Nvidia sli & stutter avoidance presentation from nvidia developers' conference 2015

The Nvidia (GDC 2015) Developers' Conference had an excellent presentation on SLI & stutter avoidance. The pdf of the presentation can be found at the following link: http://developer.download.nvidia.com.../SLI_GDC15.pdf

Valuable reading. The first clear, recent explanation I have come across from Nvidia detailing the pros, cons, differences and possible solutions to the micro-stuttering inherent in SLI configurations.

Regards,

A. Ott

 

Further to the slides, some short takeaways for end-users from the slides:

1. Set maximum pre-rendered frames to n+1 where n is the number of gpus present in the sli setup. For 2-way sli; 3 frames.
2. Use AFR2 rather than AFR 1 in all profiles unless there are image corruption issues. Scaling is better.
3. Set 'Vertical Sync Smooth AFR behavior' to on in the profile. The setting controls inter-frame dependencies to some extent.

The rest of the info is more useful to game devs. The top bits of advice:

1. "Getting testing builds to NVIDIA early (Please ) for SLI profiling, identification issues, advice."
2. "QA SLI on 3-way configuration. Needs profile or AFR FriendlyD3D.exe to scale."

Currently awaiting a response from Nvidia for more detailed information on the GDC2015 presentation.

 

Should we set MAXIMUM_FRAME_LATENCY=3 in graphics.ini as well? What are your thoughts on the other setting in the NVIDIA control panel called Virtual Reality Pre-rendered frames?

 

I left it at 0 to keep conflicts with the existing setting in the Nvidia driver 'Maximum pre-rendered frames=' to a minimum. I followed the Nvidia recommendation of n+1 pre-rendered frames for a 2 gpu sli setup (where n is the number of gpus). I don't know if the AC setting creates an extra buffer based on an internal AC setting or uses the existing Nvidia setting. This may be useful for AMD driver users. It would be worth asking the AC devs. If you use NVIDIA Inspector, you can also set 'Maximum frames allowed=' to 3 to match the 'Maximum pre-rendered frames=' setting in the AC profile. This seems reduces gpu utiliztion much like capping frame rates does according to the monitoring software in EVGA PrecisionX 16.

Nvidia recommends keeping the VR pre-rendered frames setting to "1" to minimize lag. I don't have a DK2 SDK so no personal experience with this. I will wait for the commercial release since Nvidia is working on driver optimizations for VR delivery (specifically the Oculus Rift). According to the slides, dual gpus will be necessary, one for each screen in the VR headset.

 

Does that also mean that for single GPU it's better to have n+1, i.e 2 max prerendered frames instead of 1.

 

I hope that Kunos devs will read this nvidia pdf or even contact the two SLI specialized persons named at the end of the pdf in order to solve all the SLI problems we have in AC (timing bug, micro stutters, overexposure with PP effects on...).

 

Options 1 & 3 introduce too much input lag on my system for them to be viable options.

I've managed to reduce microstutters to an acceptable level via a combination of other settings in Nvidia inspector.

Only real solution I see is some extra love from Kunos but I know they've mentioned before that they don't recommend SLI, so doubt its high on the priority list, especially when there's so much focus on building content that ports across to consoles and their single gpus.

 

...stick the cards on ebay and buy one fast GPU

 

The explanation from Nvidia's 'SLI Best Practices Guide' ( http://developer.download.nvidia.com/whitepapers/2011/SLI_Best_Practices_2011_Feb.pdf ) is as follows:

"By default, Direct3D allows buffering up to three frames‟ worth of work before Direct3D calls will start stalling. Historically, this was done to grant a performance benefit on single GPU systems by allowing overlap of CPU and GPU work. SLI configurations require this buffering of frames to achieve performance scaling in AFR mode."

"The drawback of this buffering is that it introduces additional latency between the user‟s input and its results being visible on screen (often referred to as “input lag”). The overall goal of SLI is to achieve the greatest performance benefit while minimizing perceived input lag."

So you can try reducing the setting to two or one frames to reduce input lag but may end up getting stuttering in the frame rate due to pipeline stalls or cpu-gpu synchronization issues. G-Sync was developed to address this by forgoing the use of v-sync but requires new monitors with the hardware built in. Adaptive sync is a software setting in the driver that attempts the same thing but can cause tearing and other artifacts below 60 fps on a 60 hz monitor.

 

I run two 980 ti's in SLi, but have never noticed micro stuttering. I keep pre rendered frames on 1 or 0 depending on version as input lag reduction is my primary goal. Interesting article thanks for posting it.

 

No need . With tweaking it runs good enough for me plus I spend most of my time in iRacing which is silky smooth.

 

Fair enough, you just said the exact opposite. I'm sure Kunos will be pleased to note they don't need to do anything now.

 

No, I said it was good enough for me.

People are affected to a greater/lesser degree depending on their setup and how sensitive they are to its effects.

 

In which case they should see my earlier post for the only real solution.

 

As an update, the video for the talk and slides referenced above 'NVIDIA SLI and Stutter Avoidance: A Recipe for Smooth Gaming and Perfect Scaling with Multiple GPUs' is now available at: http://www.gdcvault.com/play/1022388/

Regards

 

And here is a link to a more detailed presentation on SLI and the causes of stuttering from GDC2012.

https://developer.nvidia.com/sites/default/files/akamai/gameworks/CN/Stuttering_Analysis_EN.pdf

 

max 3 pre-rendered frames instead of 1? At 60 Hz, that's 32 ms of extra latency for a total of 48 ms.

Not the sort of thing you want to be doing for AC.

The AFR stuff seems relevant however.

Whatever Kunos did with the FPS lock in the last patch has really sorted out frame times it seems. I set 60FPS lock and Vsync and everything runs beautifully. Nice flat red line on the FPS monitor, no tearing.

This is all mainly geared towards developers, but you could learn how to properly identify and diagnose the sources of your gaming stutters and hitches related to SLI with it.

 

If you are technically inclined, GPUview will give very detailed analyses of gpu, cpu and DXxx.x events and interactions. This application and the logs it creates are used by Nvidia and AMD to identify and diganose the sources of stutters. http://graphics.stanford.edu/~mdfisher/GPUView.html
It is still upto the developers to make changes in their code to conform to sli / crossfire coding best practices to reduce the stuttering reported by the diagnostic tools (which they don't seem to use).

In regards to stuttering, the following information from Nvidia on sli specific vsync may be helpful:

"Smooth Vsync is a new technology that can reduce stutter when Vsync is enabled and SLI is active.

When SLI is active and natural frame rates of games are below the refresh rate of your monitor, traditional vsync forces frame rates to quickly oscillate between the refresh rate and half the refresh rate (for example, between 60Hz and 30Hz). This variation is often perceived as stutter. Smooth Vsync improves this by locking into the sustainable frame rate of your game and only increasing the frame rate if the game performance moves sustainably above the refresh rate of your monitor. This does lower the average framerate of your game, but the experience in many cases is far better."

 

In the nvidia control panel, when hovering the cursor over the option for maximum pre-rendered frames the message at the bottom states "This setting does not apply to SLI". So what's all this n+1 about?

I've got mine set to default anyway (use application setting) and MAXIMUM_FRAME_LATENCY=0 in graphics.ini. AFR2 is set via nvidia inspector and I've no problems with stutter whatsoever. I can hotlap at 7680x1440 just fine, but have to dial it down to 5760x1080 to maintain 60 fps with a full grid. Pretty happy with that performance tho', no complaints from me.