At 1920x1200 we’re seeing a roughly 5% across the board performance improvement for both the 7970 and the 7950. Everything except Starcraft II sees at least a marginal improvement here, with Starcraft II being the lone exception due to the previous issues we’ve run into with the 1.5 patch. The 7770 also sees some gains here but they aren’t quite as great as with AMD’s other cards; the average gain is just 4% at 1680x1050, with gains in individual games being shallower on the 7770 than they are on other cards.
Interestingly even on the 7970 the largest gains are at 1920x1200 and not 2560x1600. The latter is the more GPU-limited resolution and that’s where we’d typically expect to see the largest gains, but that’s not what’s happening here. Most likely AMD’s performance improvements are targeting shading/texturing performance rather than ROP/memory performance, in which case the highest resolutions where we’re already more likely to be ROP/memory bound would be the resolutions least likely to benefit. This isn’t necessarily the best outcome since it’s at the highest resolutions that we need the greatest performance, but since most gamers are still on 1920x1080 (even with cards like the 7970) this is admittedly a more useful outcome.
Meanwhile like most major performance drivers, even when performance is up across the board the biggest gains are seen in a handful of games, and Catalyst 12.11 is no exception. Among the games in our test suite, DiRT 3, Shogun 2, and Battlefield 3 see the greatest improvements, with the former two picking up 6-7% each.
But it’s Battlefield 3 that really takes the cake: the performance improvement from Catalyst 12.11 ranges from 13% for the 7770 at 1680 to a whopping 29% for the 7970 at 1920. This makes Catalyst 12.11 a very special driver for AMD – not only are performance improvements over 20% particularly rare, but Battlefield 3 has long been a thorn in AMD’s side. NVIDIA’s hardware has until now always outperformed AMD’s equivalent hardware here, and as BF3 has remained an extremely popular MP game it’s been one of the most important games for high-end video card buyers. In other words it has always been the game AMD could least afford to lose at.
With the 12.11 drivers AMD has completely eradicated their performance defecit in BF3, with the 7970, 7870, and 7770 being made performance competitive with (if not a hair faster than) their respective NVIDIA GTX 600 counterparts in our BF3 benchmark. AMD has told us that the specific performance benefits are map-dependent with our results appearing at the high-end of their guidance, so while not every map will see the same 20%+ performance gains, some of them will while others will be in the 10% range. Much like our overall performance averages, the largest gains are at 1920 with FXAA, while 2560 with FXAA and 1920 with MSAA will see smaller gains, once again hinting that AMD’s optimizations are on the shader/texture side rather than ROP/memory.
For our part we have long theorized that the Frostbite 2 engine’s heavy use of deferred rendering techniques – particularly its massive G-buffer – was the factor that AMD was struggling with. While these results don’t really further validate or invalidate that theory, what is clear is that AMD has fixed their Frostbite 2 performance problem. Given the fact that Frostbite 2 will be used in at least a couple more games, including the AMD Gaming Evolved title Medal of Honor Warfighter, this was an important engine for AMD to finally conquer.
For their part, AMD hasn’t told us much about what it is they’ve done to optimize their drivers. What we do know is that it’s not driver command lists (an optional DX11 feature that NV has supported for some time), so it has to be something else. AMD has briefly mentioned surface mapping and memory mapping optimizations, but it’s not clear what exactly they’ve done there and if those are the only optimizations.
Finally, these seem to be clean optimizations, as image quality has been held constant in most of our games (e.g.: Battlefield 3). The sole exception is Skyrim, and this is something we’re certain is an unrelated bug. On our 7970 we appear to be missing a lighting pass, but only on our 7970. On our 7870 with the same exact settings everything is being rendered correctly. AMD has told us that they haven’t seen this issue in-house (it would admittedly be hard to miss) so this may be some esoteric issue; due to the short preview window and our own testing time constraints AMD hasn’t had time to look into it further.
