Enabling SLI on GTX 970

[padre.ayuso]

Hi there,

 

So here is a question. I have had two Zotac GTX 970 on my machines for a while but just in the last few days I have come to find that the SLI is not enabled on my system. My motherboard is from Dell, the Product Number is 0D881F, version A07, if that makes any sense to anyone, and I don't have an SLI bridge because I thought I did not need it. On the Nvidia Control Panel I don't have the option to turn SLI on or off, so in the internet I found that I have to manually turn it on BIOS, but I have not found it either.

 

Does anyone know how to enable SLI on my system? Would it make a difference in performance? I reckon it would, since two 970's should roar my system, but I don't know, maybe they are roaring my system and I don't know it, since it has been my first try at GPU renderings.

 

Should I buy an SLI bridge for my system and that would turn it on?

Edited November 13, 2015 by padre.ayuso

[Nikolaos M]

If it's only for gpu rendering, you should leave the gpus as they are. Make sure your motherboard "sees" both of them and check your gpu rendering software to ensure you can use both of them for rendering.

 

If you are a gamer, though, then sli could help in some games (I'm not a gamer, but I know a few things about games). You certainly need an sli bridge to enable sli inside the NV Control Panel.

[padre.ayuso]

Thanks Nikolaos!

 

So I gather that I don't need to SLI them to benefit to their max. They seem to both be working when I render. The only thing that annoys me is sometimes they take longer than I'd want, and this is for large architectural interior scenes, so I end up going CPU anyhow and at times my whole machine would slow down so much it is hard to move through the viewports at all, during GPU. If you have any advice, I'll happily take it. Yes, my motherboard does recognize them both and no, I don't play videogames on this system.

[Nikolaos M]

Sli would most probably slow down your rendering performance. So, yes, leaving them as they are, is the best choice for rendering.

 

It's not so weird that viewport performance is decreased while rendering. Gpu rendering pushes gpus to their limits, so if you're using your main gpu (the one that's connected to the monitor) for rendering too, it's natural to get big slowdowns.

 

Not all gpu renderers are quick enough as I hear. I've only tried i-ray and I wasn't impressed. Octane, Redshift and FurryBall are said to be much faster. Which one are you currently using?

 

Edit: One more thing. If your scenes are very big and RAM heavy, it's possible that the memory needs for the renderings exceed your Vram amount. Gtx 970 has 4gb of vram, so a scene that's memory hungry might not "fit" inside your card's available Vram. If that occurs, the rendering passes through the cpu by default. Check your Vram usage and confirm it's under 4gb for the rendering.

Edited November 13, 2015 by nikolaosm

[padre.ayuso]

All that makes sense Nikolaos. I'm using VRay and usually do CPU renderings but wanted to use GPU for testing as this would, in theory, speed up the light and texturing tweaking.

 

Good to know that SLI is not needed, so that takes a worry off my back. Most of my scenes are not that memory hungry, but I believe that maps may take the cake here, specially when using a huge HDRI. I notice that when I use the HDRIs that are small in size (20MB) I'm able to use GPU relatively OK, but when I use the 150MB or above version of the same HDRI I may as well use CPU as GPU with my current cards don't quite make the grade and I wait for a long time just to have an image that I can work with.

 

So thank you for this data as it has helped me clear up this confusion I had. Now, one other question, if I was building a new rig and was to use these two GTX970 for test renderings, I could get another card for viewport handling while rendering, right? Like, getting a Quadro or Tesla? and don't use these for rendering, just for viewport handling? What do you recommend?

[Nikolaos M]

So thank you for this data as it has helped me clear up this confusion I had. Now, one other question, if I was building a new rig and was to use these two GTX970 for test renderings, I could get another card for viewport handling while rendering, right? Like, getting a Quadro or Tesla? and don't use these for rendering, just for viewport handling? What do you recommend?

 

Quadro, yes. Tesla, no. It's only for computing tasks. It has no outputs.

 

I haven't tried mixing cards. I use cpu rendering basically. I think it can be done with no issues. But you don't seem to have the right motherboard and a beefy psu to handle this work load. Dell Precision T7500 must be your WS. So, if I'm right, your motherboard has only 2 pciex16 (gen 2) slots. Is that right?

 

I think you must start thinking about an upgrade. Two 970s in a system old as yours is somehow weird.

[padre.ayuso]

Quadro, yes. Tesla, no. It's only for computing tasks. It has no outputs.

I think you must start thinking about an upgrade. Two 970s in a system old as yours is somehow weird.

 

Thanks for the data on the Graphics Cards.

 

LOL!!!! Yes, an upgrade is coming up very shortly, partially the reason I'm doing some research right now. Funny is that when we got these systems 7 years ago, they were "cheaper" because they were already a year or two old models. So, last year was my last upgrade of these machines with some extra RAM and SSD but we are now looking into an overhaul. 5 machines for workstations, yes, and I want something that I don't have to look back at in about 5 to 8 years, if anything only to upgrade RAM and GPU at the most. I saw the config you recommended someone else for under 2 grand, but perhaps you would recommend something a bit beefier? Wouldn't I rather go for double 8 core CPUs? with hyperthread turned on, wouldn't that give me 32 threads? And then, Xeon5 chips are more expensive than Core 7, but for still images and renderings, is it worth the extra amount? I know for slave machines, a Xeon 5 chip is better, something to do with something but unsure what. Then again, when doing my test renderings, using my current workstation, wouldn't a Xeon 5 chip be more efficient than a Core i7?

[padre.ayuso]

And one datum, not so much a question, but since I was looking in the net, a Quadro with 4GB, the K2200 costs me double as the 970 (also, the K2200 has half the amount of CUDAs being used). I asked Tesla or Quadro, because I understood that Tesla was for viewport navigation, but otherwise, I had this datum from Dimitris from a while back that I put two 970s together and it would give me the 4 times the performance of a quadro for the same price. So I did, but having been my first GPU venture, I can't compare to anything in the past. But based on the data you have given me thus far, it seems that the K2200 would have been even slower than my 970s. Based on all of this, if I get another card, it may be a TitanX as a third card for each machine, which may be able to then take the bigger jobs, and I should be able to use the power of all 3 cards? If anyone else has experimented with this, please let me know.

[Nikolaos M]

Quadros are better in very specific areas compared to the equivalent gtx's. Mostly in OpenGL based software like Catia, Solidworks etc., some very specialized programs (medical etc.) and in ray traced rendering and iray.

 

In all other gpu renderers like Octane etc. or direct3d accelerated viewports, a quadro should be on par with the corresponding gtx. So, inside software like 3ds max or Maya you should see no difference between a k2200 and the equivalent consumer card. The only difference is the price...

 

Now, about the upgrade potential. You must first decide the kind of rendering you will focus on. Building a rig for gpu rendering has a very different philosophy compared to a cpu oriented rendering machine.

 

If you decide to place your bet in gpu rendering, then you need a good motherboard with multiple pcie slots and the best power supply your money can buy. And of course, a big and spacious case with great airflow in order to manage all the heat it's goin to be produced by 2 or 3 gpus running at full speed while rendering. In this case, you don't need dual Xeons etc. A Asus-X99 WS motherboard and a 5930K (with 40 pcie lanes) is a good combination I think, plus a Corsair AX1500i to power the "beasts" inside. Theoretically, it can feed up to 4 Titan X's under full load (about 250W each, without oc of course).

 

A cpu rendering machine, on the other side, has a different logic behind it. And, splitting your build in two parts, one as a workstation and one (or more) as a render slave(-s) is the best way to work and render at the same time. You can find many builds and cpu rendering configurations inside this forum, if you search a little. There is a lot of feedback to take into account.

Edited November 14, 2015 by nikolaosm

[padre.ayuso]

Thanks Nikolaos,

 

Yes, all of that data helps a lot. Cheers to you!

[RyderSK]

Modern GPU raytracing engines (such as Octane or Redshift) do have their own equivalent of "low-thread" like in CPU raytracers, where they leave small percentage of performance to cater to viewport navigation, so your PC won't completely "freeze" for other visual demanding tasks.

 

Of course, for smooth multitasking, keeping primary GPU for performance is still benefitial (if you want to keep modeling in separate session while still rendering at full power).

 

Raytracing does not benefit from pro-cards like Tesla and Quadro range because raytracing does not require dual-precision calculations (brute force algorithms are just averaging of endless noise patterns) like scientific (medical, space, economic,etc..) calculations do. Same goes for other hardware benefits like buffered ECC ram to enable stacking of hundreds (and thousands) of units next to each other in farms/supercomputers and correcting small errors in quite rare cycles that don't affect raytracing in any way either.

 

Quadro as primary viewport card has benefit if you want full 14bit color LUT pipeline (for cinematic grading, or precise scientific graphics,etc..) or OpenGL based CAD softwares (which are getting rarer) otherwise consumer grade GTX will do equally solid job today in 3dsMax, for fraction of the price.

[Nikolaos M]

I may write in English here in the forums, but I still think "Greek".

 

Quadros are better in very specific areas compared to the equivalent gtx's. Mostly in OpenGL based software like Catia, Solidworks etc., some very specialized programs (medical etc.) and in ray traced rendering within iray.

 

In all other gpu renderers like Octane etc. or direct3d accelerated viewports, a quadro should be on par with the corresponding gtx.

 

Now it makes sense for other people too and not only Greeks (the syntax with "and" is a Greek idiom). Iray seems to favor Quadros vs GeForce cards. All other renderers don't, exactly as you said Juraj.

 

And one more thing. New Quadros (Maxwell) have a very poor FP64 performance (only 1:32 FP32), while the Keplers were much-much better in double precision (~1:3 FP32).

Edited November 16, 2015 by nikolaosm

[RyderSK]

Iray seems to favor Quadros vs GeForce cards. All other renderers don't, exactly as you said Juraj.

 

iRay is no different, it uses resources same way as other GPU raytracers. Raytracing is simply not double precision calculation, so it cannot benefit from nVidia's arbitrary feature stripping from consumer cards, even if the kernels run on their proprietary CUDA.

 

On architecture level, the performance acceleration is pretty much identical across identical cards from both ranges. I.e GTX TitanX = QuadroM6000 with slightly higher clock. And as you see in benchmark, this rings true, Titan-X is indeed few smallish percents faster than M6000.

Same memory (12gb) as well, for price difference of striking 4 thousands dollars :- ).

 

http://www.migenius.com/products/nvidia-iray/iray-benchmarks-2015

 

[Nikolaos M]

I'm aware that gpu rendering is a single-precision FP procedure, Juraj. That's why I recommend GeForce cards for this task. But I was aware of this older benchmark

 

 

And recently, in another thread, a member sent me a file to try an iray rendering with my gtx 780 (2304 cuda cores, Kepler arch.). He had a K4000 (768 cuda cores, Kepler also, just the 1/3 of the 780's), and with exactly the same iray settings I managed to shave onle only a few minutes compared to his rendering time (26'35" was mine and ~34' was his). It was quite impressive, if you take into account the amount of cuda cores in each case and the higher clocks of the 780. The k4000's time was not bad at all.

 

As for the bench results you uploaded, I understand that it's more fresh and probably refers to a newer iray version, but some results seam weird. Ex. The 980 is so much worse in iray compared to a 780Ti? I know that the 780Ti has 35% more cuda cores, but where is the architectural improvement per shader unit? I was waiting a similar performance between these two cards. I don't know what you think about that.

Edited November 16, 2015 by nikolaosm

[RyderSK]

Hey, don't take offence about the single/double remark, it wasn't even aimed at you, just general remark :- )

 

Could you please link me to the Tom's HW review directly ? I've seen tons of benchmarks by now and this one just seem strange, almost like it was testing in forced double-precision mode.

 

Regarding the superclocked 780Ti beating 980, the high margin in this case is slightly suspicious, since otherwise these cards are pretty equal in general performance. The methodology is otherwise pretty sound and robust enough, so I would take it bit more seriously than Tom's HW. It also does test the latest iRay available, although, outside of 3dsMax.

Edited November 16, 2015 by RyderSK

[Nikolaos M]

Could you please link me to the Tom's HW review directly ? I've seen tons of benchmarks by now and this one just seem strange, almost like it was testing in forced double-precision mode.

 

Sure. Here it is

 

http://www.tomshardware.de/geforce-quadro-workstation-grafikkarte-gpu,testberichte-241759-3.html

[RyderSK]

Searching bit more on that, I found discussion of that benchmark here :

 

http://forums.autodesk.com/t5/3ds-max-3ds-max-design-general/iray-titan-x-vs-quadro-m6000/td-p/5592903

 

General consensus just seems that's an really odd benchmark. Seems nVidia provided an iRay patch since for all 3dsMax versions (and thus older iRay cores) that fixes the performance for Maxwell cards, but that does not explain why Pro Maxwell cards didn't suffer the loss. Almost looks like that typical nVidia's shit...

 

Won't waste further time on this, the correct solution to this is to ignore iRay :- )

[padre.ayuso]

Searching bit more on that, I found discussion of that benchmark here :

 

http://forums.autodesk.com/t5/3ds-max-3ds-max-design-general/iray-titan-x-vs-quadro-m6000/td-p/5592903

Won't waste further time on this, the correct solution to this is to ignore iRay :- )

 

Give a high-five on that one.

[Nikolaos M]

Won't waste further time on this, the correct solution to this is to ignore iRay :- )

 

Haha..! I vote for this too. Especially after the iray test I ran for this guy here in the forums, I don't see any tempting performance gains vs cpu rendering.

 

Some guys I know in other CG forums, though, have tried out both Redshift and FurryBall and were impressed by their speed and overall behavior. I'm really interested in taking a chance with FurryBall for some car/motorcycle renderings I have in mind. Do you have any experience or feedback of any type on this renderer, Juraj?

Edited November 16, 2015 by nikolaosm

[RyderSK]

Both Redshift and Octane look like superb engines to me. Furryball is sort of odd, and they're known for less than stellar communication and benchmarking techniques ;- )

 

Redshift is outside of my league since it's basically gpu accelerated Vray (Vray in older sense, like 1.5 or 2), one that values speed over simplicity , which aligns with their animation oriented focus. Octane is more like my cup of tea, very similar to Corona that I use in workflow. I don't particularly think that even with my Titan-X I could comfortably fit my scenes within (I routinely go over 30 gb of memory usage). So for me, it's not the best solution. Not yet at least.

 

But others have great results with it and seem very content (particularly with Octane). Best to always try and find personal favourite. All the engines on the market are quite great. Of course, VrayRT is doing some serious improvements as well, shouldn't discount that one.

[erkutacar]

i got quadro 5000 once and it was the stupidest purchase ive ever made.. it has the same performance as gtx580.. if u want to enable sli for iray rendering, dont! iray doesnt support sli and can use each card individually no matter what model it is as long as its cüda! if u wanna enable it for games, sli might be turn off in the bios..