NVIDIA GeForce GTX 1080 SLI

[efikkan]

Right, all things I'm "dreaming about" assume some hypothetical new architecture ... in this case specifically, a data sync for the next frame that doesn't disturb the rendering of the current frame - sync that happens independently while the current frame is being rendered to hide at least the part of that latency ... possibly through nvlink

The thing that bothers me about multi-GPU is not only that we actually need it, but also that it's something that should be solvable. As mentioned multi-GPU scaling is up to the games, but all the stuttering problems are up to the drivers/hardware. NVLink would be nice, a big NVLink SLI bridge would be awesome. Still we would need more speed than both 1st generation NVLink and PCIe 3.0 provides in order to push the transfer times of the final frame down to <0.1ms, so that wouldn't happen any time soon.

I do have a "solution" that will work though without any hardware changes; create a new "AFR mode" which uses the primary GPU only for display, and all the rendering is done by "slaves". At least then the latency will be constant, but there still will be some latency.

Back to my previous point that we need multi-GPU. The last ~3 years or so the gaming market has changed from gamers targeting 1080p/60 Hz to 1440p/120 Hz and soon 2160p/144+ Hz. Even with the great improvements of Kepler, Maxwell and Pascal, gamers are not able to get the performance they want. We are not going to get to the point where a single GPU is "fast enough" in the next two GPU generations (Volta/post-Volta), so there is a place for multi-GPU. The problem with multi-GPU is essentially the chicken and the egg problem, but I'm pretty sure that if multi-GPU were "stutter free" and scaled well in most top games, then a lot of people would buy it. At this point it's more or less a "benchmarking feature", people who want smooth gaming stay away from it. (Multi-GPU works fine though for professional graphics where microstutter doesn't matter...)

I don't know about that, going to 4k seems to need less AA and more temporal AA to reduce shimmering with camera movement

You are right that shimmering/flicker is a problem for camera movement, especially when there are a lot of objects moving only slightly (like grass waving in the wind). TXAA combines MSAA with temporal filtering and previous pixels which essentially blurs the picture, which kind of defeats the purpose of higher resolution in the first place, since it removes the sharpness and detals. The reason for adding this technique is proper AA is too expensive for higher resolution. The advantage of SSAA(supersampling) and MSAA is that it reduces the aliasing while retaining the sharpness. The problem with all types of post-process (or semi post-process) AA techniques (incl. TXAA) is that they work on the rendered image, and can essentially just run different filters on existing data. Proper AA does on the other hand sample the data at a higher resolution and then average it out, essentially just rendering it at a higher resolution. SSAA is the best and most demanding. Even 2x SSAA almost quadruples the load, essentially rendering a 1080p image in 4K and scaling down the result. You might understand why this gets expensive in 1440p, 2160p and so on...

Oh yes, to successfully cover all scenarios, you'll need to split frame buffer into tiles (similar to what Dice did in Frostbite engine for cell implementation in PS3) and distribute tile sets with similar total complexity to different gpus ... which would increase cpu overhead but I think it's worth investigating for engine makers ...

I'm quite serious, I'm very familiar with landscape rendering. Most games renders the terrain from a first person perspective, which will make some regions quick to render and some slow. If you split vertically on the middle, you'll end up in a situation where both GPUs are not well saturated at any time, and continuously adjusting the split is not a good solution either. This is the reason why split frame rendering has been abandoned for games. Why use it when AFR scales so much better?

maybe even use integrated gpu (or one of the multi gpus asymetrically using simultaneous multi-projection) in a preprocessing step only to help speed up dividing jobs for multi gpus based on geometry and shading complexity (on a screen tile level, not on the pixel level).
Agreed

I'm just going to address this quickly. Offloading some simple stuff to an integrated GPU is of course possible, but the latency would not make it worth the effort. The integrated GPU will only perform like 2-5% of a powerful one anyway, and you have to remember that every MB of transfer between them is expensive. If you want to offload some preprocessing, physics, etc. then the result needs to be as compact as possible. If you need to transfer like 250 MB in each frame, then you'll get like 20 FPS and huge latency problems

 

[Steevo]

HB bridge scale well in firestrike benchmark.

http://www.vmodtech.com/th/article/nvidia-hb-bridge-for-pascal-gpus-review/page/2

That game is awesome to play.......


But I prefer games with more play time and interaction than sitting and stroking my penis and looking at numbers.

 

[Enterprise24]

That game is awesome to play.......


But I prefer games with more play time and interaction than sitting and stroking my penis and looking at numbers.

I just show that in synthetic benchmark it has some scaling. No need to troll.

 

[BiggieShady]

I'm quite serious, I'm very familiar with landscape rendering.

So far there is nothing to prove otherwise, so no worries.
Consequently you should quite seriously read how Dice cleverly divided deferred rendering jobs for all the SP in the cell processor in old PS3 version of Frostbite engine: http://www.slideshare.net/DICEStudio/spubased-deferred-shading-in-battlefield-3-for-playstation-3
They did it for different reasons ... to help the weak gpu with the lighting pass (multiple lights + global illumination light probes).
I'm suggesting a research following the same principle, not using exact same solution.
No need to get stuck with either horizontal nor vertical frame split in two pieces, when engine workload for a single frame is already split into passes and when tile based approach may offer better granularity for balanced job division, for example:

these are the tiles for the lighting pass rated by calculation complexity. Every pass could work similarly. Essentially doing the split frame rendering by splitting the workload rather than frame .

 

[enya64]

If the new sli bridges eliminated microstutter, then I wouldn't care if there was no improvement in frame rate. But reading other forums, sli users are still dealing with the same level of microstutter as before.

 

[Breit]

To end all the discussion about when those bridges are needed and when not, Nvidia should make the connection info about those bridges available through the driver as they are doing with all the other stats that can be tracked in realtime, like core/mem clock speed or GPU utilisation. I'm thinking of clockspeed, bandwidth and utilisation of the SLI bridge and maybe even provide a setting to change the frequency of the SLI connectors manually in the driver.

 

[Aquinus]

So, even if you solve the bandwidth problem with sending frame data to the GPU actually displaying the frame, you'll still have this weird latency problem where the frame coming from the local GPU will always be ready to be rendered before the secondary GPU. The funny thing is (when it works,) running a third GPU can sometimes reduce the amount of micro stutter, probably because the two non-primary GPUs have the same amount of latency to communicate with the primary. So consider this, to get a frame from one GPU to another is measures on the scale of milliseconds where the transfer of a frame on the local GPU is literally measured in nanoseconds. The primary will always get its frame to the active frame buffer for the display before a secondary or tertiary GPU will. This latency alone is the source of most of the woes with micro-stutter. The only way to solve that problem is to normalize the rate at which frames and produced and received.

Micro-stutter gets worse when you start pushing a GPU to its limits because the GPU can't get the frame over to the primary as quickly as if it were running at 50% and is getting the frame to where it needs to be earlier so, when you're running at well over 60FPS, using something like v-sync makes it possible to buffer frames and essentially smooth out the latency with the frame times.

With that said, if you had a device that didn't do the rendering but, handled only the display output and buffering for a monitor or monitors would be the solution. Equal latency between the devices that produce the frames to the thing that consumes the frames is what will give us the best experience but, so long as one GPU is responsible for outputting that frame, they're going to have real fun trying to get the two sources of frame data to maintain a consistent latency between rendered frames to be displayed.

tl;dr: The less you tax a multi-GPU setup, the less likely it's going to cause micro-stutter issues because the latency between frames will become more consistent.

 

[Steevo]

I just show that in synthetic benchmark it has some scaling. No need to troll.

It was more of a jab at the only reasons to have/use it is for epeen and benchmarking, so its essentially useless for 99.999999999999999999999999999% of users.

 

[enya64]

Okay I drove to newegg yesterday and picked up my two gtx 1070s and an MSI hb sli bridge, just in case there was some future benefit. Every game that I tried (Witcher 3, Project Cars, GTA 5, Mortal Kombat X, The Evil Within, Mad Max, Fallout 4) ran ABOVE 60 fps in 4K at max/ultra settings. The exceptions were Crysis 3 which hovered around 53 fps and GTA 4 which averaged 53 fps in 4K.

The 1070s are gigabyte g1 gaming cards and the overclock software I used to push these in sli is EVGA Precision X. To overclock I simply increased the power target to the max of 111%, the core clock to in increments of +25 to land at +75 ( +100 caused the Heaven Valley Benchmark to crash ), set a max temp of 81 degrees, and I haven't touch the memory clock yet. This land me at a stable clock speed of 2078 with no fluctuation during gameplay at a max temp of 72 degrees. I'm still testing because I've only had these cards for a couple of hours, but it is a huge improvement over my two 970s and it so happens that the games I play have sli profiles.

I didn't notice any microstutter, but because I was testing with vsync off on all the games there was slight screen tearing when panning the camera too quickly in a few of the games. It happened far less than with 970 sli and was less noticeable in comparison due to frame rate being above 60 compare to the previous 35fps average of the 970 sli. I've noticed through my different builds over the last 2 years that gaming in 4K introduces a higher probability of screen tearing that is more pronounced the lower the frame rate, which is solved by vsync at the cost of performance . 60 fps almost drastically lowers the visibility and occurrences of screen tearing without the need for vsync.

I'll keep testing and post any updates.

 

[InquisitorDavid]

Long time reader, first time member here.

Wanted to share some very interesting findings in this vid by Hardware Unboxed:

Big gains can apparently be had from the HB Bridge vs old Flex Bridge depending on the title (Fallout 4, Black Ops 3, Witcher 3). Thing is, the HB results could be replicated by putting 2 Flex Bridges together, so the gain is not from the HB Bridge itself, but from the linking of both SLI connectors. Can TPU confirm their findings by running those games under similar conditions? Noted the review had BO3, but only for 1600x900.

 

[Caring1]

Big gains can apparently be had from the HB Bridge vs old Flex Bridge depending on the title (Fallout 4, Black Ops 3, Witcher 3). Thing is, the HB results could be replicated by putting 2 Flex Bridges together, so the gain is not from the HB Bridge itself, but from the linking of both SLI connectors.

It's already been stated flexible bridges do not have the same gains, even if two are used.

 

[InquisitorDavid]

Stated where?

It seemed to be game-specific, so I wanted to get verification on this.

 

[Caring1]

Stated where?

It seemed to be game-specific, so I wanted to get verification on this.

Pretty sure W1zz confirmed it in one of the threads discussing the 1080's.

 

[trog100]

It's already been stated flexible bridges do not have the same gains, even if two are used.

i tried using two flexible bridges.. more for cosmetic reasons than anything else.. two look better than one.. he he

sadly it didnt work.. it produced very pronounced interference lines across the image.. i am guessing some kind of cross chatter between the two cables was taking place.. i had to revert back to just the one..

its a shame about the lack of game support.. having said that i have two of the games listed as not using the second card.. just cause and primal.. i didnt know they lacked sli support until i read about it.. they played just fine.. mind you i dont run 4K..

trog

 

[Silence48]

I am wondering if I need to invest in a high bandwidth SLI bridge. I ran Unigine heaven 4.0 benchmark on my 1080 ti SLI setup and the results seem to leave something to be desired even though they are pretty awesome.
Can you tell me what you think? You can see the video here.

 

[Caring1]

I am wondering if I need to invest in a high bandwidth SLI bridge. I ran Unigine heaven 4.0 benchmark on my 1080 ti SLI setup and the results seem to leave something to be desired even though they are pretty awesome.
Can you tell me what you think? You can see the video here.

You may get a slight increase, but better of waiting for drivers to be optimised first, as the 1080Ti is still in it's infancy as a card, especially in SLI where you might only see a small increase in performance.

 

[GhostFace0621]

I have no complaints with my GTX 1080 FE SLI setup. I've got a 1440Pmonitor and playing games at higher refresh rate with higher resolution? It just looks amazing. Also, I think that the sweet spot for gaming right now is still 1440P rather than 4K.

Also, I can't wait for explicit mGPU support to start rolling out on a lot of DX12 games. I hear that it'll utilize your GPUs better than your traditional SLI/Crossfire setup. I mean it would make sense, with explicit mGPU not needing the drivers to utilize both GPUs.

 

[trog100]

I am wondering if I need to invest in a high bandwidth SLI bridge. I ran Unigine heaven 4.0 benchmark on my 1080 ti SLI setup and the results seem to leave something to be desired even though they are pretty awesome.
Can you tell me what you think? You can see the video here.

turn one of the cards off and see what happens.. quite what part of the video you think looks awesome has me a bit baffled it all looks like a jerky load of crap to me..

trog