I have a problem with Nvidia GPU Boost.

[Ferzerp]

Why? If there is TDP headroom, the card clocks higher. This might happen at 30fps, at 60, at 100. Two applications might load the GPU to 100% (according to Afterburner, GPU-Z etc.) but the actual consumption is different. So in app A it might clock higher than in B, giving additional performance.

In practice you don't really see that. You set your offset for a max, and it pretty much runs at that max or 1 bin below it if you're running in the low 70's (or you run slower if you're using vsync or the "fps target" functionality and don't need the extra speed).

3dmark11, furmark, the witcher 2, metro 2033, wow, etc. All sit at 1215 Mhz for me the entire time. The only time it goes below are when the scene just doesn't require that much to maintain 60 fps (I use adaptive vsync).

edit: these are all overclocked amounts. If I don't overclock it, it just sits at 1098 Mhz the whole time. That's mine's default max boost.

 

[boxleitnerb]

Then the question is, why have the turbo at all and not just set the base clock at 1098?

 

[Ferzerp]

Perhaps the assumption is that people aren't going to be running them 100% the entire time they are in use. Considering how common 1920x1080 is, most people, if using adaptive vsync, and most games, would rarely need more than the 1006Mhz.

I might set my offset such that it will clock up to 1215Mhz, and if I am using the resolution most people use, it will only be at 1215Mhz in those rare cases it needs it. Instead, I use 2560x1600 and need it all the time (until my second one arrives, then I expect that they'll only rarely run at max speed).

 

[boxleitnerb]

Wait, now I'm confused:

I thought the framelimiter was the thing that leads the card to change clocks according to target fps. Now adaptive vsync does the same?

 

[Ferzerp]

Wait, now I'm confused:

I thought the framelimiter was the thing that leads the card to change clocks according to target fps. Now adaptive vsync does the same?

From what I've seen, if you're using adaptive vsync and there is a bit of headroom over your refresh rate, it will clock down because there is no need to boost. It does it all the time in wow, I'll check skyrim later and see because that's the other game I've been messing with lately that's locked at 60 fps with one card.

 

[BallaTheFeared]

These power saving features may very well make the cards work when they otherwise wouldn't but dynamic clocking and the subsequent impact on frame times is quite noticeable. Its a bit of a shame they don't have another solution really, because inconsistent performance isn't good when it comes to graphics.

Which is all the more confusing because [H] reported 680 SLI having noticeably smoother rendering than 7970 CF.

 

[Ferzerp]

I think in all threads pertaining to GPU boost, everyone should be required to post the following info.

a) I am basing my assumptions on reading reviews/previews

or

b) I am relating my first hand experience from hardware I own or have access to


The reason I say this is because none of the reviews have described this dynamic clocking in a manner that is consistent with what I and other owners of the cards report. The impression that the reviews give is that the clocks should swing wildly all over the place, which frankly, they just don't do. We then have people who read a (p)review or two and suddenly they are experts on a functionality that they really have no understanding of.

Sure, they cards act differently than we are accustomed to, but they are thoroughly predictable, and the reasons the clocks may dynamically change are pretty easy to understand. They aren't some mysterious, magical, voodoo clocked demons. They pretty much stay nearly exactly where you set them within a tiny bit of wiggle room, or clock lower when you don't really need it. Of course, I understand not everyone has the ability to leave the load %, power %, fan %, temp, clock speed, etc displayed on their keyboard the whole time.

The first week I had the thing, I spent more time watching the LCD on my keyboard than paying attention to the game I was playing. Even after that, I had a few flawed assumptions (mostly based on the stock behavior as I'd left mine stock for all of about 30 minutes).

 

[Grooveriding]

Wait, now I'm confused:

I thought the framelimiter was the thing that leads the card to change clocks according to target fps. Now adaptive vsync does the same?

It's widely game, resolution and settings variant until you get into temperatures about 70C where it progressively reduces the boost headroom by 13mhz. I'd be interested to know if there is an element at play in the drivers on a per game basis because of what I notice in BF3.

There is definitely something more at play because I am pegged at max clocks in most games I play.

For example in BF3 I am always pegged at full clocks even with a frame cap of 65 unless I am loading in or, at times, on the respawn screen. My GPU usage with the frame cap is maybe 60-80% varying. If I turn off the framecap I get framerates above 65 to as high as 80-100+. So the cards have more headroom but are still pegged at max clocks even below that cap. The same situation plays out in Skyrim using 4xmsaa/4xsgssaa and texture packs.

Some of the information says the hardware inside the GPU monitors more than just utilization, but also memory usage. Perhaps that is why I see this behaviour.

http://www.techarp.com/showarticle.aspx?artno=745&pgno=1

Kepler-based GPUs have dedicated hardware circuitry that constantly monitors various aspects of the graphics card, from the GPU's power consumption and temperature to the actual GPU and memory utilization. All that information is processed by the GPU Boost software algorithm to determine what changes, if any, should be made to the GPU and memory clock speeds and voltages.

Nvidia should give out a whitepaper giving more detail on how the boost works.

 

[Ferzerp]

I may be basing the clock down effect on a single game and may be mistaken and it's just that it does it in that single game for some reason.

 

[Ben90]

If you make your card work hard enough, for example render a very difficult scene that i.e. only runs at 10 fps and pegs your GPU usage at 99%, your card will stay pegged at 1006 Mhz because of high power and temperature. If on the other hand you render an easy scene that runs at 60 fps capped and 50% GPU usage, Boost will raise the clocks as long as it has the power and temperature headroom to do so. Here you have an example of GPU boost doing absolutely nothing useful for you.

Your opinion regarding boosting in this scenario is extremely valid and I completely agree with it. I do have a slight issue with how it was presented though. Generally the less effort the front end exerts, the more actual load gets placed on the core. I'm not talking GPU-z load numbers, but actual amount of work that gets done (and the heat that gets generated with it).

A lot of programs are designed to be as easy on the front end as possible. The most notorious of which is Furmark. No matter the framerate, the front end is basically telling the cores, "Hey remember what you just did? Do it again." Over and over and over again. A lot of programs get programmed this way on accident. Starcraft 2's menu screen being a great example.

A game engine on average is a decent amount tougher for the front end to schedule. Everything in general is just a little more complex with some mixing-it-up-action going on. One microsecond it could be tessellating a tree, the next drawing the shadow of the tree, the next applying post processing effects to the shadow. The front end needs to work a lot harder to keep the cores running smoothly.

If your game drops down to 10 frames per second from lets say 30, it means either there is a crapload of detail all of a sudden (explosion), or something stalled really bad. Most likely, both happened at once. While the front end is trying its hardest to keep the cores fed, a lot of the scene cannot get worked on until we receive our explosion texture from main memory. GPUz reports 100% load, but the cores really aren't doing that much work since most of them are stalled.

You can give millions of examples for a counter-point. But on average the longer it takes for a scene to get rendered, the GPU uses less power per unit of time than that same engine running at twice the framerate when both are 100% GPU bound.

 

[JAG87]

I see in your sig you have one of these. It's obvious you haven't bothered actually watching the clocks because everything you've posted is pure fabrication with no basis in how the function actually works in practice.

I have watched the clocks and that's exactly how it behaves. Just because you set the power to 132% and forgot that it affects how GPU Boost works, doesn't mean you know what you're talking about.

Try setting it at 100%. If you can keep the GPU usage at 99% long enough, you'll fall right back to 1006 Mhz, or whatever you card is capable of without going over 100% TDP (not all GPUs leak power the same, it is affected by many things, temperature being one of the greatest variables).

Your opinion regarding boosting in this scenario is extremely valid and I completely agree with it. I do have a slight issue with how it was presented though. Generally the less effort the front end exerts, the more actual load gets placed on the core. I'm not talking GPU-z load numbers, but actual amount of work that gets done (and the heat that gets generated with it).

A lot of programs are designed to be as easy on the front end as possible. The most notorious of which is Furmark. No matter the framerate, the front end is basically telling the cores, "Hey remember what you just did? Do it again." Over and over and over again. A lot of programs get programmed this way on accident. Starcraft 2's menu screen being a great example.

A game engine on average is a decent amount tougher for the front end to schedule. Everything in general is just a little more complex with some mixing-it-up-action going on. One microsecond it could be tessellating a tree, the next drawing the shadow of the tree, the next applying post processing effects to the shadow. The front end needs to work a lot harder to keep the cores running smoothly.

If your game drops down to 10 frames per second from lets say 30, it means either there is a crapload of detail all of a sudden (explosion), or something stalled really bad. Most likely, both happened at once. While the front end is trying its hardest to keep the cores fed, a lot of the scene cannot get worked on until we receive our explosion texture from main memory. GPUz reports 100% load, but the cores really aren't doing that much work since most of them are stalled.

You can give millions of examples for a counter-point. But on average the longer it takes for a scene to get rendered, the GPU uses less power per unit of time than that same engine running at twice the framerate when both are 100% GPU bound.

Yes I agree with what you said, but it's not really a problem for anybody running with frame limiter or target enabled, which you always should.

In that case, it's the harder scenes that have the greatest workload, not the easy scenes repeated at 1000 fps.

 

[felang]

That has nothing to do with GPU boost, it is controlled by the Adaptive vs. Maximum Performance power setting in the CP, in combination with an fps cap and or vsync. Older cards work the same way as well.

GPU boost is snake oil for the consumer, it is designed to increase the average fps in benchmarks for nvidia. It does nothing to address minimum frame rates (stressful rendering = high power consumption and temperatures). If you make your card work hard enough, for example render a very difficult scene that i.e. only runs at 10 fps and pegs your GPU usage at 99%, your card will stay pegged at 1006 Mhz because of high power and temperature. If on the other hand you render an easy scene that runs at 60 fps capped and 50% GPU usage, Boost will raise the clocks as long as it has the power and temperature headroom to do so. Here you have an example of GPU boost doing absolutely nothing useful for you.

Currently nobody can go above 1.175 volts, so no matter how high you set the core offset, nobody is reaching 132% power threshold... so your example doesn´t exist in cases where the power threshold is increased.

As long as you keep under 70 Degrees and 132% max power threshold the card will never underclock from whatever you have the cpu offset set to + max auto overclock for that specific card.

In my case, my card stays pegged at 1215 Mhz (+100 core offset)as long as it is under 70 degrees, after that it seems to lower 13 Mhz for each 5 degree increase in core temp

 

[JAG87]

Currently nobody can go above 1.175 volts, so no matter how high you set the core offset, nobody is reaching 132% power threshold... so your example doesn´t exist in cases where the power threshold is increased.

As long as you keep under 70 Degrees and 132% max power threshold the card will never underclock from whatever you have the cpu offset set to + max auto overclock for that specific card.

In my case, my card stays pegged at 1215 Mhz (+100 core offset)as long as it is under 70 degrees, after that it seems to lower 13 Mhz for each 5 degree increase in core temp

That's right, which is why if you want to understand how GPU Boost behaves, you need to leave it at 100%, not 132%. And nobody running a semi-stressful game will keep it under 70C on air, unless you're blasting the fan at 85%.

 

[Ferzerp]

I have watched the clocks and that's exactly how it behaves. Just because you set the power to 132% and forgot that it affects how GPU Boost works, doesn't mean you know what you're talking about.

Stock it just sits at 1098 all day long. Of course you raise the power allowance if you OC it. That would be silly to try to OC it with the default power limit.

Basically you're building a strawman to knock down (No, guys, if you allow it more power when you OC it you're doing it wrong! You should really make sure to limit your card as much as possible, then the clocks won't be consistent).

 

[Tempered81]

card will struggle in some scenes (30 FPS) and thrive in others (90 FPS), it seems that the dynamic overclock will activate in those less stressful scenes, right?

If you try to avoid the marketing slides from NV's press deck on the definition of "boost" and analyze it, you will see that it's geared at maximizing bins. The board components and the chip all come at variable performance levels. The chips have different leakage levels, default VID states, and efficiency metrics (all within the binned range to qualify as GK104-400 samples). The final product of chip package paired with sample board together have a goal of drawing power through the 6+6+pci-E, excluding the memory asic draw and memory phases, and for that power to fit within an acceptable range.

When the board hits its stock defined clock of 1006 and 1058mhz to meet the "gtx 680 qualification requirements" it has some headroom left which can vary depending on sampled combination of board + pwm + asic. That leftover headroom is wisely capitalized in the variable amount of boosting the final product does to maximize the performance of the 680 product within the TDP of the board. ASIC sample quality, efficiency and leakage variations are why different 680's boost to ~1060 and some to ~1270. Not everyone's GK104 boosts the same.

Your statement is exactly what I had in mind, thinking "it's only going to boost when the gpu is not stressed hard". But since all samples have a nominal voltage of 1175mv in their loaded clock states, the boost maximizes performance of various samples. Look at it as: You're guaranteed 1056, high clocks represent the lottery factor of higher quality bins.

The +offset and +power control the designed 170W and 195W qualification and limitations, but the principle still applies that the "boostclock" state will continue increasing core frequency to the maximum within the desired TDP.

 

[felang]

That's right, which is why if you want to understand how GPU Boost behaves, you need to leave it at 100%, not 132%. And nobody running a semi-stressful game will keep it under 70C on air, unless you're blasting the fan at 85%.

Ok, I get your point, but I think it´s worth mentioning that by increasing the power threshold you can basically nullify GPU Boost. I don´t understand why people are making such a big deal about it to tell you the truth.

Also, about the fan noise, the GTX 680 at max (85%) is a lot quieter than my old HD5870 at 45%. Actually I have an Antec Ninehundred which is quite noisy in itself and at 75% I struggle to hear the GPU fan over the case fans (Antec tricools) set at 40% via fan controller. Granted, I wouldn´t call my setup quiet at all... but it is possible to just set the fan at 75% to keep temp under 70 without the noise being "unbearable". IMO though, the increased noise is not worth avoiding the measly 13Mhz downclock, unless you´re benching or whatever.

 

[Ferzerp]

Just checked. Skyrim either runs 1006 Mhz/no boost (in doors and lower complexity scenes) or 1215Mhz (outdoors with more complexity), so it's not just wow. If it doesn't need it, it doesn't use it. If it needs a little more juice, looks like it will boost somewhere in between as well. So, very complex indoor scenes run slightly higher.

 

[boxleitnerb]

What TDP/GPU usage are you getting indoors and outdoors? I thought Skyrim was GPU bound indoors and rather CPU bound outdoors. At least it is here (when you let it by disabling VSync).

 

[Ferzerp]

What TDP/GPU usage are you getting indoors and outdoors? I thought Skyrim was GPU bound indoors and rather CPU bound outdoors. At least it is here (when you let it by disabling VSync).

GPU usage is pretty similar 50-60% in both scenarios. I'd have to reboot to get power usage since it bugs out and the display doesn't change from 30% sometimes. Can't do that now as the TV is being watched in the living room, and this is the media center server with the cable tuner in it.

Temps, even outside at max clock speed don't rise above mid 60's in Skyrim for me, so 55ish% fan speed.

Using the high res texture pack and inside memory usage is around 900MB, outside 1400MB.