Multi-GPU support timeline?

Question

rcroman OP

Created Jul ’21

Replies 2

Boosts 0

Participants 2

Thanks very much for tensorflow-metal.

Is there a timeline for when it will support multiple GPUs?

Alternately, is there a way to start two instances and force each to different GPUs? Then two sessions could be run in parallel.

Thanks!

Boost

Answer 1

Frameworks Engineer OP

Apple

Aug ’21

Hi @rcroman, Thanks for your interest. We don't have any update on multi-GPU support. For starting two instances and running different GPUs, you can looking into Distributed training(https://www.tensorflow.org/guide/distributed_training) with Tensorflow. Specifically on the MultiWorker strategies which allow training on multiple devices.

0

Answer 2

rcroman OP

Aug ’21

Thanks for the reply. How does one tell tensorflow-metal which GPU to use? I only ever see one GPU, even though I have two installed:

tf.config.list_physical_devices('GPU')

[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

When first defining a model, I get a message similar to the following:

Metal device set to: AMD Radeon PRO W6900X

Usually this is my faster GPU, sometimes not... it seems random, and only changes when I restart the machine. Is there a way in tensorflow-metal or PluggableDevice to 1) enumerate the actually-installed GPUs, and 2) select which I want to use?

0