Can I clock my SLI GPUs asynchronously?

[SickBeast]

I've read conflicting information on this.

I would like to run my main GPU at 950mhz on the core, with the second one running at 850mhz or so. I've read that it's advantageous to have the first SLI card clocked higher than the second because the first card does more processing.

The thing is, I have also read that the nVidia drivers will automatically downclock the faster card so that it's equal to the slowest card in the setup.

Do any of you know how this all works?

TIA

 

[apoppin]

Can I clock my SLI GPUs asynchronously?

Yes

Use Precision or Xtreme Tuner

 

[SickBeast]

Yes

Use Precision or Xtreme Tuner

Do you recommend doing this?

My old card seems to run faster at lower voltage than my new one.

Crysis 2 kept crashing on me tonight so I had to raise the voltage to 1.125 to keep the cards stable at 900mhz.

Will doing this cause microstutter or anything like that?

This seems like it's playing with fire and it makes me nervous. D:

 

[CarpeDiem99]

I've tried this before, with 2 gtx 285s, having the one clocked higher in the main slot. I wouldn't recommend it, kept getting crashes with driver in desktop, and sometimes games.

Switched the cards around, no problems. So I always keep the lower clock card in the main slot now.

 

[Zap]

I've tried this before, with 2 gtx 285s, having the one clocked higher in the main slot. I wouldn't recommend it, kept getting crashes with driver in desktop, and sometimes games.

Switched the cards around, no problems. So I always keep the lower clock card in the main slot now.

I would think that a stability problem with the higher clocked card. When it is the primary card, it does more work, yet doesn't get as much cooling due to being sandwiched with the secondary card.

 

[BFG10K]

I doubt this will cause any kind of performance gain as I’d expect the drivers to clock the cards to the lowest common denominator.

But assuming they don’t, the microstutter will be worse compared to two equally clocked cards. In theory anyway.

 

[notty22]

I also fooled with this using afterburner, it allows you to sync the o/c and set them individually. I purposely set one a 100mhz less. 900mhz and 800mhz. The driver does not attempt to change any clocks, but I noticed in games when my clocks are the same, the gpu load on both cards are always with a -+2% of each other.
With different clocks, the slower card, would run at 99%, and the faster clocked card would run at 89%. The driver (this is imo) is load leveling the 2 cards. So essentially your are capped by the slower card. Even though it never changes the clocks, the driver probably does this to avoid (as mentioned) increased stutter issues.

 

[Seero]

I've read conflicting information on this.

I would like to run my main GPU at 950mhz on the core, with the second one running at 850mhz or so. I've read that it's advantageous to have the first SLI card clocked higher than the second because the first card does more processing.

The thing is, I have also read that the nVidia drivers will automatically downclock the faster card so that it's equal to the slowest card in the setup.

Do any of you know how this all works?

TIA

It can be done but it should not be done. Since load are evenly splitted towards cards, if one of them is slower, than you will either experience lag every second frame, or FPS based upon the slower card. There are no performance statistics about the video card in bios on SLI, so the system don't know that the first card is faster than the second and therefore should take more load.

 

[apoppin]

It can be done but it should not be done. Since load are evenly splitted towards cards, if one of them is slower, than you will either experience lag every second frame, or FPS based upon the slower card. There are no performance statistics about the video card in bios on SLI, so the system don't know that the first card is faster than the second and therefore should take more load.

Such shocking misinformation!


Try it. It works great and there are definite advantages (compared to minor disadvantages).

i don't see extra microstutter and the performance increases if you increase just one card's clockspeed. See if it suits your PC. Only testing will tell. It's like a recipe for a cake - you never know until you eat it

 

[Seero]

Such shocking misinformation!


Try it. It works great and there are definite advantages (compared to minor disadvantages).

i don't see extra microstutter and the performance increases if you increase just one card's clockspeed. See if it suits your PC. Only testing will tell. It's like a recipe for a cake - you never know until you eat it

I tried. The following are what I got:
Bad FPS.
video driver crashes.
sever flickering.

btw, which part is mis-informative?

 

[apoppin]

I tried. The following are what I got:
Bad FPS.
video driver crashes.
sever flickering.

btw, which part is mis-informative?

You are misinforming the people reading your posts about this subject. None of the above happens with asynchronous clock speeds in SLI or CrossFire.

Yours is a bad experience and it is not the way it works. The graphics drivers make the asynchronous speed experience generally excellent although there may occasionally be issues in specific games and with conflicts between very specific BIOSes of mismatched Nvidia's partner cards.


i have been using mismatched clockspeeds ever since AMD introduced them - years ago. And Nvidia has been allowing this since GeForce 185.85 almost a full year ago.

It works awesome. i tested HD 4890+HD4870 CrossFireX and HD 4870-X3 Tri-Fire with varying clockspeeds and it worked great back then. i will repeat this again in the near future with HD 6990 + HD 6970 (actually i already have with good results) and with GTX 580 SLI.
:thumbsup:

 

[Seero]

You are misinforming the people reading your posts about this subject. None of the above happens with asynchronous clock speeds in SLI or CrossFire.

Yours is a bad experience and it is not the way it works. The graphics drivers make the asynchronous speed experience generally excellent although there may occasionally be issues in specific games and with conflicts between very specific BIOSes of mismatched Nvidia's partner cards.


i have been using mismatched clockspeeds ever since AMD introduced them - years ago. And Nvidia has been allowing this since GeForce 185.85 almost a full year ago.

It works awesome.
:thumbsup:

I said it can be done, but it should not be done due to the reasons you stated. There are issues in some games and occasional conflicts on BIOSes. Back then the driver shows all card running at the lowest common denominator, it no longer does that, but that doesn't mean it will be stable.

Depending on your level of tolerance, crashing may appears to be scary to some.

Yes, I can have SLI setup by profile games to use single video card. That way it seems to be problem free, but that defeats the SLI purpose.

 

[Rhoxed]

I have used different OC's on my cards for years now. I do not notice any increase in stutter, and it actually does improve performance. - I do this for 2 systems.

 

[SickBeast]

Thanks everyone for the replies.

It's interesting that this thread has given me the same conflicting information that I had before, essentially leaving me just as confused as I was before.

Tonight I'm going to give this a shot. The thing is, I have both cards at 900mhz right now and they run nicely at that speed at 1.1v and 55C under load. I'm thinking about boosting one of them to 950mhz, but to do that it requires 1.2v and it will surely create more heat. My one card ran at that speed for months no problem, but it makes me a little nervous doing it in SLI, especially in light of some of the horror stories on here.

Thanks everyone for your input. I'll see how it goes. I'm going to use Unigine Heaven as my benchmark.

It's interesting in that my first GPU is typically at 100% usage whereas the second one varies between 80% and 100%. Perhaps having the primary card running 20% faster will be "optimal", and I can undervolt the other card a bit to compensate.

 

[apoppin]

Play many games. Not just one synthetic benchmark.

(it is good advice anytime)

 

[notty22]

I don't think everything that applies to Nvida SLI applies to Crossfire also.
They don't scale the same , obviously. Also AMD allows different shader count gpu's to be crossfired, not just clock speeds.

I agree , there is confusion. But I'm certain of what I stated, from my testing.

Here is a comment from TH, during the review of the 6790.
This seems to go against what most people believe concerning AMD crossfire. That 1 faster card (5770 and 5750 for ex.) can help achieve higher fps, in a crossfire pairing.

http://www.tomshardware.com/reviews/radeon-hd-6790-barts-gpu-geforce-gtx-460,2917-14.html

While we have only a single Radeon HD 6790 reference card for testing, we’ve simulated CrossFire performance by pairing it with a Radeon HD 6950. AMD has previously confirmed that cards in CrossFire are limited to the specifications of the lesser card, so we matched clock speeds and recorded the following results:

pretty certain they meant 6850.

edit: Also, I'm not certain that reviewer is correct. I'm just presenting his comment. There is confusion surround things.

 

[apoppin]

I don't think everything that applies to Nvida SLI applies to Crossfire also.
They don't scale the same , obviously. Also AMD allows different shader count gpu's to be crossfired, not just clock speeds.

I agree , there is confusion. But I'm certain of what I stated, from my testing.

Here is a comment from TH, during the review of the 6790.
This seems to go against what most people believe concerning AMD crossfire. That 1 faster card (5770 and 5750 for ex.) can help achieve higher fps, in a crossfire pairing.

http://www.tomshardware.com/reviews/radeon-hd-6790-barts-gpu-geforce-gtx-460,2917-14.html
pretty certain they meant 6850.

edit: Also, I'm not certain that reviewer is correct. I'm just presenting his comment. There is confusion surround things.

i did a lot of exploration of CrossFire-X scaling with HD 4870 (512MB and 1GB versions), HD 4890 and HD 4870-X2 in many configurations. i am not allowed to link to my own articles (PM sent) but i will quote myself from the conclusion of CrossFire-X eXplored:

Here we finally see CrossFire-X scaling exposed. Two “mismatched” cards in FrankenFire do not default to the slower card’s speeds but the slower one generally contributes to the overall performance as the load is balanced between them as well as the drivers are capable with that particular game.

i am preparing to do a follow-up evaluation in June using HD 6990, HD 6950 and HD 6970 all in Crossfire-X combinations as well as SLI asynchronous clocking (i already tested GTX 580 SLI at the same core speed as GTX 590).

 

[SickBeast]

Ok, so I tried it and I was disappointed. I gained 2% in Heaven when clocking the one card at 950mhz vs the other at 900mhz. That's in line with what I would expect from boosting the one core by 5%; half of it is 2.5%.

So, there appears to be no advantage of having the main card faster than the other one, at least from my limited testing.

I'm going to leave the cards both at 900mhz for the time being. I might try for 950mhz on both if I get brave enough.

 

[Sylvanas]

I don't think everything that applies to Nvida SLI applies to Crossfire also.
They don't scale the same , obviously. Also AMD allows different shader count gpu's to be crossfired, not just clock speeds.

I agree , there is confusion. But I'm certain of what I stated, from my testing.

Here is a comment from TH, during the review of the 6790.
This seems to go against what most people believe concerning AMD crossfire. That 1 faster card (5770 and 5750 for ex.) can help achieve higher fps, in a crossfire pairing.

http://www.tomshardware.com/reviews/radeon-hd-6790-barts-gpu-geforce-gtx-460,2917-14.html
pretty certain they meant 6850.

edit: Also, I'm not certain that reviewer is correct. I'm just presenting his comment. There is confusion surround things.

By default the driver sets clockspeeds as per the default clockspeeds of each card, however you can clock them manually since Cat 8.1 and achieve a minor benefit. The architectures have to be similar though ie a xx70 and xx50 are usually based off the same die just with reduced clocks/shaders and generally play nice when Crossfired'd asynchronously. You couldn't go 6970 and 5850 for example.

You are however limited in terms of framebuffer no matter what. A 1GB card and a 512mb card for example, the framebuffer for both will be 512mb so you effectively lose half your ram on the 1GB card.

 

[aka1nas]

I doubt this will cause any kind of performance gain as I’d expect the drivers to clock the cards to the lowest common denominator.

But assuming they don’t, the microstutter will be worse compared to two equally clocked cards. In theory anyway.

IME, SLI has never been smart enough to override clocks on async-clocked cards. In the past, many overclocking utilities weren't sophisticated enough to independently clock the cards when overclocking, which I think has given many the impression that the driver was always syncing them. Several of the popular monitoring utilities at the time were already able to individually report clock speeds dynamically and could be used to confirm that async-clocked cards were not being synced. Using cards with different default clocks would give this scenario, which were also less common until the last several GPU generations (Yes, you could get something like a Geforce 3 Golden Sample model back in the day, but we didn't have every vendor with 3-5 overclocked SKUs for the major card models like we do now).

I also don't see why the microstutter would be any worse. Individual frames aren't going to take the same amount of time to process anyway on either card, and the PCI-E bus isn't being clocked differently, so you aren't doing anything to affect transit latency. I think the difference that you can realistically get in clocks between the two cards would be overwhelmed by that difference in frame rendering times.