We benchmark real TeraFLOPS that training Transformer models can achieve on various GPUs, including single GPU, multi-GPUs, and multi-machines. It helps you to estimate how many machine times you need to train your large-scale Transformer models.

The real performance depends on multiple factors, including your hardware, cooling, CUDA version, transformer models, hyper-parameters such as batch sizes, and implementations. We list the numbers we got on both personal PC and cloud instances. We also provide Jupyter notebooks for you to benchmark on your machines and workloads:

• Understanding Transformer layer performance
• Training BERT and GPT-2 with (multi-)GPUs

Measure the TFLOPS for various micro-benchmarkings. Results are from running micro_bench.ipynb.

You need a CUDA-enabled pytorch to run workloads. We recommend you to use the latest version CUDA and pytorch for better performance. One easy way is using nvidia docker. Once installed, you can find latest tag of the pytorch image, for exmaple, 22.07-py3, then run

sudo docker run --gpus all -it --rm -p 8888:8888 -v ~:/workspace \
	--ipc=host --ulimit memlock=-1 --ulimit stack=67108864 \
	nvcr.io/nvidia/pytorch:22.07-py3

After the docker is running, execute jupyter notebook in the docker's shell to open this notebook.