Only the PC version of the test is being made available today, but the associated technical documentation tells us a great deal about how the cross-platform testing will be handled. The new 3DMark has three tests: Ice Storm, Cloud Gate, and Fire Strike. On the PC, all three tests run using DX11's ability to target specific API frameworks: Ice Storm runs in DirectX 9_3, Cloud Gate uses DirectX 10, and Fire Strike is a DX11 benchmark.
Ice Storm forms the core of the cross-platform benchmark test. Once available for Android and iOS, it will run using OpenGL ES 2.0. Target resolution is 1280x720 with 128MB of GPU memory. The benchmark runs at this resolution off-screen, which is part of what makes it useful for testing multiple graphics cards and displays.
The first Ice Storm test, GT1, is light on pixel shading but pushes vertex shading capabilities heavily. 530,000 vertices and 4.7 million pixels are processed per frame. That's well within the capabilities of even midrange handheld graphics; Apple's SGX 543MP4 has a peak fill rate above four gigapixels per second.
The second test, GT2, switches things around. Vertex processing is low, at 75,000 vertices per frame, but per-frame pixel counts jump to 12.6 million. The two tests are designed this way so as to test the two GPU capabilities separately and isolate weak points in any particular solution.
Finally, there's a physics test with four simulated worlds. Each world has two soft bodies and two rigid bodies that collide with each other, as shown above. One thread per CPU core is used, and graphics overhead is kept to a minimum. This test uses the open source Bullet library.
Ice Storm uses the same rendering engine on all platforms.
Cloud Gate targets DirectX 10.0 and keeps the 1280x720 resolution Ice Storm used. It adds 1024K shadow maps, a low-quality depth of field effect, and bloom. Total targeted GPU memory is 256MB. Futuremark has stated that this test is aimed at modern integrated graphics solutions and older PCs.
As with Ice Storm, GT1 and GT2 are designed to measure specific performance targets. GT1 emphasizes geometry processing and uses simple shaders, while GT2 uses mathematically complex pixel shaders but handles less geometry.
The physics test in Cloud Gate uses 32 worlds, with four soft bodies, four joints, and 20 rigid bodies per world. All physics are computed on the CPU. One thread per world is also dedicated to particle simulation (again, running on the CPU).
This is the test PC owners will be most interested in. Fire Strike has two benchmark modes: Normal mode runs in 1920x1080, while Extreme mode targets 2560x1440. GPU target VRAM for normal mode is 1GB and the benchmark uses tessellation, ambient occlusion, volume illumination, and a medium-quality depth of field filter.
Extreme targets 1.5GB of VRAM and increases detail levels across the board. This version of the benchmark is explicitly designed for CrossFire/SLI systems. GT 1 focuses on geometry and illumination, with over 100 shadow casting spot lights, 140 non-shadow casting point lights, and 3.9 million vertices calculated for tessellation per frame. Only 80 million pixels are processed per frame.
GT2 emphasizes particles and GPU simulations. Tessellation volume is reduced to 2.6 million vertices, but the number of pixels processed per frame rises to 170 million.
We haven't had time to run 3DMark across a range of GPUs yet, but we like what the technical documentation describes. Designing a single benchmark that can scale across tablets, smart phones, and desktops is no small feat. Futuremark's goal was to build a product that would make it easier to compare 3D performance between a Galaxy S3 and an iPhone 5, or a Windows RT tablet vs. a conventional x86 device. As a reviewer, that's incredibly useful.
We'll have further coverage available in the not-too-distant future, particularly when the iOS and Android flavors of the benchmark are available. If you're curious to see how your system scores, the Basic edition of the test is free.