NVidia SLI Guide

Because for TAA you need a previous frame data to perform TAA and with AFR the previous frame is actually being rendered in parallel with the current one. Any temporal technique will kill AFR rendering approach unless it's aware of the AFR rendering and modifies itself accordingly - which most games don't do since their main tech target are consoles and single GPU PCs, not the ~1% mGPU market.

 

This is simply not true, with enough PCIe bandwidth or custom SLI profiles many modern DX11 games scale relatively well with even native 4K + TAA.
And in the future NVIDIA NVLink might entirely solve the problem.

 

Only if you buy the high bandwidth NV link bridge j/k

 

This has nothing to do with bandwidth.

 

Good SLI scaling with high native resolutions combined with TAA has everything to do with PCIe bandwidth these days.
Certainly good AFR friendly games are also a prerequisite, but the difference between PCIe 3.0 x16/x16 and PCIe 3.0 x8/x8 can mean a good 70-80% positive scaling versus single GPU performance in some UE4 games.

 

I'm curious, since you seem well informed about this, is there any article on the subject of SLI scaling with PCI-E x16/x16 and x8/x8 you'd recommend? I'm talking about something that benchmarks specifically these types of games where you say there's a very significant difference. Testing with and without a HB bridge on Pascal would be nice to see as well. I'm only asking because I can't find anything myself and I'm running 1080 SLI, 4K, but with an "old" 4790K, so it's pretty relevant to me.

 

Blaire actually did extensive testing in the 3DCenter SLI thread.
This is in german, but benchmark numbers are benchmark numbers
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=11059897&postcount=2261

Most graphics card reviewers skip this part of the testing, since they simply don't know about the bandwidth limitations IMO.

 

Thank you! Luckily for me I can speak German too. Those results are very interesting and at the same time downright depressing, a 6850K with 40 lanes is almost $800 where I live, I don't much like the idea of having to drop that much on a CPU alone, just for some PCI-E lanes too.

Sadly the only game I own out of the bunch is Witcher 3, but I have to say I can't replicate the (very poor) results that were shown in that benchmark with PCI-E x16/x16 2.0, which should be equivalent to my x8/x8 PCI-E 3.0 in terms of bandwidth. I haven't done any specific benchmarking since TW3 doesn't have any built-in tool to provide reliable results AFAIK, but the performance difference is clearly visible, I'm getting ~57-70FPS with SLI enabled and ~33-42 with SLI disabled, in a quick run in Toussaint around Crovo Bianco. The higher numbers are probably due to the fact that I have aftermarket 1080s running at ~2GHz, but in any case I don't see such poor SLI scaling with half the bandwidth. This is with a HB bridge. I have also tested with 1 and 2 old flex bridges: using a single one essentially kills any scaling, but with 2 of them connected the performance is the same as with the HB bridge, so Pascal definitely makes use of both links. I don't have any LED bridge to try.

Anyway, I hope Zen is decent and comes with enough PCI-E lanes, as I really don't want to dump $800 into Intel's pocket for a 40-lane CPU in my near future.

 

Well, please keep in mind that these benchmarks are done with native 3840x2160 resolution (on a "real" 4K monitor) combined with temporal AA.
This should not be confused with 1080P + 4xDSR or 1440P + 2.25xDSR, which both produce the same 3840x2160 resolution but uses far less PCIe bandwidth.

(1920 x 1080) x 4 = (1920 x 1080) x (2 x 2) = 3840 x 2160
(2560 x 1440) x 2.25 = (2560 x 1440) x (1.5 x 1.5) = 3840 x 2160

However then again native 3840x2160 also produces much higher sharpness and image quality than just the DSR equivalents.
Also in the case of Fallout 4 and Witcher 3, the bandwidth limitations with 4K + TAA can be fixed with custom SLI profiles.
With Watch Dogs it is necessary to "only" use MSAA instead of TXAA at native 4K resolution, to avoid running out of bandwidth with less than PCIe 3.0 x16/x16.

Fallout 4
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=11203814&postcount=2396

The Witcher 3
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=10674669&postcount=2008

 

I have a native 4K monitor too and played through Blood & Wine at native 4K. There were some drops under 60 with HairWorks enabled and maxed but nothing close to a 47FPS average, I played with VSync enabled and the game ran at 60FPS the vast, vast majority of time. I always had in-game AA enabled, which AFAIK is temporal in TW3.

This is a screenshot with TW3 maxed that I took just now, no custom SLI profile on 375.70

As you can see my FPS is way beyond the 48FPS max reported with PCI-E x16 2.0 on 3dcenter. Just riding around Toussaint and Beauclair I haven't seen it drop under 55, though with HairWorks enabled I know getting in some fights can definitely push it below that (I played with HW off).

I wish I had some of the other games to test with as this is making me quite curious, though not to the point where I feel like buying them.

 

Just use "0x280021F1" profile in Witcher 3
This is in the above link.

That is enough to solve these bandwidth problems with 4K+TAA in TW3.
Regarding that 48FPS result, that was achieved with the official profile using a custom benchmark.
Don't worry about the actual numbers, just the differences.

And you do know, that it is possible to have 99% usage on both GPUs and still be bandwidth limited?

 

I just tried it out, I'm getting 1-2FPS extra, but it does seem to be consistent. Still, no major change in any case. A bit of free performance is always nice though.

 

Interesting, perhaps there has been some improvements regarding bandwidth in the latest drivers then

 

Yeah, maybe they did change something. Just to be clear though, when you mention TAA in TW3 you mean the built-in "Anti-aliasing" postprocessing option, right?

 

Correct, it is the only AA option in Witcher 3.
Actually it is some type of custom temporal FXAA mode.

 

Yeah, that's what I thought. I always ran the game with in-game AA enabled since disabling it never showed any real performance increase in my case. I remember when I had CF instead of SLI (on the same system otherwise) that it didn't really work too well at all and I always kept it disabled.

About the driver version - since TW3 was the first game I played after upgrading to 1080 SLI I was able to check which driver I had downloaded at the time, it's 368.69. When I played through Blood & Wine I was using 368.69 and was getting pretty much identical performance. If NVIDIA somehow drastically alleviated the b/w bottleneck it happened between 368.39 (3dcenter results) and 368.69.

 

Anyone got a recommended SLI Bits config for Planet Coaster? Theres alot going on this game and im only getting 22-30fps most of the time.

BTW on Witcher 3 I found that "TheWitcher3 SLI Performance Fix [Updated-September2016] .zip" works better then the "TheWitcher3_2-WaySLI-Fix.zip" get a good 5-10fps extra on a 4k ultra config i run on the 970s

 

This is also the one linked in my spreadsheet
"0x280021F1" is needed to fix all slowdowns, where the older "0x280020F1" profile still causes slowdowns when fighting certain monsters using the 368.xx and later drivers.

As far as Planet Coaster, no idea about this one.
Did you try the standard profiles such as "0x080000F1", "0x080020F1", "0x080040F1"?
I'm assuming this is a DX11 title?

 

Yes. It uses the Cobra Engine which is used by Elite Dangerous. Going to try 0x000000F5 Edit: 0x80000F1 works amazingly almost too good to be true. 55-60fps where it was 22-30fps before. Like perfect scaling.

 

Sounds excellent
Please check for any visual glitches, flickering or other problems though.
If you make a couple screenshot comparisons with FPS numbers, I would love to add it to the spreadsheet.