Part 4: Rendering Quality, New Anti-Aliasing and Anisotropic Filtering Modes

TABLE OF CONTENTS

1. Part 1. Theory and architecture
2. Part 2: Practical examination
   • Video cards' features
   • Testbed configurations, benchmarks
   • Synthetic test results
3. Part 3: Game test results (performance)
4. Part 4: Render quality comparison in games
   • Synthetics: D3DAFTester and DX9FSAAViewer
   • Far Cry
   • Doom 3
   • Half-Life 2
   • Splinter Cell: Chaos Theory
   • F.E.A.R. SP Demo
   • Serious Sam 2 Demo
   • Conclusions on Part IV

So, we have come to know the architecture of Radeon X1x00 chips, synthetic test results and performance, demonstrated by the new video cards from ATI in games. The fourth part of our review will cover the last aspect: render quality of Radeon X1x00 video cards compared to the previous Radeon generation and the immediate competitors from NVIDIA.

At first, we shall take freedom to get back to Part I and recall the new image quality features, brought by 5x0 series chips. All the new features can be divided into two groups:

• those we can use as soon as we install the card (new texture filtering and antialiasing algorithms, support for functions, which were previously unavailable in Radeon cards)

• those (let's hope for the best) to be used only in new games, which are not yet out.

The key innovations:

1. DirectX 9 Shader Model 3.0 support (there is a nasty, but not fatal reservation: no support for vertex texture fetch).

2. FP16 format support: full support for reading FP16 data into a frame buffer (including any blending operations and even MSAA). FP16 texture compression (however, hardware FP16 texture filtering is unavailable), including 3Dc.

3. New RGBA (10:10:10:2) integer data type in a frame buffer for higher quality rendering without FP16.

4. New high-quality algorithm for anisotropic filtering (a user is given a choice between a faster or higher-quality anisotropy options), improved trilinear filtering.

5. Adaptive antialiasing algorithm, which allows the chip switching dynamically from multisampling (MSAA) to supersampling (SSAA) with "transparent" textures — it's a counterpart of Transparency AA implemented in GeForce 7800 chips.

The last two points obviously belong to the first group of innovations. The third point belongs to the second group (an application must be able to use the new data format), and the first two points are somewhere in between. That's because, on the one hand, games with SM3 support already exist. But we haven't yet encountered games, where SM3 support affects the render quality.

FP16 rendering is another story: none of the games using FP16 rendering on NV4x+ chips worked correctly with our RX1800XT video card in FP16 mode (that is with enabled HDR). What is it? A bug in the driver or games should be "taught to recognize" new ATI chips? We don't have an answer. The only thing for us to do in this situation is to await developments, new driver versions, and game patches.

Testbed

• AMD Athlon 64 3700+ (2.2 GHz, 1 MB L2)
• EPoX 9NPA+ SLI (nForce 4 SLI, PCI Express x16/x8+x8)
• 1024 MB of PC3200 RAM
• ATI Radeon X1800 XT, 625/750, 512 MB, Catalyst 5.10 Beta (8.173)
• ATI Radeon X850 Pro, 505/530, 256 MB, Catalyst 5.9
• NVIDIA GeForce 7800 GTX, 430/470/600, 256MB, ForceWare 81.84 Beta

Synthetics: D3DAFTester and DX9FSAAViewer

But before we proceed to the quality comparison in games, let's start two tiny, but damn useful utilities: D3DAFTester (by Demirug) and DX9FSAAViewer (by Colourless). These utilities help evaluate changes in texture filtering and antialiasing quality with your own eyes using special scenes.

D3DAFTester

• We disabled Catalyst A.I before taking screenshots on Radeon video cards.
• The GeForce 7800 operated in High Quality mode.
• The other driver settings were by default.

We had hardly started our tests, when we ran across an unpleasant peculiarity of activating High Quality anisotropy on the RX1800 via the standard Catalyst Control Center. The fact is that in the given version the CCC checkbox, which activates HQ anisotropy, is inactive until you force any degree of anisotropic filtering, that is until you deselect the checkbox that passes anisotropic filtering control to an application. Such CCC behavior was unacceptable for us.

Forced anisotropic filtering is not always similar to anisotropic filtering enabled by an application. As a rule, an application enables anisotropic filtering only for those textures, which really need it, while forcing anisotropy makes a chip do anisotropic filtering for absolutely all textures in a frame. This difference may result in performance as well as in image quality differences.

Fortunately, after a number of experiments and consultations with colleagues, we found out that if you selected the High Quality AF checkbox and then selected the Let the application decide checkbox (HQ AF control checkbox would become inactive again), the application would gain HQ AF control! We hope that such an obvious blunder in CCC configuration mechanics will be corrected in future Control Center versions. As for now, in order to activate HQ AF and then control it from applications, we have to follow the above mentioned "voodoo staff"...

	Radeon X1800 HQ	Radeon X1800
1x
2x
4x
8x
16x

	Radeon X850	GeForce 7800
1x
2x
4x
8x
16x

First of all, let's compare images from the first table row: they help evaluate the quality of trilinear filtering of three video chips. Comparing the usual trilinear filtering by the RX1800 with filtering by RX850 and GF7800 shows that the linear interpolation precision for trilinear filtering on the R520 is increased versus the R480. But it's still lower than in NV4x/G70 chips. To do justice we should note that you will see no difference in trilinear filtering quality in real games even between the RX850 and the GF7800.

Further comparison of screenshots shows that the algorithm of "regular" filtering in two Radeon cards is absolutely the same, while the HQ mode in the R520 demonstrates significant quality improvements in pixel LOD algorithms: there are no pronounced radial lines, typical of the "regular" filtering mode.

The 2x mode reveals an interesting peculiarity: only in this anisotropic filtering mode the anisotropy quality of the G70 is similar to the quality of the new HQ anisotropy in the R520. What concerns the regular filtering mode, it's the same in Radeon cards and is inferior to HQ anisotropy in the R520 as well as to the 2x anisotropy in the G70.

In 4x and 8x modes, all cards start lagging behind noticeably from the HQ anisotropy of the R520. By the way, we have already seen this "clover" somewhere... Ah, that's practically a copy of the NV2x anisotropy, still considered a model of anisotropy by many users! According to performance tests, it's very fast — bravo, ATI! Such anisotropic filtering is praiseworthy!

The G70 in 16x mode and Radeon chips in regular anisotropy mode are quite predictable, but something strange happens with the R520 in HQ anisotropy mode: the "clover" turns into the "rectangle", while MIP levels become badly "mixed". However, even in this case the HQ anisotropy in the R520 is still much better than the anisotropy in its immediate competitor.

But before singing the praises, we should test the new anisotropy algorithm of the R520 in games — that's what we shall do on the next pages. And now let's have a look at antialiasing modes.

DX9FSAAViewer

• When we were taking screenshots, all drivers were configured by default.
• We changed only antialiasing modes.

Note: Due to a nasty bug in ForceWare 81.84 beta, we didn't enable Gamma Correct Antialiasing for the GeForce 7800 GTX. When this option is enabled, the GF7800 starts using OG SSAA (GF256 and GF2GTS used to work in this mode) instead of RG MSAA, which has a disastrous effect on the performance as well as on the antialiasing quality. This reservation applies not only to the DX9FSAAViewer test, but also to all game tests in the fourth part of our review.

By the way, do you remember that AA gamma correction is always active in Radeon cards and you cannot disable it.

+TA = + Transparency Antialiasing — as NVIDIA was actually the first to integrate this function into its chips, we shall use the term offered by this company; especially as it's more correct than "Adaptive Antialiasing", proposed by ATI.

Click an image to open the full DX9FSAAViewer scene: here you can see not only the arrangement of sub-samples, but also the antialiasing quality of inclined lines. Click! :-)

	Radeon X1800	GeForce 7800
1x
2x
2x+TA
4x
4x+TA
6x/8xS
6x/8xS+TA

	Radeon X850
1x
2x
2x+TA	-
4x
4x+TA	-
6x/8xS
6x/8xS+TA	-

You can see well that usual AA modes in the Radeon X1800 remain as they used to be in the Radeon 9700. Antialiasing of transparent textures copies the standard modes, but for texture pixels rather than triangle borders.

What concerns quality, everything remains unchanged:

• 2x mode is identically bad on all video cards. It's just a temporary measure to increase performance;
• 4x mode is absolutely the same on both Radeon cards (of course, except for the X1800 being able to smooth transparent textures) and approximately comparable between Radeon and GeForce, even though GeForce cards don't use gamma correction;
• 6x mode in Radeon cards is an absolute quality/performance champion compared to the 8xS mode in GeForce — we look forward to NVIDIA's implementation of MSAA with more than 4 sub-samples...
• 8xS mode in GeForce cards is an obvious survival of times past: it's of lower quality (and performance!) than the 6x mode in Radeon cards; after the appearance of Transparency AA, it lost its ace of trumps — antialiasing of transparent textures.

But enough of synthetics, let's proceed to real games. Be warned that all the next pages contain a lot of high-quality graphics - they will take much time to load, if you have a slow Internet connection.

Read on: Far Cry

Write a comment below. No registration needed!