Overclock.net banner

NVLInk (NOT SLI)

4095 Views 21 Replies 12 Participants Last post by  Mckayman
Recently, I have rebuilt my rig with used parts. I went from a x58 platform with a xeon 5690x 6c 12t, to a x79 platfrom with a xeon e5 2697v2 12c 24t 30mb Cache. I am an Engineer, so I figured the I would go with more cores and cache for CAD software and rendering.

I bought (2) MSI RTX 2070 SUPERS and then an ASUS ROG RBG NvLink 4 Slot Bridge ( I could not find the MSI bridge ) My mobo an ASUS P9X79 DELUXE has PCIE 3.0 x16 slots in locations that require me to use the 4 slot, the other PCIE 16x Slots do not the fully 16x capacity. ( You can see the pins only go up to 8x in the 16x lane. )

I did some digging on the NVLink, for i knew right of the get go it is not the same type of interface as SLI. My understanding is NVLink is the same type of platform the Quadro Cards have used, and every sales rep at each computer store I went to, told me it was told it was not recommended to buy (2) RTX 2070 SUPERS as for the "SLI" did not have the performance gains than previous generations. After a month of time went buy with only one RTX 2070 SUPER of normal use and benchmarking, I couldn't take it anymore. I had to have two!

So how is it? HOLY SMOKES! The performance gain has been incredible! In 3DMark has shown results of almost 100% gains on the graphics scores! Frame rates for DX11 Skydiver are went from 180-200 to 400+!
DX12 Time Spy are when from 60-75fps to 120-140fp! Time Spy Extreme 4K went from 30fps to 60fps. For the price of 1100 for both cards and the bridge. I would say these results would be equal to if not the same as a RTX 2080 Ti which is still hovering in the 1300-1400$ Range.

I will follow up with results and provide photos. Now, I am curious about the bandwidth of the bridge itself. ( The ASUS ROG is 40GBs ) some of the nvidia ones for the Quadro cards are 50GBs to 100GBs. I am wondering if finding a compatible with the higher bandwidth would even allow for more.

What do you think? Are there any more items or test you would do with these two cards to show what type of gains and performance increases you would do?
1 - 20 of 22 Posts

· Registered
Joined
·
2,958 Posts
The gains are gone when you don't play 3dmark, that's why you were told not to buy it. Only people that should be using nvlink /sli are those with 2080tis or rtx titans. You have to remember these are gaming cards, in the last 3 years maybe 3-5 new games support nvlink/sli and the scaling is rarely incredible. There are a few exceptions but those are rare.
 

· Tech Enthusiast
Joined
·
12,390 Posts
I'd avoid it at pretty much all cost but that's primarily because I don't care about bench mark scores much these days.
 

· Been there, done that
TwelveOnTap
Joined
·
276 Posts
Discussion Starter · #4 ·
Well if you would like to see performance gains for CAD software, I can write up a report. However, might not be to exciting, buts that's what my intent is on this machine. ;)
 

· Banned
Joined
·
5,111 Posts
Well if you would like to see performance gains for CAD software, I can write up a report. However, might not be to exciting, buts that's what my intent is on this machine. ;)
I'd like to see that. It would nice to be see the usefulness of NVLink outside of artificial benchmarks. It would be nice to see some actual games benchmarks as well.
 

· no regurts
Joined
·
645 Posts
The gains are gone when you don't play 3dmark, that's why you were told not to buy it. Only people that should be using nvlink /sli are those with 2080tis or rtx titans. You have to remember these are gaming cards, in the last 3 years maybe 3-5 new games support nvlink/sli and the scaling is rarely incredible. There are a few exceptions but those are rare.
I agree - and I've only been doing SLI starting with 1080's (then 1080 Ti's, Titan Xp's, failed to hack Titan V's NVLinks into working, and now Titan RTX). When it helps, it is usually pretty nice (the only two games I play that come to mind are Witcher 3 and most recently Jedi Fallen Order) but most games won't utilize the second card. This is especially true with DirectX12 titles since that API requires explicit game programming to use multiple graphics cards. In DX11, as long as the game can do full screen mode you *should* be able to get SLI going for it. The games that Nvidia has already made SLI profiles for are the best ones to use it with though - chances are if Nvidia hasn't managed to create an SLI profile for a game - there's probably a good reason for it.
 

· Vermin Supreme 2020
Joined
·
39,546 Posts
benchmarks show epic gains.

games rarely do, and many users end up fighting with it more than not, if they play a wide selection of titles. certain genres have support, like simming, for example, but all in all, in the grand history of mGPU, as of right now, for 90% of gamers, they should consider it dead.

NOW THEN .... if you're into benching. go flipping ham. most benchmarks have near perfect scaling. for gaming though, have fun with troubleshooting nearly inherent stutter(minimum FPS tanking) & latency injections, and have fun trying to figure out why the hell you aren't getting above 50% utilization, even with the bits and driver swaps, while everyone else (usually without proof) claim they're having a "perfect" experience (again, on the gaming side of things).

NVLINK tech itself is still not being utilized on a game engine rendering level though. it's all sli still.
 

· Banned
Joined
·
8,594 Posts
I agree - and I've only been doing SLI starting with 1080's (then 1080 Ti's, Titan Xp's, failed to hack Titan V's NVLinks into working, and now Titan RTX). When it helps, it is usually pretty nice (the only two games I play that come to mind are Witcher 3 and most recently Jedi Fallen Order) but most games won't utilize the second card. This is especially true with DirectX12 titles since that API requires explicit game programming to use multiple graphics cards. In DX11, as long as the game can do full screen mode you *should* be able to get SLI going for it. The games that Nvidia has already made SLI profiles for are the best ones to use it with though - chances are if Nvidia hasn't managed to create an SLI profile for a game - there's probably a good reason for it.
As more games utilize Vulkan and DX12 we'll see less SLI profiles. AMD already stopped supporting crossfire since the developers can make use of mgpu themselves.

https://developer.nvidia.com/explicit-multi-gpu-programming-directx-12

Background

Since the launch of SLI, a long time ago, utilization of multiple GPUs was handled automatically by the display driver. The application always saw one graphics device object no matter how many physical GPUs were behind it. With DirectX 12, this is not the case anymore. But why start doing something manually that has been working automatically? Because, actually, for a good while before DirectX 12 arrived, the utilization of multiple GPUs has not been that automatic anymore.

As rendering engines have grown more sophisticated, the distribution of rendering workload automatically to multiple GPUs has become problematic. Namely, temporal techniques that create data dependencies between consecutive frames make it challenging to execute alternate frame rendering (AFR), which still is the method of choice for distribution of work to multiple GPUs. In practice, the display driver needs hints from the application to understand which resources it must copy from one GPU to another and which it should not. Data transfer bandwidth between GPUs is very limited and copying too much stuff can make the transfers the bottleneck in the rendering process. Giving hints to the driver can be implemented with NVAPI or by making additional Clear() or Discard() calls for selected resources.

Consequently, even when you didn’t have explicit control over multiple GPUs, you had to understand what happened implicitly and give the driver hints for doing it efficiently in order to get the desired performance out of multi-GPU setups. Now with DirectX 12, you can take full and explicit control of what is happening. And you are no longer limited to AFR. You are free to invent new ways of making use of multiple GPUs that better suit your application
 

· no regurts
Joined
·
645 Posts
benchmarks show epic gains.

games rarely do, and many users end up fighting with it more than not, if they play a wide selection of titles. certain genres have support, like simming, for example, but all in all, in the grand history of mGPU, as of right now, for 90% of gamers, they should consider it dead.

NOW THEN .... if you're into benching. go flipping ham. most benchmarks have near perfect scaling. for gaming though, have fun with troubleshooting nearly inherent stutter(minimum FPS tanking) & latency injections, and have fun trying to figure out why the hell you aren't getting above 50% utilization, even with the bits and driver swaps, while everyone else (usually without proof) claim they're having a "perfect" experience (again, on the gaming side of things).

NVLINK tech itself is still not being utilized on a game engine rendering level though. it's all sli still.
Yeah idk I've never had any issues with it at all in any game - either it works or it doesn't. Haven't noticed any stutter/latency/etc but I also don't play hundreds of games, probably less than 50. The few where I've really wanted SLI do scale quite well (Witcher 3 and Jedi Fallen Order come to mind).

I've been running NVLink on 2080 Tis for over a year now (sig link), and so far, I have not had any problems...mind you, I don't play quite as many games as others seem to. What is exciting though is that NVidia (rather quietly) added CFR / checkerboard rendering support last November to the RTX line.

https://www.youtube.com/watch?v=HX5pNpVMShA
Yeah same boat for me - don't play a huge variety of games and it either works or just isn't supported at all. I still need to try enabling that checkerboard rendering though to see if it helps in something like World of Warcraft.
 

· Registered
Joined
·
4,359 Posts
i played shadow of the tomb raider/metro exodus/fc4 & 5/jedi fallen order/finally give a chance to star wars bf2 using sli and scaling was pretty good on 4k i aim for IQ and 60fps is my goal as i game on a 65' hdtv qdot samsung...
surely so 8x/8x vs 16x/16x is where the gains bus usage at


so framecap is a must everytime i use dual gpu setups when i can.... i surely use riva turner and framecap the crap out of them been doing it for years making sure i stress the IQ+resolution instead of letting the gpus to go rampage mode creating an unpleasant experience or using low resolutions where the cpu is a concern of a bottleneck even so on single gpus..

so my frametimes are a straight dead flat line @ 60 very responsive awesome experience....

but you know i didnt spent $2400 or 5k on just gpus lol the second 1080ti only cost me $420 second hand when they announced the RTX cards when people went ballistic for that 25% of the 2080ti more in average... which i bought for my emulator rig then changed minds... still the sucker "2080ti" its not @ [email protected] gpu in every game out there by herself with everything cranked up so more lol's
XD
 

· Been there, done that
TwelveOnTap
Joined
·
276 Posts
Discussion Starter · #12 ·
I no longer game anymore. Just CAD and Renders. Apparently, Autodesk has taken use of the NvLink in 3DMax. Next project comes up ill be able to confirm some results.
 

· Registered
Joined
·
1,330 Posts
You won't see any gpu rendering benefits of NVLink on your 2070s, only 2080 and up support NVLink functions for software level memory pooling, unfortunately. That said, you will still get both cards to render scenes, but will still go out of core if you max out the vram on one card with the 2070s, which will slow renders down. Basically, you can take the NVlink off and you won't notice a difference with your renders on 2070s.
 

· Disgruntled Dev
Joined
·
1,258 Posts
As I understand it, the nice thing about NVLink offers pro cards is the pooling of VRAM (whereas SLI/CFX never have, and never would). While Nvidia's new consumer cards are using the NVLink standard, the adapter isn't the same, so the bandwidth isn't there to actually pool VRAM of say two 2080Tis. I could be wrong, and I can't say I have any hands-on experience with it myself.

I've also read that someone was in fact able to have both cards and their VRAM pooled and utilized in Linux, but that could have been hearsay as well.

I'm at a point now where I've been asked to do some CAD/rendering of products, so I may throw my second 980Ti back into my rig just to get the extra benefit for that until I build a new rig. From a gaming perspective, a had an improvement in overall frames at 4K by removing the second card. This was years ago, and perhaps scaling/drivers have improved things in the few games I do play. But my 980Ti struggles pushing FPS and other open-world games at 4K as it is. I can push Escape from Tarkov at 4K anywhere between 35-70FPS, with high texture, and everything else turned down and still have a pleasant visual experience. But, the lower FPS can cause issues when things go from 0 to 60 in an instant as they often do in that game. Switching to my 2K Dell with a 75Hz refresh rate, I get a HUGE boost in FPS, which I didn't expect from the 980, but that goes to show how viable of a card it can still be.
 

· Banned
Joined
·
5,111 Posts
You won't see any gpu rendering benefits of NVLink on your 2070s, only 2080 and up support NVLink functions for software level memory pooling, unfortunately. That said, you will still get both cards to render scenes, but will still go out of core if you max out the vram on one card with the 2070s, which will slow renders down. Basically, you can take the NVlink off and you won't notice a difference with your renders on 2070s.
That sucks...
 

· Registered
Joined
·
1,330 Posts
As I understand it, the nice thing about NVLink offers pro cards is the pooling of VRAM (whereas SLI/CFX never have, and never would). While Nvidia's new consumer cards are using the NVLink standard, the adapter isn't the same, so the bandwidth isn't there to actually pool VRAM of say two 2080Tis. I could be wrong, and I can't say I have any hands-on experience with it myself.

I've also read that someone was in fact able to have both cards and their VRAM pooled and utilized in Linux, but that could have been hearsay as well.

I'm at a point now where I've been asked to do some CAD/rendering of products, so I may throw my second 980Ti back into my rig just to get the extra benefit for that until I build a new rig. From a gaming perspective, a had an improvement in overall frames at 4K by removing the second card. This was years ago, and perhaps scaling/drivers have improved things in the few games I do play. But my 980Ti struggles pushing FPS and other open-world games at 4K as it is. I can push Escape from Tarkov at 4K anywhere between 35-70FPS, with high texture, and everything else turned down and still have a pleasant visual experience. But, the lower FPS can cause issues when things go from 0 to 60 in an instant as they often do in that game. Switching to my 2K Dell with a 75Hz refresh rate, I get a HUGE boost in FPS, which I didn't expect from the 980, but that goes to show how viable of a card it can still be.
Memory pooling over NVLink works for geometry and VDB volumes on 2080 and 2080ti as long as the rendering software implements support for it. Currently, beta versions of some software are using it for both and I believe only VRay has it on their production versions. I've used it on Redshift and Octane so far, not sure if Arnold has gotten there yet. Puget Systems has a little app that can test your NVLink bandwidth as well. 2080Tis sit at 100GB/s and i believe 2080s are half that.

Link for that app is here if anyone is curious (the article is a bit old, so lots has happened in software since this was written): https://www.pugetsystems.com/labs/s...dro-and-GeForce-RTX-Cards-in-Windows-10-1266/
 

· Disgruntled Dev
Joined
·
1,258 Posts
Memory pooling over NVLink works for geometry and VDB volumes on 2080 and 2080ti as long as the rendering software implements support for it. Currently, beta versions of some software are using it for both and I believe only VRay has it on their production versions. I've used it on Redshift and Octane so far, not sure if Arnold has gotten there yet. Puget Systems has a little app that can test your NVLink bandwidth as well. 2080Tis sit at 100GB/s and i believe 2080s are half that.

Link for that app is here if anyone is curious (the article is a bit old, so lots has happened in software since this was written): https://www.pugetsystems.com/labs/s...dro-and-GeForce-RTX-Cards-in-Windows-10-1266/
Ah! That's what it was! As I also understand it, due to less pins on the consumer cards, that also cuts into bandwidth (50%) compared to the pro cards. I was under the impression that the reduced bandwidth was one of the limiting factors in the consumer cards being able to pool VRAM. It seems as though this isn't the case.
 

· Registered
Joined
·
1,330 Posts
Ah! That's what it was! As I also understand it, due to less pins on the consumer cards, that also cuts into bandwidth (50%) compared to the pro cards. I was under the impression that the reduced bandwidth was one of the limiting factors in the consumer cards being able to pool VRAM. It seems as though this isn't the case.
One NVLink max bandwidth is 100GB/s, so the 2080Ti does take full advantage of it.

Only way to get past that is to get the Quadro GV100's which uses TWO NVlinks for two cards, doubling the bandwidth.

https://blog.exxactcorp.com/nvlink-bridge-comparison-quadro-rtx-vs-gv100-whats-the-difference/
 

· Registered
Joined
·
2,586 Posts
I no longer game anymore. Just CAD and Renders. Apparently, Autodesk has taken use of the NvLink in 3DMax. Next project comes up ill be able to confirm some results.
Hi there

Depending on renderer you can mix GPUs, I'm using or running multiple GPUs or 4*GPUs setup with Asus RTX 2080Ti Strix, EVGA GTX 1080Ti,EVGA GTX1080 and Manli GTX1080 for rendering and I can only comment on my tests in Blender, my Asus RTX 2080Ti Strix will render Blender Classroom in 75 seconds with OptiX and 4*GPUs(RTX 2080Ti Strix, GTX1080, GTX1080 Ti and GTX1080) will finish that in 42-45 seconds without the OptiX and Asus RTX 2080Ti Strix without the OptiX will render that in 2 minutes and 19 seconds that's in Blender Cycles

In E-Cycles same scene it will render in 1minute and 44 seconds that's on Asus RTX 2080Ti Strix and without the OptiX and same scene with 4*GPUs(Asus RTX 2080Ti Strix, GTX1080, GTX1080 and GTX1080Ti) will render in 32-36 seconds and just two GTX1080 will render that scene in E-Cycles in 1 minute and 29 seconds

NVLink memory pooling doesn't work in every renderer, some renderers offers Out of Core memory which means it will use yours system memory rather than GPU

Depends on scene as well, in many cases 8GB is more than enough but in many cases I'm using different renderer like is Corona Renderer which is CPU based and its quite fast


Hope this helps

Thanks, Jura
 

· Disgruntled Dev
Joined
·
1,258 Posts
One NVLink max bandwidth is 100GB/s, so the 2080Ti does take full advantage of it.

Only way to get past that is to get the Quadro GV100's which uses TWO NVlinks for two cards, doubling the bandwidth.

https://blog.exxactcorp.com/nvlink-bridge-comparison-quadro-rtx-vs-gv100-whats-the-difference/
Yup, that's what it was. I read up on it a while back and then forgot most of the details.

I also assumed that ALL pro cards were using dual NVLink, but what I had read basically compared NVLink on a 20-series card to the Big Daddy GV100 and nothing in between.
 
1 - 20 of 22 Posts
Top