AMD/ATI DirectX 10 Series: RADEON HD 2900 XT (R600)

Part 3. Game benchmarks

Testbed configuration

Intel Core2 Duo (775 Socket) based machine
- Intel Core2 Duo Extreme X6800 (2930 MHz) (L2=4096K) CPU
- EVGA nForce 680i SLI motherboard on NVIDIA nForce 680i
- 2GB of Corsair DDR2 SDRAM 1142MHz (CAS (tCL)=5; RAS to CAS delay (tRCD)=5; Row Precharge (tRP)=5; tRAS=15)
- WD Caviar SE WD1600JD 160GB SATA HDD
Windows XP SP2 DirectX 9.0c; Windows Vista DirectX 10;
Dell 3007WFP 30" monitor
Drivers: ATI 8.37; NVIDIA 158.18/158.19.

VSync disabled.

Test results: performance comparison

Testbed configuration

We used the following benchmarks:

Splinter Cell Chaos Theory v.1.04 (Ubisoft) - DirectX 9.0, Shaders 3.0 (with and without HDR), maximum settings
Call Of Juarez (Techland / Ubisoft) - DirectX 9.0, shaders 3.0 (HDR), (demo, maximum settings. Batch file included.
Call Of Juarez DX10 Benchmark (Techland) - DirectX 10.0, Shaders 4.0,
FarCry 1.4 (beta) (Crytek/UbiSoft), DirectX 9.0, shaders 2.0b/3.0 (with and without HDR), 3 demos from Research (run with -DEVMODE option), Very High settings, HDRRendering=1 for HDR tests
PREY 1.01 (3D Realms Entertainment / Human Head Studios / 2K Games) - OpenGL, Shaders 2.x, demo003 (40MB), maximum quality. Batch file included.
3DMark05 1.20 (FutureMark) - DirectX 9.0, multitexturing, trilinear filtering
Serious Sam II 1.068 (Croteam / 2K Games) - DirectX 9.0, Shaders 3.0 (with and without HDR), batch file, standard demo0002, maximum settings. We thank Voznyuk Valery Mikhailovich AKA Px for writing a batch file for this game.
F.E.A.R. v.1.08 (Multiplayer) (Monolith / Sierra) - DirectX 9.0, Shaders 2.0, maximum settings, Soft Shadows disabled.
Company Of Heroes (Relic Entertainment / THQ) - DirectX 9.0, Shaders 2.0, maximum quality. Batch file. In game invoke graphics settings and click the test button.
3DMark06 1.02 (FutureMark) - DirectX 9.0c, multitexturing, trilinear filtering
RightMark3D (iXBT.com / www.Digit-Life.com) - DirectX, synthetic benchmark, Shaders 2.0/2.0b/3.0, available here.

Test results are available in Excel tables in RAR and ZIP archives

Those of you who are into 3D graphics will understand the charts themselves. And newbies and those who have just begun choosing graphics cards can read our comments.

First of all, you should browse our brief references dedicated to modern graphics card series and underlying GPUs. Note clock rates, shader support, and pipeline architecture. You should understand that 8800 series made the "pipeline" term obsolete. Now there are stream processors handling games' shader instructions, texture processors handling pixels (texturing, filtering). In other words, everything's separated. And only after texture operations are complete data is written to memory buffer, like always. Before it has been like that: take a triangle, process it with a shader instruction set, texture it, filter it, and send it along. This has come to an end. Now drivers together with special thread processors decide which stream processor will do what and when. The same can be said about texture processors. The R600 family operates in a similar way.

Reference of ATI RADEON X1300-1600-1800-1900 series

Reference of NVIDIA GeForce 7300-7600-7800-7900-8800 series

Secondly, you can browse our 3D graphics section for the basics of 3D and some new product references. There are only two companies that release graphics processing units: ATI (recently merged with AMD and obtained its name) and NVIDIA. Therefore all information is generally separated in two respective parts. You can also read our monthly 3Digest, which sums up all graphics card performance charts.

Thirdly, we'll see how the examined graphics card performs in benchmarks. We conducted two test sets: a complete one under Windows XP (since today it's the most popular gaming OS) and a brief one under Windows Vista (today this OS is still raw, so you should use it just to examine DirectX 10).

Now some words about overclocking. While the drivers prevent 2900 XT from overclocking without an 8-pin power connection, there's an utility from AMD that enables to bypass this. Thus we have also obtained results at 850/2000 MHz clock rates.

Remember that today many games are bottlenecked by CPUs and systems in general, so even the most powerful graphics card might not be able to show its potential in AA 4x and AF 16x modes. Therefore, examining previous top-performing products, we have added a new HQ (High Quality) mode, which stands for:

ATI RADEON X1xxx: AA 6x + adaptive AA + AF 16x High Quality
ATI RADEON X2xxx: AA 8x + adaptive AA + AF 16x.
NVIDIA GeForce 7xxx/8xxx: AA 8x + TAA (MS with gamma correction) + AF 16x

Above are screenshots that demonstrate the HQ quality.

Note that R600 has AF on the level of High Quality AF produced by the X1xxx series and it's nearly identical to AF of NVIDIA GeForce 8800 series. So, there's no sense in making some dedicated screenshots for it.

And another matter: we will compare HD 2900 XT only with GeForce 8800 GTS, according to the market positioning of these products.

Testing under Windows XP

FarCry, Research (No HDR)

Test results: FarCry Research (No HDR)

Since without the HQ mode this game is only bottlenecked by CPU, we'll only compare the HQ mode in which 2900 XT demonstrated some very good results. Especially at 1600x1200 providing good gameplay.

Overclocking boosted performance only by 15%. Perhaps, CPU bottlenecked that. Besides, shader unit clock rate might have increased in the same proportion as well. (We remember how abruptly the performance of 8800 series reacted to a fast increase of shader unit clock rate during overclocking).

FarCry, Research (HDR)

Test results: FarCry Research (HDR)

The load increased in this test, so even without HQ mode we witness the superiority of R600. Though, along with a strange and significant defeat of X1950 XTX. Perhaps, we should blame drivers.

F.E.A.R.

Test results: F.E.A.R.

In this game that puts some very heavy load on graphics cards R600 is clearly the winner. But why its performance drops in AA+AF mode? It would seem that 512-bit bus should have provided AA nearly "free of charge".

AMD officially said that was because of raw drivers. Another reason was the disbalance between the power of shader unit and that of texture unit that ATI is still implementing. 16 texture processors is just not enough for such a his large amount of shader processors.

Splinter Cell Chaos Theory (No HDR)

Test results: SCCT (No HDR)

A similar picture. But since this game puts even heavier load on shaders, the novelty showed definite advantage even with AA+AF.

Splinter Cell Chaos Theory (HDR)

Test results: SCCT (HDR)

The same picture again. In the HQ mode the new product outperforms 8800 GTS nearly twofold!

Call Of Juarez

Test results: CoJ

In this game, that had already become a problem for NVIDIA and its 8800 series, R600 stumbled upon some issues as well. First of all, it lost completely. But it's early to judge, because there are some serious quality problems related to shadows and artefacts. Let's wait until they fix these in new drivers.

Company Of Heroes

Test results: CoH

And this test is just the opposite. In it R600 outperforms its rival the most! But overclocking boosted performance by just 11%.

Serious Sam II (No HDR)

Test results: SS2 (No HDR)

And still the reason of 2900 XT demostrating lower performance with AA+AF must be the drivers. So, there might be a cure. On this example we can see that AA+AF load gives a greater advantage to R600.

Serious Sam II (HDR)

Test results: SS2 (HDR)

A similar picture. Though we remember this game is very critical to performance of ROPs and texture processors.

Prey

Test results: Prey

And again the new graphics card lost some performance with AA+AF. Drivers again? Or still an architectural disbalance? I guess we'll get the answer after new drivers are released.

The new product performs naturally good in synthetic tests. It's not a secret that performance of any new graphics card is honed in benchmarks like 3DMark nearly in the first place.