So are these 4 CUs being reserved to compensate for Jaguar's potential performance shortfall?
They're not being reserved.
They just have to be used more explicitly in your code. If I'm writing my rendering code, I make library calls, and the library and hardware take care of the rest. The GPU splits the work up per vertex or per group of fragments, or whatever, and schedules the threads and executes them.
These 4 CUs won't operate under that regime. They won't mix in 'automagically' to the work you do through the graphics libraries. You'll have to set up the work explicitly, separately, for those 4 CUs.
The work could be whatever you want that a CU can do - graphics or GPGPU. So they're not being 'reserved' per se, they're not being prescribed for any one particular use. However the motivation for splitting the CUs out of a hardware managed regime is so that if you want to do GPGPU, you don't have to leave the scheduling and context switching up to a hardware balancer that might not do things optimally when mixing graphics and compute tasks.
"I think"