Posted this over on the Redshift forum as well as the Deadline forums and haven't received any useful feedback from either. For USD renders, Redshift is using an attribute called "redshift:global:RS_gpuX_enable" that Deadline (at least as of 10.1.14.5) doesn't seem to be aware of. Regardless of whether I override GPU Affinity per worker, or set it in the Deadline Scheduler, all available GPUs are used instead. In Solaris, the render settings->Redshift->Config actually defaults to "Set or Create" and has all GPUs enabled. Forcing this to "Do Nothing" doesn't seem to affect the behavior.
I'm aware that it's possible to work around this by setting up groups and forcing the render settings config to specific GPUs, but this isn't ideal particularly if you don't have a consistent number of GPUs per render node. Deadline should be able to manage this.
Just curious if anyone has experienced something similar, has gotten it to work, or has a better solution. This is with 18.5.499, RS 3.0.43.