MP 7,1 Infinity Fabric Link use

[PowerMike G5]

I've been thinking of upgrading my Mac Pro with another Pro Vega II GPU, for use with various 4K/8K video projects in Adobe CC. Mainly working with a lot of 10-bit 4:2:2 XF-AVC, Red Raw, Canon Cinema Raw Light, etc. A good mix of media codecs. Color grading mainly through Lumetri Color, but use various GPU-accelerated effects and plugins like NeatVideo.

I've been researching Infinity Fabric, but can't seem to find too much information on its usage. I see that it provides a far faster interconnect between the 2 GPUs, but I'd assume that an application has to be programmed to take advantage of IFL specifically, correct?

Does anyone know if Adobe Premiere Pro treats 2x Pro Vega II GPUs with IFL as one GPU, similar to how it sees Crossfire on the Windows side? Or if it sees it as 2 separate GPUs, will it at least take advantage of IFL, or does it treat the 2 GPUs and just 2 separate GPUs connected via separate PCIe slots.

Any insight would be helpful in determining if this is a worthwhile upgrade.

 

[DrEGPU]

Apps would need to be written to specifically take a data of the IF links, but I don’t know if adobe apps do that. Try contacting them?

 

[PowerMike G5]

Thanks. I've reached out to AMD and will try to reach out to Adobe as well.

I'm not sure if a lot of users here are using the setup above, but will try to report back any knowledge I find about this.

 

[deconstruct60]

Does anyone know if Adobe Premiere Pro treats 2x Pro Vega II GPUs with IFL as one GPU, similar to how it sees Crossfire on the Windows side? Or if it sees it as 2 separate GPUs, will it at least take advantage of IFL, or does it treat the 2 GPUs and just 2 separate GPUs connected via separate PCIe slots.

Apple Metal (and the GPU driver stack ) does not treat it as one GPU, so it is hard to image any app goes through the gyrations of trying to create that "virtual" GPU out of that given all the overhead the app would have to manage for itself. In some limited contexts can left Metal do a portion of the work of chopping up and distributing and colasing the work. However, even that is an explicit set of commands in the code required.

There are some additions to Metal that make two GPU connected with IFL coordinate their memory sharing in a convenient fashion that improves performance. ( the mem copies and/or read/writes go through IFL rather than the default through PCI-e data paths). That has to be explicit code written by the app though and only invoked after checking the links are present.

Any insight would be helpful in determining if this is a worthwhile upgrade.

The question to ask app developers (or their support "knowledge base" is one if the app does workload scaling over multiple GPUs at all. Second, if that code has been enhanced to incorporate the addtional Metal changes to take advantage of it. ( Kind of a corollary. If say only using OpenCL to do GPGPU compute tasks then I don't think Apple extended the IFL link aware additions to that ; in part because they have deprecated it. )



P.S. As a whole Premiere Pro won't. Specific subcomponents may. Also depends upon which parts of program (and/or extensions) being used.