 | |||
|
|
|
|
|
 |
3DMark 06 Overview and Performance January 18, 2006 Alan /.effect Dang, Pfucata.com |
|
 |
|
| |
Summary: FutureMark's latest benchmark ushers in Shader Model 3.0 gaming like never before, with the addition of high dynamic range lighting in 2 of the benchmark's 4 game tests. 3DMark 06 also begins to address physics processing. Join us as Alan goes over the changes that have been introduced in this new benchmark, as well as evaluating the performance of 13 different ATI and NVIDIA GPUs. Who comes out on top: ATI or NVIDIA? Only one way to find out!
 Introduction | Page:: ( 1 / 11 ) |
[image]
| <% print_image("01"); %> |
With the exception of 3DMark 2003, which generated some heated debate between journalists, manufacturers, and FutureMark due to the manner in which the shaders were written, FutureMark has otherwise had a proven track record of developing benchmarks that are unanimously considered to be fair, objective, and useful ever since Final Reality
3DMark 05 was a superb benchmark, bringing sophisticated but realistic shaders and tools to help end-users and journalists evaluate the performance of various graphics cards or system tweaks. With 3DMark 06, FutureMark has brought high dynamic range lighting (HDR).
We’ve had a chance to play with a pre-release version of 3DMark 06 and take a look at the benchmark design of 3DMark 06. Spoilers for the game demo are present.
| Proxycon returns in 3DMark 06 | Page:: ( 2 / 11 ) |
Test 1: Return to Proxycon
Return to Proxycon is once again a part of the 3DMark benchmark as a Shader Model 2.0 test. To maintain its effectiveness as a benchmark tool with today’s faster GPUs, FutureMark has increased the resolution, the number of objects and lights. Additionally, all of the tests in 3DMark 06 use a new type of depth shadow map called Cascaded Shadow Maps in which there is are five 2048x2048 shadows maps used. One of the lights in Return to Proxycon uses the CSM approach. Traditional point lights with 1024x1024x6 cube depth maps/shadow maps are used for everything else. There are a total of 26 light sources.
[image]
| <% print_image("02"); %> | <% print_image("03"); %> | <% print_image("04"); %> |
| <% print_image("05"); %> | <% print_image("06"); %> | <% print_image("07"); %> |
[image]
| <% print_image("08"); %> | <% print_image("09"); %> | <% print_image("10"); %> |
| The return of the fireflies | Page:: ( 3 / 11 ) |
Test 2: Firefly Forest
The forest scene from 3DMark 05 also returns in 3DMark06. Like Return to Proxycon, FutureMark has added more stress to the graphics card through incremental changes. You now have two “fireflies” instead of just one. The moonlight is done using CSM, while the fireflies are masked point lights with a 1024x1024x6 cube depth map/hardware shadow map.
[image]
| <% print_image("11"); %> | <% print_image("12"); %> | <% print_image("13"); %> |
| <% print_image("14"); %> | <% print_image("15"); %> | <% print_image("16"); %> |
[image]
| <% print_image("17"); %> | <% print_image("18"); %> | <% print_image("19"); %> |
Essentially, for the first two tests, FutureMark has take the 3DMark 05 scenes and made some incremental changes to toughen them up. The next two tests have more changes.
| 3DMark 06’s SM 3.0/HDR Test | Page:: ( 4 / 11 ) |
Shader Model 3.0/High Dynamic Range Lighting Test: Game Test 3 Canyon Flight
Although the Canyon Flight test was seen in 3DMark 2005, the Canyon Flight test in 3DMark 06 is substantially different. Most obvious is the introduction of SM 3.0 technology. This test now incorporates high dynamic range rendering, and exceptionally smooth shadows. These new shadows are created using a 16 sample randomly rotated grid. This isn’t a technique currently used by developers due to the computational intensity, but it certainly produces a very nice looking set of shadows.
[image]
| <% print_image("20"); %> | <% print_image("21"); %> | <% print_image("22"); %> |
| <% print_image("23"); %> | <% print_image("24"); %> | <% print_image("25"); %> |
| <% print_image("26"); %> |
[image]
| <% print_image("27"); %> | <% print_image("28"); %> | <% print_image("29"); %> |
| Game Test 4/CPU Test | Page:: ( 5 / 11 ) |
The second SM 3.0/HDR Test: Game Test 4 Deep Freeze
This test is new to 3DMark 06 and represents an abandoned Antarctic base. Unfortunately, the screenshots do not do justice to the actual test. In this test, a setting sun casts dynamic shadows over a wide range of objects. The lighting is produced by blending two cubemaps: one for diffuse and one for specular light. HDR lighting is also used substantially.
[image]
| <% print_image("30"); %> | <% print_image("31"); %> | <% print_image("32"); %> |
| <% print_image("33"); %> | <% print_image("34"); %> | <% print_image("35"); %> |
| <% print_image("36"); %> | <% print_image("37"); %> |
New CPU Tests
Whereas previous versions of 3DMark 05 used software rendering to test CPU performance, 3DMark 06 takes a more realistic approach of challenging the CPU with exceptionally tough physics and path-finding algorithms. These are likely to be accelerated with PPU in the future.
[image]
| <% print_image("38"); %> |
In 3DMark 06, the final composite score will include CPU performance as part of the test.
| Test Systems | Page:: ( 6 / 11 ) |
System Setup
AMD Athlon 64 FX-60 (for high-end cards)
AMD Athlon 64 3500+ (for mainstream cards)
ASUS A8N32-SLI Deluxe (for high-end cards)
ASUS A8N-SLI Deluxe (for mainstream cards)
2GB OCZ DDR400 SDRAM
High-end Cards
ATI Radeon X1800 XL 256MB
ATI Radeon X1800 XT 512MB
CATALYST 5.13
NVIDIA GeForce 7800 GTX 512MB
NVIDIA GeForce 7800 GTX
NVIDIA GeForce 7800 GT
Driver version ForceWare 81.98
Mainstream Cards
ATI Radeon X1600 XT 256MB
ATI Radeon X800 XL 256MB
ATI Radeon X800 256MB (12 pipelines at 400/350)
Sapphire Radeon X800 GTO2 Limited Edition (Running with 16 pipelines at 400/490)
CATALYST 5.13
NVIDIA GeForce 6800 GS 256MB
NVIDIA GeForce 6800 256MB
NVIDIA GeForce 6600 GT 128MB
XFX GeForce 6600 DDR2 256MB
Driver version ForceWare 81.98
250GB Maxtor Hard Drive Maxline III SATA Hard Drive w/16MB Cache
Windows XP Professional SP1
DirectX 9.0c
Benchmarks
3DMark 06 1.02
| Mainstream Cards 3DMark 06 | Page:: ( 7 / 11 ) |
Shader Model 2.0 Overall Score
Game Test 1
Game Test 2
Shader Model 3.0/HDR Overall Score
Game Test 3
Game Test 4
Notes
First of all, it’s important to note that we’re not presenting the overall 3DMark 06 score in any of today’s graphs. As we mentioned previously, the overall score includes not only the 3D performance of 3DMark 06’s various game tests, but also the CPU’s performance in 2 CPU tests as well. Considering 3DMark’s usage as a graphics rather than CPU or system benchmark, we question why all of a sudden FutureMark has decided to integrate the CPU tests into the overall score, when they didn’t do this previously in 3DMark 03 or 05. 3DMark’s overall score is used frequently among IHVs such as ATI and NVIDIA in marketing materials to demonstrate the performance of their latest GPUs, it will be interesting to see if the overall score continues to be used in this way.
Anyway, on to the results.
The GeForce 6800 GS takes top performance in 3DMark 06's SM 2.0 score as expected, but the surprise is how well the X1600 XT does in comparison to the X800 XL. Given the neck and neck performance between the two cards, game test 1 and 2 are probably compute-limited rather than memory bandwidth limited tests. The X1600 XT also does surprisingly well with HDR as it outperforms the 6800GS at both 800x600 and 1024x768. Once you move to 1280x1024, however, you can see that the GeForce 6800GS is faster than the X1600 XT in HDR Test 2. This is probably the point where memory bandwidth is likely coming into play.
| High-end cards in 3DMark 06 | Page:: ( 8 / 11 ) |
Shader Model 2.0 Overall Score
Game Test 1
Game Test 2
Shader Model 3.0/HDR Overall Score
Game Test 3
Game Test 4
Notes
Although the mainstream GPUs struggle with 3DMark 06, the high-end GPUs do much better. A 7800GTX at the GPU-intensive screen res of 1600x1200 is faster than a 6800GS at 800x600. Under these tests, the strength of the GeForce 7's shader performance is fairly clear.
As a disclaimer, we don't truly know the benchmark well enough yet -- but these are our initial impressions.
| 3DMark 06 2xAA/8xAF | Page:: ( 9 / 11 ) |
Shader Model 2.0 Overall Score
Game Test 1
Game Test 2
Shader Model 3.0/HDR Overall Score
Game Test 3
Game Test 4
| 3DMark 06 2xAA/8xAF | Page:: ( 10 / 11 ) |
Shader Model 2.0 Overall Score
Game Test 1
Game Test 2
Shader Model 3.0/HDR Overall Score
Game Test 3
Game Test 4
Notes
When we move to tests with anti-aliasing turned on, things change substantially. Although the GeForce 7800GTX is one of the fastest performing GPUs in real-world games it cannot do HDR and anti-aliasing simultaneously. ATI can. As a result, NVIDIA puts in a zero score for the HDR/SM3.0 with anti-aliasing tests.
Looking at the tests more carefully, we can see that the X1800 XL has very similar scores for HDR test 1 and 2 at 1024x768. In fact, with anti-aliasing turned off, the second HDR is faster. When the level of anti-aliasing increases from 2xAA to 4xAA, the difference between HDR Test 1 and 2 grows. HDR Test 1 ends up running about 50% faster. This again supports the idea that HDR 2 is more memory bandwidth limited. Once you move to 1600x1200 with 2xAA or 1280x1024 with 4xAA, memory bandwidth is a problem for both tests.
| Conclusion | Page:: ( 11 / 11 ) |
Of course, with every new benchmark release is the potential for controversy. The integration of CPU performance into the overall 3DMark score is one that is most likely to cause some debate in the future. While 3DMark 06 does not support the Ageia PPU, it does use the PhysX engine for the CPU tests. We expect future revisions of 3DMark 06 to incorporate hardware accelerated physics, and it will be a challenge to ensure that the benchmark appropriately weights the contribution of the GPU and PPU.
Overall we feel 3DMark 06’s release is a “good thing” for the industry because we now have another standardized tool in our benchmarking suite that we can use for evaluating the latest and greatest computer hardware, but clearly it will take more than just looking at a single number to make meaningful interpretations. The Radeon X1600 XT’s performance for instance, absolutely stunned us. While previous tests we’ve run with the card in shipping games have indicated it’s an average performer, our results with 3DMark 06 today suggest otherwise.
We’ve only begun benchmarking with the latest build of 3DMark in the past 18 hours, so we haven’t had enough time to come to any firm conclusions on how the X1600 performs so strongly, while it can be said that the X1800 XT disappoints, often finishing behind the GeForce 7800 GTX 256MB in many tests, even though real-world testing with today’s latest games indicates the opposite. We’re sure conspiracy theories will begin popping up in various forums shortly.
In the meantime, we’ll continue to focus on testing the latest graphics hardware with real shipping titles, and use 3DMark 06 only on occasion from time to time. Perhaps with 3DMark 06’s new emphasis on CPU performance, 3DMark 06 would be better served as a system benchmark used to evaluate a given PC’s performance as a gaming platform rather than solely as a 3D benchmark, in some ways similar to the way 3DMark 2001 was used by enthusiasts to judge the success of their overclocking endeavors. We’ll just have to wait and see.