Does SLI add latency?

ManuelG. Does SLI add latency?

I guess this is mostly for ManuelG, but anyone who knows is more than welcome.

Respawn has, unofficially, claimed that there's no sli support yet for Titanfall because of all the latency that sli adds, which Respawn "isn't a fan of". They have already fixed Crossfire support, which would imply that the latency sli adds must be relatively severe?

I can't permalink the post, but it's 5 up from the bottom of the page: http://www.respawn.com/forum/viewtopic.php?f=24&t=13776&start=100
This is, apparently, a convo with a Respawn Dev explaining why sli is still broken.

Having heard some of their excuses as to why the game is in the state that it's in, I'm ready to accept this as horse****, but was wondering if there's any truth to it?

 

Yes, SLI does add a frame of input latency.

On it's own, you might not notice it, especially if you're running 120hz+ but as part of the input/display chain, (input device, OS, software, GPU, monitor) little things can add up.

I am very sensitive to these things and while I haven't actually tried SLI, I avoid it because of this.

 

SLI introduces a bit more frame latency because one GPU has to wait for the second GPU to process all the information.

 

I would say yes. I had a problem on far cry 3 you would get a small glitch every so often tried all sorts to get rid of it no joy until I disabled sil. Fixed the problem with just one card running.

 

Both SLi and CrossFire have an additional frame of input latency since they use AFR, Alternate Frame Rendering. Each card is still rendering a frame in the same amount of time, but you double your framerate by pacing the shown frames such that the second card starts rendering the next frame while the first card is halfway through rendering the first frame, and they flip.

In single GPU scenario:

1) Input is provided then polled just before the frame is rendered, and the result is displayed at the end of the first frame. You get a best case input-frametime latency of 1 frame.

2) Input is provided just after the frame starts rendering, so it is taken into account in the next frame, and the result is displayed at the end of the second frame. You get a worst-case input-frametime latency of 2 frames.


In multi-GPU:

1) Input is provided then polled just before the frame is rendered, and the result is displayed at the end of the first frame. You get a best case input-frametime latency of 1 frame.

2) Input is provided just after the frame starts rendering, so it is taken into account in the next frame that is rendered by the second card starting halfway through the frame rendering of the first card. The result is then displayed at the end of the second frame. You get a worst case input-frametime latency of 1.5 frames.


You might be thinking now that latency has indeed improved. But notice that in the multi-GPU scenario, FPS is double that of the single GPU scenario BUT the frametimes, as we have mentioned previously, are the SAME.

Let's assume the single GPU is rendering at 30FPS. Frametimes are 1/30 = 33.33ms.

1-2 frames (33.33ms) of delay at 30FPS vs. 1-1.5 frames (33.33ms) of delay at 60FPS ----> 2-3 frames (16.67ms) of delay at 60FPS.

Let's clock down both GPUs to match the performance of a single GPU. Now that they have the same FPS, it's 1-2 frames vs. 2-3 frames. There's your additional frame of lag.

Conclusion is that a multi-GPU setup that is as fast as a more powerful single GPU, producing the same framerate, will have 1 additional frame of lag. Raise your FPS and that frame of lag becomes less noticeable. If you're using multi-GPU at 30FPS, then...that's not a very good idea as you now have 33.33ms additional delay over and above the huge delay existent at 30FPS already.

Remember that the input lag chain is much larger than most expect.
http://www.blurbusters.com/gsync/preview2/

 

It depends more on how the PCIe lanes are distributed on the motherboard - and if a PCIe multiplexing chip has been implemented.

Back in the early SLI days when you could only run SLI on Nvidia chipsets (Nvidia 780i, 750i and Intel's P55 and Skulltrail D5400XS chipset), these motherboards had a bridge chip (PCIe lane multiplexer) on it called the N100 (Intel D5400XS) or N200 (780i, 750i, P55).

The N100 only allowed 2-way SLI and the N200 allowed 3-way SLI.

Some people claimed that x58 motherboards with the N200 performed worse in SLI than those without.

The main point being that you only need the N200 chip for 3 way SLI and many people only ran 2 cards - so who wants the extra latency the N200 brings?

All modern Intel boards (e.g. X58. Z67, Z77, X79, Z87) with a PLX bridge chip (PEX8608/PEX8747) will have a small amount of latency.

So Yes it adds latency, but you can buy a motherboard without a PCIe multiplexer to reduce the latency.

In my experience it's not enough to get you killed in an FPS, and the extra frames outweigh the small amount of latency.

 

That latency is a different form of latency altogether. It's data transfer latency and slows down the data that reaches the cards. It does NOT add frames of latency which is inherent to a multi-GPU setup even in ideal conditions.

You do know those chip latencies are imperceptible and do not show up as video latency, right?

 

Whether sli introduces latency or not, its not enough latency to ruin your gaming experience, very much far from it.

 

I can see it adding latency if your doing AFR ,we can see frame variance goes up in general with SLI .

Now I wonder if Nvidia did true SLI, like 3dfx did would this improve this as each card renders only portion of each frame .

 

You know what I'm not a fan of Respawn Entertainment?

Weak as piss excuses given in lieu of proper support for a AAA PC title.

It's totally inexcusable at this point, especially given that SLI support used to be a hell of a lot better than it is now.

Sadly, TF was a totally unremarkable and overhyped game and adding SLI support for it at this point is probably pointless as not many people play it anymore.

 

So if simply having 2 cards adds latency, a frame, then is the excuse legitimate that sli is still broken when crossfire isn't? They managed to fix crossfire but blame sli being broken on sli.

 

Excuse is not legitimate, no.

 

Howdy y'all, long time since I've been here.

So, with a multi-GPU setup, if you're running a game at 60fps, this extra data transfer latency wouldn't normally be an issue. However, with the way Titanfall is coded, could this possibly cause a problem with V-Sync and mouse input, seeing as mouse input is affected by framerates (without forcing raw mouse input)?

I haven't messed with SLI since the days of 3Dfx (RIP). I also miss my Voodoo 5 5500, lol.

 

Thank Ubisoft

 

It's been a while since I tried SLI and Crossfire, but they're right about the latency.
As long as the technology works, the added latency can't make up for the gain in smoothness and performance. But if it doesn't work correctly it'll be worse with SLI/CF than without it.
Haven't bothered with it for quite a while though but I'd always prefer a single card to a multi GPU one.
The added costs for the second GPU, stronger CPU, larger PSU & better cooling aren't even part of the equation then

 

Oh, no. This is definitely NOT the case. If SLI, even in AFR mode, worked that way there would be no worthwhile performance benefits.

Actually, in an application that properly exploits SLI, there's an interplay between the software and driver. The software signals the driver when it's done with various resources needed by the next GPU to draw the next frame. To leverage AFR performance, you actually tend to alter the render task order a little bit to make it possible for GPU1 to begin work on the next frame before GPU0 finishes the prior.

Each frame is both a producer and consumer of data in modern games thanks to post-processing effects. What you've described is a scenario where frame n+1 cannot start until frame n finishes, but even with post-processing that scenario is ficticious. By re-aranging when certain data is produced and consumed (as well as completely removing artificial delays such as waiting for vblank), a little bit of overlap in frame begin/end time can be established and this is in-fact why SLI produces smoother and higher framerates (very similar to the pre-rendered frame limit, which increases CPU/GPU paralellism). It unfortunately increases latency variability.

The REAL problem is that with this optimization strategy in place, inter-frame latency becomes more unpredictable. This is actually what I object to latency wise with SLI, constant latency is manageable, but AFR's oddball behavior can be very unpleasant.

 

nice necro.

 

Nice input.

 

Usually bad form to necro a thread that's over a year old proper way is to make a thread and link to this one if you must.Just forum etiquette on every one I've been on.Most forums actually lock them automatically after a certain time.

 

Well he's right, so...yeah.

As jbmcmillan mentioned, it's mostly pointless to revive threads that are 1 year old.