Performance - Ceteris Paribus & DX11
The fight was fair up to now...both combatants used the same weapons, same kit, same everything. But reality isn't fair, and when one opponent has an advantage, he leverages it! As such, in this section we're no longer confining Cypress to DX9/10/10.1, and we're allowing it to use DX11 where available:
We lost track of where the DX11 argument ended up online, but it definitely pays dividends in the scenarios we're looking at. We were careful to ensure that both GPUs pushed out the same pixels (well, as much as possible considering differing APIs and shaders), which is to say we didn't enable any of the extra DX11 features which were available in Stalker or Unigine (we'll tessellate and harden our shadows upon contact later immediately, patience please!). This is also the reason for which Dirt 2 is missing, in spite of it being DX11 as well: it's almost impossible to get equal conditions between the two pathways there.
When allowed to leverage its DX11 advantage, Cypress actually manages to reach the vaunted performance doubling goal that was supposedly established for it during the development cycle...at least when AA is disabled. This is encouraging in the perspective of future DX11 titles, and how they'll treat owners of DX11 hardware. On the other hand, Unigine shows that this is not a free lunch: without developers actually opting for using performance enhancing features/practices, a simple API change won't do much as the guys behind Unigine opted for image quality enhancements exclusively - which nicely brings us into the closing bit of the pure performance enumeration, which is the DX11+"eye-candy" scenario:
We lost track of where the DX11 argument ended up online, but it definitely pays dividends in the scenarios we're looking at. We were careful to ensure that both GPUs pushed out the same pixels (well, as much as possible considering differing APIs and shaders), which is to say we didn't enable any of the extra DX11 features which were available in Stalker or Unigine (we'll tessellate and harden our shadows upon contact later immediately, patience please!). This is also the reason for which Dirt 2 is missing, in spite of it being DX11 as well: it's almost impossible to get equal conditions between the two pathways there.
When allowed to leverage its DX11 advantage, Cypress actually manages to reach the vaunted performance doubling goal that was supposedly established for it during the development cycle...at least when AA is disabled. This is encouraging in the perspective of future DX11 titles, and how they'll treat owners of DX11 hardware. On the other hand, Unigine shows that this is not a free lunch: without developers actually opting for using performance enhancing features/practices, a simple API change won't do much as the guys behind Unigine opted for image quality enhancements exclusively - which nicely brings us into the closing bit of the pure performance enumeration, which is the DX11+"eye-candy" scenario:
Cypress vs RV790 0xAA/0xAF 
| 0xAA/0xAF | 4xAA/16xAF | 8xAA/16xAF | |
| DIRT 2 | 76 | 70 | 69 |
| Stalker:CoP | 47 | 28 | * |
| Unigine Heaven | 38 | 27 | 23 |
frames per second
The old TANSTAAFL rule still holds: added IQ costs. Unigine, which goes probably the furthest in terms of adding immediately obvious features when using DX11, by tessellating like no one ever tessellated before, is also the heaviest hitter with regards to performance costs (note that Stalker prior to the 1.06.02 patch was considerably worse performance wise, and also note that we had to mess with the benchmark program sent out by GSC in order to use the patch - as such, Stalker results aren't comparable with what you may see in other places).
Now that we've taken care of the slightly more boring part, we should move on the the more interesting one: bottleneck hunting.
