Sapphire Radeon HD 6950 Dirt 3 Edition Video Card Review

Company: Sapphire
Author: James Prior
Editor: Charles Oliver
Date: August 15th, 2011

Performance Analysis


Average frames per second (FPS) results from the Adrenaline Racing benchmark tool, Finland stage. Results of 1920x1080 in DirectX 11 with ultra preset and varying levels and types of AA.

First we have single, dual, and triple CrossfireX performance, using stock clocks. In game AA was varied between 0xAA, 4xAA and 8xAA with MSAA/Box, EQAA and EDAA anti-aliasing modes selected in Catalyst Pro Control Center (CPCC). Mouse over data points to see details.

As you can see the performance hit for 4xAA and 8xAA is pretty constant and quite low - only 7FPS (~11%) hit for a single 6950 at 1080p. The hit decreases as more cards are added, but only slightly (~9% for 2x and 3x GPUs).

Crossfire scaling isn't bad at all, 75% increase in performance adding 1 GPU and adding two GPUs gives 167% increase, which is a 50% improvement moving from two to three cards. Best of all, the 8xMSAA performance of a single card is near the 60FPS average mark, delivering a fun and enjoyable high quality visual experience. Two GPUs can't quite hit the magic 120FPS average but instead deliver over 90 at 8xMSAA; three GPUs is enough for 120Hz vSync, easily, with 8xMSAA.

EQAA starts to make the Cayman Pro core work harder, pushing a single card down towards 45FPS average - still playable and fun, but not optimal for competitive play. Two cards are in the 95-85FPS average, depending on AA level, plenty of performance, and three GPUs is still well over 120FPS (around the 130FPS mark with 8xEQAA). Enhancing the image using EQAA costs between 9 to 9.5% against MSAA.

EDAA drops the performance the most, although the performance drop from 4xEDAA to 8xEDAA (12xMSAA to 24xMSAA) is smaller than the EQAA one. Applying EDAA is a ~20-25% performance hit vs. no AA, and 17% drop vs. normal 4xMSAA/8xMSAA.

Next under CrossfireX we used CPCC to force 16xMSAA and 16XEQAA, with 16xAF added as well; and Morphological Anti-Aliasing performance, with 2xMSAA in game set and 16xAF.

Forcing 16xMSAA and 16xEQAA doesn't have much effect with two cards, showing that the override is ineffective although it is doing something to the image, just not 16xAA samples worth. The performance drop is likely the Anisotropic Filtering hit, which is quite significant with 3 GPUs for some reason; maybe it's more than AF? The resultant image looks somewhat filtered for aliasing, indicating perhaps it's an adaptive or selective application of AA occruring.

Morphological AA has the biggest impact on performance and image quality, a 40% performance hit vs. no AA for a single card, dropping performance into the mid 30FPS range. Two cards brings performance back up to almost single card no AA levels, close to the magic 60FPS marker, but a bigger hit in performance relative to two-GPU no AA - 46% down. With three GPUs we're down 47% vs. no AA, but performing in the 80FPS range. MLAA works quite well but not enough to decide to use it over in-game 8xMSAA and turning on EQAA or EDAA, if you've got the performance to spare - and this is really not where AMD targets MLAA, it is supposed to offer better image filtering options for games that lack all the post processing options we expect.

The 15% engine and 6% memory overclocking of the Sapphire HD 6950 DiRT3 made almost no difference in performance, less than 1/2FPS on in the average results.

Crysis 2

Next up is Crysis 2, benchmarked using the Times Square timedemo in the Adrenaline benchmark tool. DirectX 11 API with high resolution textures, edge AA off and varying levels of MSAA were benched. We also benched with AMD CPCC Tessellation factor set to cap at 16x subdivisions (denoted as Tf 16), to see if Tessellation has an effect on performance in this game. We also ran the tests with no in-game AA, but MLAA and CPCC forced 16xAF enabled. Mouse over the data points for more detail.

The framerates are in the slideshow territory for the single GPU, although managing an average FPS in the 30's, it's not gonna feel fluid and smooth enough when thegoing gets tough. In-game AA is pretty cheap, 9% to add 2xMSAA or 4xMSAA. MLAA is a 33% hit, looks pretty decent but messes up some of the lighting flares and diffusions.

For two GPUs the performance is much better, a very nice 82-85% increase depending on AA level. MLAA is more expensive with 2 cards, now costing near 40%, and dropping average FPS away from the 60FPS sweet spot the two GPU cards inhabited. Three GPUs performs 1-2FPS lower than 2 GPUs. Most obviously, three GPUs is not worth it for Crysis 2 at this time - less performance than with 2 GPUs, Ugh.

The results of the tessellation cap are interesting, it certainly helps performance but not enough to be really meaningful by only adding 2-4FPS at most to any individual result. There's no in-game discernable difference in quality, unless you stop playing to line up on a wall or object and hunt for the missing poly's. Once you get familiar enough with the game you might see it more and more but as you have to positively override and select this tessellation cap, to get 2-4FPS you might just not bother.

Because of the single card performance results, we decided to repeat the tests at 1680x1050 and 1280x720, to see if playing Crysis 2 on a single GPU would be a viable goal:

1680x1050 works well for a single card, hovering around the 45FPS average that's usually pretty fun to play. Again, the same 1-2FPS boost in performance, by capping tessellation factor. MSAA is a ~9% drop, as before, and MLAA is a 34% hit from no AA.

2 GPU scaling improves, now in the 85-89% range depending on AA level, and a steady ~8% hit for adding MSAA. Now we're up above the 75FPS average mark, getting to where we could think about adding more in-game IQ or online competitive play. MLAA + 16xAF is a 38% hit vs. no AA, still. 3 GPU scaling is a steady 3-5FPS worse than 2 GPUs.

Now we're down at the console target resolution, a single GPU can keep us above 60FPS average even with 4xAA, nice. Tessellation capping get's us an extra FPS or so, a interesting phenomena that doesn't scale with resolution (and therefore geometry) as you would think if it were a real bottleneck. MLAA is a 30% hit here, and just above the 45FPS marker for an enjoyable casual FPS experience.

2 GPU scaling is between 77% and 80%, down from 1680x1050. Performance is almost up to the 120FPS average mark now, for silky smooth gaming and responsiveness. Adding MSAA is a 4-5% hit, with MLAA +16xAF a 32% hit. 3 GPU performance is way down from 2 GPUs, splitting the difference between a single card and two cards (what smells like porpoise hork?).

Overclocking the HD 6950 was pointless for DiRT 3, but what about the much more challenging Crysis 2?

The 15% engine and 6% memory overclocks (with +20% power tune limit) get us ~10% more performance at all resolutions and AA levels, except for MLAA. The overclock puts the 1280x720 performance around the 70FPS average mark for smooth gameplay, and no AA 1920x1080 is around the 45FPS mark where you might actually enjoy playing it.