View Single Post
  #7  
Old 09-13-2011, 06:36 PM
Blackdog_kt Blackdog_kt is offline
Approved Member
 
Join Date: Jan 2008
Posts: 2,715
Default

I had seen this linked before but only just now took the time to read it. Just...wow. Multi-hundred dollar GPU setups that companies know face such a problem and they don't bother telling customers about it, and we're sitting here complaining about a $50 game

On a serious note, the summary for those who don't want to read the entire article.
  1. As long as the average FPS is in an acceptable range, maintaining the illusion of movement in a game depends more on having subsequent frames render in a similar amount of time each, and less on the absolute FPS numbers. FPS can't measure this because it averages the results over an entire second, so they use a new unit to measure it: milliseconds needed per frame rendered, so if this catches on in further reviews expect it to get an acronym of its own like mspf or something
  2. Stutters are perceptible "breaks" in this illusion, caused by having subsequent frames render in amounts of time that differ more than a certain threshold: if mspf between subsequent frames differs more than a certain value, stutters occur.
  3. Dual GPU setups are essentially a master-slave arrangement. While both GPUs render an equal part of the picture if AFR is used (alternate frame rendering, GPU1 renders frame 1, GPU 2 renders frame 2, etc), only GPU1 is connected to the monitor: GPU2 has to push the data "through" GPU1 before it reaches the monitor and this raises the mspf for GPU2. Every second frame takes longer to render than the previous one.
    In other words, dual GPU setups have an built-in "stutter generator" by hardware design, it's just that we don't always perceive it because it doesn't always go above that certain threshold.
  4. Fraps (and maybe other FPS measuring tools too?) is inaccurate in measuring this, because it measures FPS by looking at when the frame is passed on to DirectX, not when it's actually rendered and displayed, so it misses that loop which institutes the delay.
    And now for the points that are especially relevant to CoD and flight sims in general...
  5. A high vRAM usage exacerbates this problem.
  6. The problem is more prone to occur on game engines that track time more accurately.

We know CoD's textures are quite large in size (high vRAM usage) and we know all sims are very time-sensitive. Tracking the ballistics of weapons down to individual bullets/cannon rounds in relation to the target, especially in an age where hitboxes are no longer used, means a high demand for precision in how time is measured: the game engine has to "slice down" the time in smaller pieces and the finer control over time the engine has, the more pronounced the problem becomes.
But we absolutely need that time control for accurate ballistics, FM and DM, not to mention consistency in multiplayer combat.

Ironically enough, nVidia's solution to this is to lower overall FPS by inserting a metering system in SLI setups that will delay frames that are rendered faster than others to produce a more even sequence (since they can't speed up the slow ones), which is something people who've been flying the microsoft civilian flight sims have been doing for a long time, either by setting the in-game FPS limiter or by using external ones.

It also explains perfectly well why my FPS with the latest beta are lower but the game is looking smoother.

In conclusion:
  • FPS measurements on their own don't mean much, as stutters can occur even on high FPS. Plus, FPS are measured inaccurately by taking measurements from inside the rendering loop instead of measuring the entire cycle once it's completed: the point where the problem occurs is further down the "pipeline" than the point our FPS measuring tools get their data from, so they miss it completely.
  • The things a flight sim needs to do demand a certain design of the foundation game engine that is more susceptible to these effects.
  • Lowering the graphics load and especially reducing vRAM usage makes the problem harder to spot because the difference in mspf between individual frames is lowered (which, as far as CoD is concerned, means turn your texture size down a notch).
  • It's an innate hardware design issue first and foremost and impossible to optimize 100% for every single game.

So, unless you're running quad SLI/Xfire on 100FPS with a sufficiently low and steady mspf between subsequent individual frames, sell your multi-GPU setups and put that money on a beefy single GPU - single card setup, it will run better and get you some money back too that you can spend on another upgrade
Reply With Quote