Test Descriptions

Even though the 3DMark has an awesome demo, it is a 3D performance benchmarking program. When the benchmarking test is run, a series of tests are performed that each collect measurements of the performance of the system. The measurements are collected according to the 3DMark Testing Methodology. From the test measurements, the 3DMark score and other test results are calculated. The tests are the following:

3DMark2000 has two game scenes, but there are three detail level settings on each game test: low, medium and high.

Game Test 1 is a helicopter flight simulator / shoot’em up.

Game Test 2 is a 1st / 3rd person shooter type of game, based on Max Payne’s radiosity lighting and exit rendering technology.

CPU Speed - Tests the system's ability to handle geometry in a realistic, low fill rate situation.

Fill Rate - First test has four textured geometry layers, and the second one uses one geometry layer with four textures, rendered with multi-texturing.

High Polygon Count - Polygon throughput test is a theoretical test that shows the system’s polygon throughput with different light counts.

Texture Rendering Speed - The test runs through a tunnel, which is mapped with many 256x256 textures.

Bump Mapping - The test runs through different kinds of bump mapped materials: three emboss bumps and one environment mapped bump test.

Image Quality Tests - These tests help you to analyse the image quality of your 3D accelerator and compare it to other accelerators. This test is included only in 3DMark2000 Pro.

3DMark Overall Score

3DMark Overall score is calculated by summing up the result of both game tests and all of their detail levels multiplied by three. The formula is as follows:

OverallScore = (Game1LowFPS + Game1MediumFPS + Game1HighFPS + Game2LowFPS + Game2MediumFPS + Game2HighFPS) * 3

The unit for 3DMark Overall Score is '3D mark'.

Game Test 1 - Helicopter

Game Test 1 is a helicopter flight simulator / shoot'em up. The landscape is code-generated at load time, and is identical each run time. Landscape geometry, number of trees and amount of objects and explosions scale with the detail level.

Statistics: Average in View Cone
Detail Level Objects Lights Triangles Vertices Textures
Low 23 4 6 997 11 601 2.74 MB
Medium 37 5 13 520 22 875 2.87 MB
High 39 5 53 026 71 292 2.81 MB

The figures above are averages from the benchmark run, and the scenes have in total more triangles and textures than the above figures. Also please note that all lights do not light all objects or vertices, but have limited falloff ranges.

Game Test 2 - Adventure

Game Test 2 is based on Max Payne's radiosity lighting and exit rendering technology. Characters are skinned (not DX7 skinning) and have distance LOD's. Static geometry (barrels, light posts and so on) also have distance LOD. Specular gouraud is used heavily.

Statistics: Average in View Cone
Detail Level Objects Lights Triangles Vertices Textures
Low 22 4 9 392 16 356 3.14 MB
Medium 33 7 17 084 24 023 3.31 MB
High 44 8 29 941 36 983 3.41 MB

The figures above are averages from the benchmark run, and the scenes have in total more triangles and textures than the above figures. Also please note that all lights do not light all objects or vertices, but have limited falloff ranges.

CPU Speed Test

The purpose of the new CPU Speed test is to measure real-life non-fill rate limited game performance. It runs fixed frames from both of the high-detail level game tests in an 640x480 viewport on screen to minimize the effects of fill rate. You can run this on any of the optimized software pipelines and also with hardware T&L (the test uses the pipeline selected from display settings). The test can only be run when 16 bit color, textures and Z are selected.

To directly compare performance of CPUs using the CPU Speed Test please ensure that you select the optimized CPU Transformation & Lighting pipeline before running the test. The optimized CPU and Hardware Transform & Lighting pipelines may have slight differences due to the nature of Hardware T&L and the APIs used by 3DMark. Hence, users should not directly compare the performance results between optimized CPUs and mixed Hardware T&L pipelines.

Currently the CPU is usually used for Transformation and Lighting. However, in the future, the high-end gamer market segment may evolve towards using a Hardware T&L graphics cards. For future applications to deliver the best gaming experience, the CPU will evolve towards supporting an increasing amount of AI, physics modeling, and custom lighting that can make games even more cool and lifelike.

Result: CPU 3D marks (Game 1 FPS + Game 2 FPS) * 5

Fill Rate Tests

Fill Rate
The test consists of 4 geometry layers. Each geometry layer consists of 2 polygons, filling the screen completely. Every geometry layer has the same texture. The test uses separated geometry so that the "Multi-texturing" features could not be used.

--------------- layer 4 additive
--------------- layer 3 additive
--------------- layer 2 additive
--------------- layer 1 copy

Fill Rate using Multi-Texturing
This test is similar to the above, except that it attempts to use "multi-texturing". There is only one geometry layer having four textures mapped onto it. The textures are constructed as a single material and optimized to use as few passes as possible on different hardware. If the hardware supports single pass dual texturing, this test uses it. In theory if there would be hardware that could do four textures in single pass, this test would will take advantage of it. If no single pass multitexturing capabilities were found, the test reverts to doing multiple passes.

--------------- layer 1 additive + additive + additive + copy

The Fill Rate tests use Z-buffer, no mipmaps, bilinear filtering. The texel / pixel ratio varies when resolution is changed, just as in games when they are run in higher resolution.

A texel is a "texturized pixel". In the Fill Rate tests, each physical screen pixel gets four textures on it, so each individual pixel has four texels on it. For example: A 1024 * 768 Resolution screen has 786,432 individual pixels on it. This test fills each of them 4 times. So the screen has: 4 * 1024 * 768 = 3,145,728 texels on it.

Results: MTexels/s (1000*1000's of texels per second)

High Polygon Count Tests

The High Polygon Count test has 30 toruses with 5000 triangles each = 150 000 triangles simultaneously on screen, an amount we think will be typical in games in one years time. It is run in with different light counts:

- 1 directional light
- 1 directional and 3 point lights
- 1 directional and 7 point lights

The test never rejects any polygons on bounding box check basis, but stresses the clipper and backface cull portions of the pipeline normally. The geometry is static, only the object matrices are scaled. The object geometry is created optimally, in such a way that only one vertex per new triangle needs to be calculated. One 256x256 texture is used (mipmapping is enabled).

Result: Ktris/s (1000's of triangles per second)

Texture Rendering Speed Test

The Texture Rendering Speed test has changed since 3DMark 99 Max. It was redone for 3DMark2000 to reflect a more realistic situation in current and future games.

There are four tests: 8MB, 16 MB, 32MB and 64MB. The 64MB test requires 128MB of system memory and you must have enough AGP memory available (not all accelerators support addressing over 64 MB of memory). You may need to increase AGP Aperture Size in your computer BIOS to get enough AGP memory to run this test.

The test runs through a tunnel, which is mapped with 256x256 textures:
- 8 MB: 48 textures in 16 bit color, 24 in 32 bit color
- 16 MB: 96 textures in 16 bit color, 48 in 32 bit color
- 32 MB: 192 textures 16 bit color, 96 in 32 bit color
- 64 MB: 384 textures 16 bit color, 192 in 32 bit color

The tunnel is always the same length. Texel/pixel ratio therefore increases as the texture amount increases as well, just as it would in a normal game that uses lots of textures. Mip-mapping is used.

Result: FPS (Frames Per Second)