AMD Radeon HD 7970 Launch Review

Product: AMD Radeon HD 7970 Video Card
Company: AMD
Author: James Prior
Editor: Charles Oliver
Date: December 24th, 2011

Image Quality Updates

The AMD Radeon HD 7970 doesn't introduce any new filtering methods but, like Cayman, does support EQAA, in addition to MLAA and the traditional standard & edge detect modes. Partially Resident Texture support can make for better hi-resolution images, as seen in AMD's Leo demo, made the by the creators of AMD's Toy Shop demo. Watching Leo at the AMD SI tech day was nothing short of jaw dropping. Sitting in the front row and watching a pre-rendered quality animation being real time rendered on three 60" HDTVs was an incredible experience. The demo uses tessellation and new indirect light methods processed by compute, as well as PRTs, to deliver a very high quality image. The Leo demo will be available sometime later, once the usual legal bumpf has been taken care of.

Anistropic Filtering

Following the discovery and coverage of the Anistropic Filtering (AF) shimmer (see our article here for more information), AMD found ways to improve their quality in that area. Several teams collaborated internally to investigate, analyze and determine the cause of the problem, with anecdotal reports of a 'crazy Russian programmer' ripping the AF kernel apart and rebuilding it to provide a better implementation of Anisotropic Filtering.

AMD claims to have improved their image quality in the megastress test that is's filter tester moving grass texture, and retained their angle independence for great texture quality. Testing this ourselves we see that there is a significant improvement in image quality, a lot less shimmer is discernible.

Highlighted area shows where excessive shimmer occurs

However, it is not gone or matching the application ALU perfect reference. Shown above is the area where slight shimmer can be observed, far less than seen in Cayman or Cypress but still there. Adjusting LoD through negative ranges shows that there is a lot of work being done in a large range, highlighting where AMD's filter is interpolating (see picture below). This isn't matched by the reference render although shimmer is noted here as well. This adjustment is a test case that should never be seen in any game and is done for illustrative purposes only, not as demonstration or proof or wrongdoing or poor quality.

This area is subject to AMD's processing, when artificially highlighted

Testing the angle dependence confirms that is as claimed, with no variance in texture filtering by angle.

In specific games with high noise textures used on floor, wall or ceiling objects, shimmer is not observed easily but is one of those items that once you see you can't unsee it. AMD has more work to do on their Anisotropic Filtering, and gladly admits it - they found enough issues and fixes to put in to the next three generations of GPUs, so hopefully along that path they'll get the shimmer issue resolved and perhaps look at new AF techniques like High Quality Eliptical Texture filtering using an Elliptical Weighted Average filter as presented at SIGGRAPH this year.

Testing Eyefinity 2.0

Using Eyefinity in a 3x1 Portrait configuration, we still found issues like vertical screen tearing near bezel boundaries. It's slight but when moving windows between panels on the display, you might notice an offset or tear as the windows moves through the area near the bezels. It may be this is due to difference in inputs, as we had two DP connected displays and one DVI. When the multi-stream display hubs appear in Summer 2012, this should stop being as much of an issue. AMD demonstrated working prototype hubs that did the job well, despite the long bring-up time things are looking up. We did not see any mouse cursor corruption or disappearance, but did have several display port link failures. AMD indicates these are usually traced to bad connections between mini-DP to full-DP adapters, and dust, however cleaning and reseating connections didn't give us much difference in link failure rate. Some games just don't want to fire up in Eyefinity under Crossfire, like Supreme Commander 2. Running games in Eyefinity for the first time after playing them in 1080 was sometimes problematic, with even titles like BattleField 3 requiring the user to set a lower than native res and non-full screen mode to get to be able to see the full menu options as the scaling and position of the screen image wasn't fully visible. After that, setting full Eyefinity resolution and full screen mode worked fine.

Performance scaling


To test Anti-aliasing performance hit, we used DiRT 3 and various AA modes to see the % of no-AA performance. Tests were performed at 1920x1080 with Ultra preset, benchmark using Finland race. Normalized results are below.

The results indicate 2xAA is around 4% performance hit, ~7% for 4xAA, ~15% for 8xAA and ~35% for MLAA.

Anisotropic Filtering

To test Anisotropic filtering performance hit, we used Aliens Vs. Predator benchmark and varied AF with all other settings on high, no AA. Normalized results are below.

The results indicate that 2xAF is a 1% performance hit, 4xAF is a 3% hit, 8xAF is a 5% hit and 16xAF is a 7% hit.


To test tessellation performance, we used Crysis 2 and the Adrenaline benchmark tool Times Square level. With Performance setting and no AA/no Edge AA, we varied AMD Tessellation cap level to see if it affected performance, using application controlled tessellation as the baseline. Normalized results below:

The results show that capping tessellation to 4x increase performance by ~3%. A cap of 8x increases performance by just under 5%. 16x tessellation cap improves performance by ~1.5%, as does a 32x cap, where a cap of 64x (the maximum tessellation factor) provides identical performance to the app controlled benchmark.