhole n300 accelerator card. For the step-wise incremental optimisations, see the various kernels in the src directory:

• matmul_single_core.cpp
• matmul_multi_core.cpp
• matmul_multicore_reuse.cpp
• matmul_multicore_reuse_mcast.cpp

Build

Clone the repository, initialise the tt_power submodule and build the code using the provided Makefile. Note, for power draw measurements, we require openMP for threading in the host code.

git clone git@github.com:markxio/tt-matmul.git
git submodule init
git submodule update
make all

Setup python for utils/average.py:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Run

Call the run script which sets run arguments and writes performance data to output/perf_avg.csv. See the provided run.sh script for details:

./run.sh