AI Training System Requirements

Training

GPU

A GPU is strongly recommended for training deep learning models. Training performance on a GPU can be up to 20x faster than on a CPU, and can shorten training time from weeks to hours.

*GPU training requires an NVIDIA GPU with compute capability 5.0 to 9.x (Maxwell, Pascal, Volta, Turing, Ampere, or Ada Lovelace architecture), and your driver must support CUDA Toolkit 11 or higher. We recommend you update your NVIDIA driver to the latest version here. Learn more about CUDA here.

Below is a list of recommended GPU options. In general:

• More CUDA cores leads to faster training and application times
• Architectures (later is better): 1000=Pascal, 2000=Turing, 3000=Ampere, 4000=Ada Lovelace
• The more RAM the better. More important than core count. 8GB+ is recommended for deep learning training, 12GB+ for Spotlight.
• HBM2 > GDDR6X > GDDR6 > GDDR5X > GDDR5 in terms of memory bandwidth
  

Applying

GPU

A GPU is also recommended for applying deep learning models. Application performance on a GPU can be up to 20x faster than on a CPU.

CPU

Since applying often takes seconds, compared to hours for training, a CPU can more reasonably be used if necessary. Please refer to the general System Requirements for more information on CPU requirements and recommendations.