One scenario was using multiple FirePro cards in a system with a single quad-core processor for a render farm, where the workloads wouldn’t bog down the system because the CPU was too busy feeding the GPUs with information, or a reverse scenario in which the CPU is left idle when it could be assisting with the workload.
In addition, render farm systems could have different GPUs and use the Mantle API to assign code to specific GPUs that may run through certain types of code better than others. I can see this particular feature being used for Bitcoin mining on Radeon GPUs where a system may have one large, powerful GPU along with several weaker ones.
Another example was using the Oculus Rift. Andersson actually mentions “VR headsets”, but he alluded to the Rift when he mentioned that in a dual-GPU system, one could use each GPU to draw information for the two screens in a VR headset, removing issues with Eyefinity and stuttering, while having the GPUs run in tandem with slightly varying data sets to render a game in 3D, using a shared memory pool and matching up their thread queues to allow things to run smoothly.