Treat the multi-GPU as single GPU and memory size

AI & Data Science Deep Learning (Training & Inference) Frameworks
1

Hi

Do you have any way to use the multi-GPU as a single GPU and memory size for machine learning with Tensorflow framework?
So, if used the DGX-Station, there are four GPUs(V100 32GB). In this case, user will use this WS as 128GB, 20,480 cuda and 2,560 tensor cores.

Best regards.
Kaka

2

There is no transparent way to make TensorFlow tranparently span all 4 GPUs like a single logical device. You would need to explicitly assign various sections of your graph to different GPUs using with tf.device() scopes.