AMD's open GPU documentation site

Where did you read 'without any hardware support' there? Both XAA and EXA would be totally
pointless in that case of course. The benchmarks where they're slower than ShadowFB exercise
paths that the driver doesn't accelerate yet (mostly related to the RENDER extension - that
can be fixed now that AMD released 3D documentation), so the acceleration architectures have
to pay a price for trying to accelerate things instead of just doing them with the CPU in
system memory in the first place. It shouldn't be hard to find other benchmarks where hardware
acceleration is much faster than ShadowFB. Which of the options works best in practice depends
on the usage pattern, so everybody has to find out for themselves.