Can PyMC3 (or even PyMC4) make use of the AMD Radeon "Navi" 5500M GPU in a Macbook Pro 16?

Questions
,
1

Hi gang,

As per title… I’m thinking of getting an MBP 16 and curious whether the GPU is usable by pymc3/theano or pymc4/TF2.

For pymc3 it would seem to need a combination of OpenCL (unclear that OpenCL is even supported by Navi), and possibly incurs unnecessary overhead anyhow (Leveraging the GPU in PyMC3, Gpu much slower than cpu)

For pymc4 though… I’ve heard that Google have a project to internalise CUDA to allow TF2 to run on non Nvidia chips - is it at all possible that TF2 would run on this GPU within the next year or so?

I should add that this would be a ‘cherry on top’ scenario. As/when I have big models/datasets then cloud is the only practical choice

Cheers, Jon

2

At this point in time, none of those developments are nearby to the present, not only because there is no known standard way of parallelizing samplers (as opposed to neural network architectures), but because there is no good reason, afaik, to provide such support to AMD hardware, when Nvidia is available.

1 Like
3

Thanks, that’s basically as I expected :smiley: