7XXX series, frame time, and power control settings...

[Bensam123]

So... in a long running angry RMA with Gigabyte, I finally got a 7870 back. Originally, I sent a 6970 in for a RMA roughly two months ago. I got a defective video card back after a month and took another week to get them to 'upgrade' me to a 7850, which wasn't a upgrade at all, and then finally to a 7870 for their error. Two months in RMA using a 4870 while all this is happening is a long time.

Anyway, after getting it back, I noticed the responsiveness of my card is roughly the same as my old one. It was producing more frames, but didn't feel all that responsive. Which sorta corresponds to the results in the latest benchmarks. AMD cards have more variance and tend to produce more spikes. Spinning the camera in Guild Wars 2 for instance does this a bit. It was better then my old graphics card, but not by much (GW2 isn't a shining example of a smoothly coded game though). I was playing around in my overdrive options because I disliked the fan speed settings... Well in addition to finding out I can run the thing at 20% and it only gets up to 62c under full load, apparently turning up the power control settings to +20% makes the card super responsive.

I was honestly pretty amazed. Even more so when I started up WoW today and realized that the game is so fluid it's like **** water. TR hasn't addressed this at all and it makes me wonder if they've even tried it lately. While this may not be a traditional avenue to persue, I could see this making a very big difference in benchmark results, especially those focused around latency. Looking at GPUZ running in the background it adds less then 1amp onto power draw if I'm reading it right (VDDC in). Most gamers don't give a crap about power envelopes and I would say most don't even know about this setting or what a big difference it makes. I'm guessing the 7xxx series has some issues with power throttling when it's being pegged, specifically around very intense scenes... generally those where you need performance the most. The more I'm using it like this, the more I'm finding out the card is generally more responsive in both games and on the desktop; overall use has improved quite a bit.

Keep in mind my framerates haven't improved at all, but what has most definitely improved is the overall fluidity and responsiveness of the games I'm playing.

[MattVonX]

intresting,

so you just used overdrive in the CCC settings and increased power to the gpu by 20%?

did it effect fps? or more just a feeling of smoothness?

[Flying Fox]

You can probably try doing what TR does with frame time captures in Fraps. Then you can see if the spikes have been smoothened?

[killadark]

now thats an interesting find
can anyone back him up by trying it on their 7xxx series cards?

[mortifiedPenguin]

I already have the power threshold at +20% on my 7970s but I can try putting it back down to 0 and see what happens. Probably should try it with and without Crossfire. You'll have to wait 8 hours, though.

You can probably try doing what TR does with frame time captures in Fraps. Then you can see if the spikes have been smoothened?

Definitely. While valuable, I don't think pure "feel" is a benchmark that can be reported on unless performed using a true controlled double blind experiment with an appropriate sample size. In other words, you aren't allowed to say "I can see/feel it, so should you" . Anyway, I cooked up a few spreadsheets a while ago that appear to produce the same graphs as TR's reviews so the data processing should be relatively quick. Are there any demands on game and test duration? Resolution (I normally run 6048x1200 so my defaults are atypical)?

[mortifiedPenguin]

So, I did GW2. Ran around the Asura gates in Lion's Arch repeatedly for 1minute while recording frame times. Max settings @ 1920x1200.


As far as smoothness goes... I'm not really sure we can say there is a winner. FWIW, +20% power (aka Overdrive) seems to have a slight advantage as far as the total number of frames produced.

If somebody can suggest a good one time upload site for the CSV data, I can upload that too.

[kumori]

What are the units on the Y Axis. Is this measuring frame render time?

[mortifiedPenguin]

What are the units on the Y Axis. Is this measuring frame render time?

Sorry, that was pretty basic of me not to label the axis. X is frame number. Y is frame time in ms.

[Bensam123]

Looks like you have more the one set of data with no labels as to what each color represents...

Interesting, I actually made this post before the big AMD oscillation debacle that followed this December and forgot about it. I made a few posts in both the articles and sent Damage a message on the forums to which he simply said "I don't currently believe a basic malfunction with a power-saving mechanism...". But I still think there is more performance to be found in the 7xxx series by increasing powertune to +20%. Unfortunately almost no one is looking at this.

This definitely was a feel thing when I reported it as it wasn't something I measured in frame rates, but I'm pretty sure this effect could be captured with millisecond frame rates.

Thanks for the bump Matt.

[kumori]

The graph does actually have a legend if you horizontally expand your browser window:

Green = Single Card No Overdrive
Purple = Single Card with Overdrive
Red = CrossfireX with Overdrive
Blue = CrossfireX No Overdrive

[mortifiedPenguin]

The graph does actually have a legend if you horizontally expand your browser window:

I didn't realize the image overflowed. Just click on it.

I also just realized the Overdrive numbers had a mild overclock applied . If you guys wish I can run the numbers for it again.

[Chrispy_]

I wonder if it affects some cards and not others.

As I'm sure you're aware, different dies have different levels of success, which is why they're binned in the first place. It's possible that a "good" die will never run out of power, whilst a "good enough" die will occasionally need to encroach into power-throttling.

I might try a 90-second test run in Borderlands2 when I get back home this weekend, especially since I run at the same settings Scott used in his articles and secondly because I've already commented about how my results are wildly different to the articles. I'm guessing the real test would be at powertune -20, 0, and +20.

[Bensam123]

Perhaps, I don't think this is because powertune is encroaching on the thermal envelope, but rather powertune is simply throttling the GPU in certain scenarios while it's well within the PCIE spec to save on energy. I'm going to make a short test using a similar path as Damage tomorrow in GW2.

[ultima_trev]

Bensam, you are on a roll man! First the core unparking, now this.

After turning up Power Tune to +20% in CCC, I noticed much less latency in Crysis Warhead, Crysis 2 and Metro 2033 (the biggest stutterers of the games I play) on my HD 7850.

[Bensam123]

Yeah, the original post is from September too, I just didn't get any responses to it back then. I still haven't done a run through, but I plan on it.

You can definitely feel it plays differently regardless of quantitative data though, right?

[ultima_trev]

There are a few more visible drops using the Metro Benchmark utility with core parking and power tune at +0. Not sure which is the more the culprit though. Unigine Heaven definitely is smoother, and that's just a flythrough tech demo... if it was an actual game the difference would be massive. Again, not sure if Powertune is to thank or core parking. However, there were definitely less dropped frames in Windows 7 for me... enabling power tune +20 and disabling core parking in Windows 8 made my gaming smoother than it was in Windows 7... My guess is that Windows 8 is more aggressive with the power saving features.

[Bensam123]

Interesting, you may want to report that in the core parking thread as we don't have a lot of W8 users to compare to. A lot of people believed that core parking wasn't a issue in W8 due to an improved scheduler and power management...

[Bensam123]

So, I tried making my own attempt at measuring this by following the testing in the latest guide as close as possible. Scott doesn't detail the exact location and I messaged him, but didn't receive a response. After some sleuthing it turned out to be a run between Fields Waypoint and Orchard Waypoint, Shaemoor Fields, Kryta, following the road as closely as possible.

I set my settings for exactly the same as in in latest 7950 review. Interestingly enough LoD and Shadows weren't on their highest setting in the TR review. I left those the same as the TR example though.

I was originally going to do this in Gdocs since it would be quite easy for other people to access and parse at their own bidding, but as it turns out gdocs has issues with large sets of data (mine had 20k rows) and was slow as molasses.

So I ended up parsing it in Libre Office to make graphs and then did some analysis in SPSS.

My System:
i5-3570k@4.2
7870@1100/1200
Resolution was set at 1680x1050
Driver 12.11b11

Somethings to note: I'm on a populated realm so various people were wondering around in this zone. Occasionally a battle happened, but there was more of that in the 20% run through then the 0%. I left the camera at the default angle when warping into the point and ran from one end to the other, then back again. So I went over the road three different times, Fields > Orchard > Fields > Orchard.

Subjectively the system felt more responsive on +20% (but that's qualitative once again).




Std.dev is worse off on the +20% results, but that takes into account drops to lower frame times as well as higher ones (not just higher frames). Mean for +20 is lower as well as the min and max. The differences seem largely trivial, but a higher std. dev may be indicative of it being more responsive. The red line stepdown in the middle of the graph compared to the smooth one for blue was due to the way I spun around. I did a short burst with the +20 twice, compared to just holding down the turn key with the 0, this may also have influenced results.

It's quite hard to measure results like this in a live game, I'll have to try doing a circular run around one of the bigger cities tomorrow as cities seem to stress graphics more.

[Bensam123]

Curiously, the skewness of the +20 is .2692 and the 0 is higher at .2829. Skewness being a indicator of the direction of the distribution. A histogram confirms this.