Hardware Requirements

The 2021.10 Document Understanding ML Packages are different from previous releases in the way they use GPUs and CPUs.

Previous releases would train the same model architecture regardless of whether the Training Pipeline is run on a GPU or on a CPU. However, 2021.10 Document Understanding ML Packages include an optimization which reduces the size of the model when training on CPU, to accelerate the training process.

As a result, training on GPU is only five times faster than on CPU (previously it was 10-20 times faster). This also makes it possible to train models on CPU with up to 5000 pages (previously it was 500 maximum).

Please be aware that training Document Understanding models on GPU requires a GPU with at least 11GB of video RAM to run successfully.

Use the below table to check the compatibility between the ML Packages, CUDA version, and GPU driver version.

ML Packages version	CUDA version	cudDNN version	NVIDIA driver (lowest compatible version)	Hardware Generation
2021.10	CUDA 11.1	cuDNN 8.0.5	R450.36.06	Ampere, Turing, Volta, Pascal, Maxwell, Kepler

CUDA is backward compatible, meaning that existing CUDA applications can continue to be used with newer CUDA versions.

More information about compatibility can be found here