Ihatethedukes
New Member
Hi all, I'm putting together a thread that discusses and, I think, approaches the concept of microstuttering from a much better angle than approached before.
Introduction
The current view floating around the internet is that multi-GPU cards are susceptible to what is called microstutter. The prevailing mechanism put forth by people who believe in multi-GPU microstutter is that the frames being output from each GPU in AFR rendering mode (alternate frame rendering) are not evenly spaced out and leave relatively large gaps in between frames rendered, giving an illusion of lower FPS than are actually output.
As you can see from rage3d.com's figure that the mechanism is from two gpus rendering in parallel, finishing their frames at nearly the same time and displaying them close together then you're left looking at the last frame much longer than you would were the two rendered and displayed at regular intervals. It is this 'pairing of outputs' (PoPs) that supposedly leads to the visible 'choppiness' of microstuttering gameplay.
There are a few articles around that attempt to approach the issue of ustutter from:
http://www.rage3d.com/reviews/video/ati4870x2cf/index.php?p=1
and
http://translate.google.com/transla...page_1/fromTicket_/key_//tm.htm&sl=auto&tl=en
While both works are admirable, neither really tell a good story of ustutter. The older rage3d is easily the better of the two, and while their attempt at quantifying their ustutter by binning frame time is a start it falls very short of its goal because frametimes are FPS dependent, meaning, the higher the FPS the more will be binned into the faster bins.. The more recent donanimhaber article lacks a very basic control of a single GPU running the same benchmark. I look the liberty of running a single 5870 in the same benchmark for comparison.
Single 5870
Their 5970
Remember, I did not cherry pick these results I just took the first 30 frames or so to illustrate that a 5870 has a similar rendering irregularity. Here's the whole run.
This result alone makes me doubt the validity of the PoPs mechanism for ustutter. So, I'm developing A QUANTITATIVE METHOD for determining the validity of the PoPs mechanism to ustutter.
METHODS:
I am defining that microstutter is the inconsistent time intervals between rendered frames nothing more, nothing less. This can exist on both single and multi-GPUs (the mechanism is in dispute more than its existence). I propose that we impose the limit of microstutter as being a >50% difference of the moving average delay. This 50% number was chosen because it allows for 'noise' or natural differences in difficulty of rendering one frame and the next so that they don't show up as microstutter but still allows for the slight delay inbetween the PoP frames.
A moving average, averages the frame in question with the ones surround it (let's say +/-60frames). This gives us a local FPS average that isn't effected by a higher FPS 1 minute later in the benchmark to use as the 'theoretical best case' to compare the real delay too. Then we compare the actual frames one by one to their moving average. I will be considering that any variation over 50% of the value of the moving average = microstutter. EXAMPLE:
Lets have a worst case scenario. You've got a system that either produces frames at 5ms and 30ms apart from one another alternately. The moving average for 30frame +/- is the expected 16.6667. (I chose 60PFS arbitrarily). We then compare the ACTUAL frametime of 30 as follows:
(30-16.6667)/16.6667=0.8 (80%)
(5-16.6667)/16.6667=-0.7 (70%)
We then count the amount of times we see a >50% difference. If we count >50% as 1 and <50% as 0 we can graph it and see what areas there are a lot of peaks at 1 (1peaks).
Since this would have 100% of the delays as >50% then it'd be RIDDEN with microstutter.
An issue with this is, how do we just have to decide on what consists of 'real' microstutter?
I'll reiterate that the mechanism in which this inconsistency comes about is in dispute. People insist that it's in multiGPU formats that it comes about. Because of this, we can simply compare the amount of 1peaks of a single card to the amount of that a dualGPU creates and see if it's elevated in dual GPU formats. The reason for this approach is that it's quantitative rather than subjective, FPS blind and doesn't rely on subjective analysis.
We need identical benchmark data from a 5870 and a 5970 (meaning same time/settings/benchmark/place in benchmark).
Results
The results are interesting to say the least.
5870 - Average score per frame: 0.438
5970 - Average score per frame: 0.307
Analysis
The single GPU 5870 actually has MORE ustutter than the dualGPU 5970. Nearly ~44% of the frames rendered with the single GPU deviated from the local average render time by >70%! The 5970 on the other hand only had ~31% of its frames rendered with a deviation of >70% that of the average render time. To address the possibility that the 5970 has fewer, but larger variances while the 5870 has more smaller variances, the threshold for ustutter was increased to 90% variance and the 5970 had ~7% occurrence while the 5870 had ~11%. This suggests that the 5870 has both more and larger ustutter events than the dual GPU 5970.
This data refutes the myth that multi-GPU systems create ustutter.
Introduction
The current view floating around the internet is that multi-GPU cards are susceptible to what is called microstutter. The prevailing mechanism put forth by people who believe in multi-GPU microstutter is that the frames being output from each GPU in AFR rendering mode (alternate frame rendering) are not evenly spaced out and leave relatively large gaps in between frames rendered, giving an illusion of lower FPS than are actually output.
As you can see from rage3d.com's figure that the mechanism is from two gpus rendering in parallel, finishing their frames at nearly the same time and displaying them close together then you're left looking at the last frame much longer than you would were the two rendered and displayed at regular intervals. It is this 'pairing of outputs' (PoPs) that supposedly leads to the visible 'choppiness' of microstuttering gameplay.
There are a few articles around that attempt to approach the issue of ustutter from:
http://www.rage3d.com/reviews/video/ati4870x2cf/index.php?p=1
and
http://translate.google.com/transla...page_1/fromTicket_/key_//tm.htm&sl=auto&tl=en
While both works are admirable, neither really tell a good story of ustutter. The older rage3d is easily the better of the two, and while their attempt at quantifying their ustutter by binning frame time is a start it falls very short of its goal because frametimes are FPS dependent, meaning, the higher the FPS the more will be binned into the faster bins.. The more recent donanimhaber article lacks a very basic control of a single GPU running the same benchmark. I look the liberty of running a single 5870 in the same benchmark for comparison.
Single 5870
Their 5970
Remember, I did not cherry pick these results I just took the first 30 frames or so to illustrate that a 5870 has a similar rendering irregularity. Here's the whole run.
This result alone makes me doubt the validity of the PoPs mechanism for ustutter. So, I'm developing A QUANTITATIVE METHOD for determining the validity of the PoPs mechanism to ustutter.
METHODS:
I am defining that microstutter is the inconsistent time intervals between rendered frames nothing more, nothing less. This can exist on both single and multi-GPUs (the mechanism is in dispute more than its existence). I propose that we impose the limit of microstutter as being a >50% difference of the moving average delay. This 50% number was chosen because it allows for 'noise' or natural differences in difficulty of rendering one frame and the next so that they don't show up as microstutter but still allows for the slight delay inbetween the PoP frames.
A moving average, averages the frame in question with the ones surround it (let's say +/-60frames). This gives us a local FPS average that isn't effected by a higher FPS 1 minute later in the benchmark to use as the 'theoretical best case' to compare the real delay too. Then we compare the actual frames one by one to their moving average. I will be considering that any variation over 50% of the value of the moving average = microstutter. EXAMPLE:
Lets have a worst case scenario. You've got a system that either produces frames at 5ms and 30ms apart from one another alternately. The moving average for 30frame +/- is the expected 16.6667. (I chose 60PFS arbitrarily). We then compare the ACTUAL frametime of 30 as follows:
(30-16.6667)/16.6667=0.8 (80%)
(5-16.6667)/16.6667=-0.7 (70%)
We then count the amount of times we see a >50% difference. If we count >50% as 1 and <50% as 0 we can graph it and see what areas there are a lot of peaks at 1 (1peaks).
Since this would have 100% of the delays as >50% then it'd be RIDDEN with microstutter.
An issue with this is, how do we just have to decide on what consists of 'real' microstutter?
I'll reiterate that the mechanism in which this inconsistency comes about is in dispute. People insist that it's in multiGPU formats that it comes about. Because of this, we can simply compare the amount of 1peaks of a single card to the amount of that a dualGPU creates and see if it's elevated in dual GPU formats. The reason for this approach is that it's quantitative rather than subjective, FPS blind and doesn't rely on subjective analysis.
We need identical benchmark data from a 5870 and a 5970 (meaning same time/settings/benchmark/place in benchmark).
Results
The results are interesting to say the least.
5870 - Average score per frame: 0.438
5970 - Average score per frame: 0.307
Analysis
The single GPU 5870 actually has MORE ustutter than the dualGPU 5970. Nearly ~44% of the frames rendered with the single GPU deviated from the local average render time by >70%! The 5970 on the other hand only had ~31% of its frames rendered with a deviation of >70% that of the average render time. To address the possibility that the 5970 has fewer, but larger variances while the 5870 has more smaller variances, the threshold for ustutter was increased to 90% variance and the 5970 had ~7% occurrence while the 5870 had ~11%. This suggests that the 5870 has both more and larger ustutter events than the dual GPU 5970.
This data refutes the myth that multi-GPU systems create ustutter.
