Its been 20 months since the release of 3DMark03 and that’s a long time in the graphics industry. Many cards have come and gone as has the 15k barrier and so it was getting time for someone to really push the latest graphics cards to their limits. That company is Futuremark and today the NDA lifts on their latest product, 3DMark05. We at Driverheaven have been lucky enough to have access to the benchmark in advance of its launch and have run Nvidia and ATI’s latest and greatest cards through the benchmark with some rather surprising results. Read on to find out exactly whats new in this version of 3DMark and also to see how the results panned out…
3DMark05, what is it?
You will not be surprised to hear that this new version of 3DMark is a tool that provides accurate and impartial performance figures for 3D graphics performance. In 3Dmark03 Futuremark coded to use DirectX 9 functions in a limited manner due to the availability of components which supported these features. In 3DMark05 Futuremark take advantage of much more DirectX 9 functionality and really push the latest hardware to its limits. Because of this you are going to need a DX9 capable graphics card before you can even run the benchmark…so those of you with Radeon 9200 series and below and Geforce 4 series and below can forget about the 300+mb download.
The reason for supporting DX9 cards only is that 3DMark05 extensively uses (/tests) Shader Model 2 and 3 (inc. 2.0a and 2.0b) by using them for all Vertex and Pixel shader processing. Futuremark believe that future games will offer both Shader Model 2 and 3 and this is why 3DMark05 offers both. Interestingly you can force 2.x on the Geforce 6800 series to compare to your other results (by default the Geforce 6800 uses Shader Model 3).
On the complexity of each scene, Futuremark tell us that 3DMark03 used several hundred thousand polygons per frame on average where as 3DMark05 uses over one million rendered polygons per frame on average. Quite an increase!
What are the tests?
There are 4 sections of tests which we are going to look at today. Game Tests, CPU Tests, feature tests and batch tests.
All of the tests are real time rendering, meaning that they are not a playback of events previously recorded and function much more like a real game than other benchmarks. As far as the engine goes, 3DMark05 uses an all new engine, long gone is the MAX-FX engine used in 3DMark01, in 3Dmark05 the engine, like games, uses the CPU to prepare and optimise the content that will be rendered by your graphics hardware. Additionally the engine builds shaders for each element in HLSL format and then they are configured to best fit the shader profile of the graphics hardware being used. The final point on the engine specifics for now is the shadow system…the shadow system (Perspective shadow maps) uses the CPU and memory quite extensively. This added work on the CPU means that the benchmark is much harder on your system than 3DMark03 and in theory should allow the engine to scale much better with the whole system than 3DMark03 did.
Game Test 1: Return to Proxycon
Features used: Blinn-Phong Reflections, Dynamic/PSM shadows, Per Pixel Specular, Normal Mapped Bump Mapping, High Detail Geometry, bloom effect and directional/point lighting.
RTP is a continuation of the story from 3Dmark03. The engine is designed to work like a FPS game would and contains 2 sections. The first is a large indoor area with several characters involved in a battle and the second is a corridor scene. One of the main differences between the two (other than the draw distance / volume of action onscreen) is the type of lighting used. The main lighting in the large indoor area is directional where as the corridor is mainly point lighting.
Game Test 2: Firefly Forrest
Features used: Real time light scattering and Fog, Dynamic Shadows (2 light Sources), animated high detail geometry, diffuse/diffuse detail, normal/normal detail maps, specular maps, Directional and point lights.
Game test 2 continues the nature scene trend from previous 3DMarks. The scene shows a forest environment with a large volume of independently moving vegetation. The scene is lit as if it were night and a light source (firefly) makes its way around the scene testing the point lighting and perspective shadow maps.
Game Test 3: Canyon Flight
Features used: Light scattering and depth fog (map), multi-layered normal layer bump mapping, reflection and refraction, pixel shaded water with depth fog, directional lighting, Per-Pixel Fresnel
The scene shows an airship moving through a canyon which is half filled with water. There are some very complex goings on in this scene. For example the scene is rendered 6 times to get the correct reflections, refractions and shadows. The water shader is an advanced version of that used in 3Dmark03. The canyon wall shader is also very demanding. The shader almost fills the instruction limit for a PS2.0 shader when combined with the shadow rendering. Finally the sun is done with HDR which looks nice when mixed with the sky’s procedural light scattering.
There are 2 CPU tests in 3DMark05, both of which are retreads of the game tests. The first CPU test is the Canyon Flight test, the second is Return to Proxycon. In these tests a 640x480 resolution is used along with software vertex shaders which reduces the load on the graphics hardware. Additionally post processing is disabled. The tests also uses fixed frame rendering to ensure that the workload is comparable across systems and as well as the rendering the CPU calculates the flight path of the ship in Canyon.
CPU score is calculated as (cpu1 * cpu2)^0.5*1500
The feature test section of 3DMark05 do not have any relevance to the final score achieved in the benchmark. They are instead used for informative purposes to show how your graphics hardware performs with certain functions. The areas tested are Fill Rate, Pixel Shaders and Vertex Shaders.
Fillrate is split into single and multi texturing and are worked slightly differently (less nice looking) to previous versions of 3DMark so as to reduce the graphics memory bandwidth bottleneck and give more accurate fillrate results.
The pixel shader test uses a section of rockface from Gametest 3 and can be bandwidth limited due to the complex shader and multiple lookups to large textures.
Vertex Shader testing is split into two tests. The first consists of several sea monster models which test transformation and lighting. Specification wise there are several million vertices per monster which all require transformation and illumination. The second test renders a section of grass where each blade is rendered independently and as little work as possible is passed to the CPU so as to reduce its impact on the results. Additionally the scene is rendered from a distance so as to reduce the fillrate bottleneck.
This is an interesting one, and again it doesn’t affect the scores obtained by your hardware. In these tests futuremark test areas of graphics card drivers which are not normally optimised. (Basically the normal method is to render as large batches as possible). The batch tests run 128meshes of 128x128 quads through tests with 8, 32, 128, 512, 2048 and 32768 triangles per batch. Futuremark believe that only the last 2 should be optimised normally.
Before we look at the Image Quality we received when using 3DMark05 and the results obtained its worth pointing out that there are a few more features in the application.
Firstly there is a detailed IQ section where you can grab screenshots, or video of the tests, this – like the benchmark- is configurable and the options are shown below. There is also the option to test filtering using the inbuilt texture filtering and anti-aliasing tool. These are mostly important for hardware reviewers or enthusiasts however definitely a nice to have for everyone (dependant on your version of 3DMark05).
3DMark05 also adds graph creation to its impressive list of features. The Graph mode is again more useful for the reviewers and enthusiast out there however is an excellent tool for comparing results for various different cards
Finally there is the Demo. Each of the game tests is actually a small part of the demo. For example Return To Proxycon is actually much longer and shows a bit more behind what all the fighting is about. Add to this that the demo shows off more of the excellent quality graphics and I recommend that everyone gives the demo at least one viewing.
Overall the IQ in 3DMark05 is stunning, it’s probably the best looking engine that we’ve run on our systems (including Source and Doom3). The lighting effects from the sparks/gunfire in Proxycon are amazing and the water effect in Canyon is spot on. We’ve added a selection of images which were taken using the standard benchmark settings below. Please note, to accurately compare image quality, these are high resolution uncompressed BMP files, dial up users may wish to skip.
ATI (below): Taken using 8.07 with standard benchmark settings
Nvidia (below): Taken using 66.51 with standard benchmark settings
I did of course say overall the IQ is stunning… there is one exception and personally we’re still undecided on whether this is a good thing or not. The effect in question is in Game Test 3. On certain parts of the rockface you will see flickering textures. Since we had access to an earlier build of the benchmark it has been really distracting and looks quite like a driver bug. This isn’t the case though, instead it’s a cause of the lighting/shadow method. The flickering happens where the surface is nearly perpendicular to the light direction. Futuremark tell us that in these areas the flickering is caused by the worst case projection of the PSM shadows where there is just not enough resolution in the depth map caused by a compromise between how large the depth maps used are and how much shadow artefacts to allow.
Reference 6800 Ultra 400/1100mhz
Connect3D X800 XTPE 520/560mhz
Athlon FX53 (Socket939)
Windows XP SP2
Forceware 66.51 beta
Also squeezing in with a new approved driver on 28/9 was Nvidia, their 66.29 scored approx 4600marks on our Ultra where as the newer 66.51 scored over 200points higher which is a nice increase. The newer driver also fixed a few minor bugs on loading screens.
In the overall scores its clear to see the lead with the latest approved driver is firmly with ATI. With a 1000 point advantage over Nvidias fastest and even ATI’s lower clocked XT scoring 700marks higher. Of all the cards on test here the Asus 6800 Non-Ultra scores lowest however due to the enhanced clocks of this model it scores around 400 marks higher than most other manufacturers 6800 NU models (based on some quick tests) and therefore is the Non Ultra to put your cash into. Looking in more detail at the individual game tests we see that the results are as expected , ATI’s cards lead the way in all 3 game tests.
We’ve been very impressed with our initial look at 3DMark05, the tests look gorgeous, the results are always repeatable and the interface is as clean and well laid out as could be expected.
We’ve mentioned a number of times throughout this article that 3DMark05 does things in the same way that the latest and future games do, looking at the results obtained this appears to have been a success. The difference in performance between ATI and Nvidia’s top cards follows the trend set by the Source Engine, FarCry and many other DirectX 9 games. It will be interesting to see how things change with updated drivers however it will take some effort on Nvidias part to catch up and to be honest they really should concentrate on real games over a benchmark so hopefully that’s what will happen.
As far as benchmarks go, 3DMark05 is “the” graphics test to run for comparing products or showing off your latest kit to your friends, two thumbs up from us.
Let the race to 10k begin…