Vantage is designed to be a gaming benchmark, not just a GPU benchmark, and to that end, has four distinct subtests with a few feature tests thrown in for good measure. The four tests are Jane Nash, New Calico, AI and Physics, each doing something very different.
3DMark Vantage Jane Nash
Jane Nash tests high complexity dynamic objects, cloth, water rendering techniques and Anisotropic materials. You might also note the logo on the boat when thinking about pushing boundaries. It simulates a lot of what you will find in an indoor based game.
New Calico hits the points that Nash doesn't, lots of moving objects, little skinning, shadow maps and instancing. It also does a bunch of ray-tracing effects, Parallax Occlusion Mapping, True Impostors and Volumetric Fog. Think outdoor and space games here.
3DMark Vantage New Calico
Together, the two GPUs complement each other quite nicely. There is no hardware in existence that can do both sets of tests at the same time, and putting enough of each in to run at decent speeds defeats the purpose. By the same token, a scene with a few million hugely detailed unique objects that runs at .0000000000416FPS doesn't buy you much either, even if it will run at 30+FPS on hardware common in 2016.
The CPU tests start out with AI and follow it up with Physics. Both of these are inherently non-graphical, but a black screen for five minutes followed by a number does not make a good benchmark. To solve this, both of them use the graphic engine from the GPU tests with a lot cut out.
There is no post processing or complex shaders other than the bare minimum required for the engine. The geometry is vastly simplified, shadows are right out, and only what you see is rendered. Basically, they are putting up pictures but that has nothing to do with the simulation score.
3DMark Vantage AI test
The AI test is simple enough to describe, there are a bunch of planes racing through a canyon vying for the best path through a series of rings. Computing that path in 3D is not an easy task, and avoiding collisions while simulating the plane physics as well only adds to the load.
Vantage does a stochastic search of the space to pick a path for the planes to follow. More CPU power allows the planes to plot a better course by trying more options. It is also somewhat random in nature which makes it a much better benchmark.
The engine spawns one thread per CPU, drawn from a thread pool. If a CPU is free, it gets hit with a thread, and each one results in a path. The number of paths computed is divided by the time, and that is more or less the score.
The physics test measures the other big problem facing CPUs in gaming, and it is gaining more importance as time goes on. The test itself uses the Ageia Physx API, and will use a (the?) PPU if available. This one makes me a tad nervous because results can be pretty radically skewed by the accelerator, but there is an easy disable option should you not want acceleration.
3DMark Vantage physics test
Physics looks similar to the AI test with planes flying through hoops and posts while spraying volumetric smoke. Without the AI portion, they crash into each other and the gates, breaking into 12 rigid bodies in a correct physical way. The gates are modeled as pressurised cloth or tofu (elastic foamy substance), and hitting them will do the 'right' thing as well. Additionally, the smoke will dissipate correctly over time.
If you are thinking lots of particles and objects, you are right. Interestingly, that count is dependent on the number of CPU cores, with one pair of gates per core. If you have a PPU, it will have one gate pair per core minus one, plus for for the PPU. Using rough math, that would give a PPU about 4x the power of a modern CPU core.
Scoring this test is simple, how many ops can you get in a time period. The more ops the higher the score, nothing complex there. One thing to think about though is that Nvidia bought Ageia a few months ago. Given that both GPU companies have been caught red handed massaging benches including older versions of 3DMark, this is kind of scary. With Nvidia owning the API, they can cheat readily and mask it in 'upgrades'. Watch this test VERY closely, it puts too much control in the hands of a company that regularly abuses power.
Like earlier versions, there are feature tests, in this case six of them. They are in order Texture Fill Rate, Color Fill Rate, POM Shader, Cloth Simulation, Particle Simulation, and Shader Math. Other than POM, they all should be more or less self explanatory.
POM is Parallax Occlusion Mapping puts up a 4K * 4K height map and lights it with 7 lights, 4 point and 3 directional. The map should block the light from reaching the camera, hence the occlusion in the name.
The four main tests are all combined into two sub-scores, which are in turn combined into one main score. The graphics tests are S(graphics) = C(GT1)F(GT1) + C(GT2)F(GT2). The C term is a weighting constant and the F is the FPS result. similarly, the CPU score is S(CPU) = C(CPU1)O(CPU1) + C(CPU2)O(CPU2) where O is the Ops measured.
The end result can be summed up with 3DMark Vantage Score = (WG + WC) + ((WG/SG) + (WC/SC)). WG is a weighted graphics score, WC is the weighted CPU score, and SG/SC are the numbers from above. If the weighting part makes you confused until your head hurts, bear with us for a bit.
The reason for the weighting is simple, there is not just one 3DMark Vantage test, there are four. From bottom to top they are Entry, Performance, High and Extreme. Each has a different weight for the scoring, with the low end giving CPU a higher weight and high end emphasizing the GPU more. If you want the exact numbers, they are below.
3DMark Vantage scoring weights
The scores will all be tagged with the letter of the test run, E, P, H or X followed by a 4 digit number. Once the XtremeSystems guys get their teeth into things, it won't be long before we see 5 digit numbers, that is for sure.
The preset tests are run at 1024*768, 1280*1024, 1680*1050 and 1920*1200 respectively. High and Extreme use AA instead of Trilinear with 8x for High, 16x for Extreme. Similarly, the harder the test, the harder individual features are pushed. The same tests are run, they are just thrown more complex data to crunch.
In another change, there will be four different versions of the benchmark ranging from free to almost $500. They are named Basic, Advanced and Professional, with a 'Trial' version of basic for free.
Trial lets you run and submit one score without cost while Basic lets you submit all the scores you want on single one of the four presets. It only costs $6.95 or about .43 Euros if you wait another week. Advanced is the same thing but gives you access to all four presets and custom settings for a mere $19.95.
The last one is Professional, and it is probably something most readers will never need. The main thing it does is allow for commercial use, the lesser versions do not license you for this. It also comes on a CD, has technical support and a CLI for scripting. All of these things add up to a price of $495, so unless you need those things, stick with Advanced, it gets you 99.9% of the way there for 1/25th the price.
The only problem we have found with the benchmark is that it is DX10 only, thus requiring Me II, the broken OS. Due to the enforced malware in that OS, it makes it an untenable proposition, but we do understand why it was done, you can't get DX10 any other way. Since this is a DX10 benchmark...
In the end, Futuremark seems to have done the right things for the right reasons. As with any benchmark, there will be quibbles, bitches and optimizations. We are on the first step of a new learning curve, and the community knowledge base is sure to grow exponentially in the next few days.
Grab your copy here and get started. µ
Sign up for INQbot – a weekly roundup of the best from the INQ