RTX 2080 super in nvlink/sli vs RTX 2080 ti

xionis_2946116

I'm having a little buyers remorse and I'm wondering if I made a mistake. Lastnight I purchased 2 rtx 2080 super cards as well as a nvlink and I'm not noticing too much of a improvement over when I was running a single rtx 2080 super. I still have 30 days to return them to the store though... So should I take them back and buy a 2080 ti instead? Or are there additional settings that I need to check so that Daz will perform better using these 2 cards?

duckbomb

I'm not an expert, but I usually just go for the highest VRAM on-card.  The speed differences are negligible to me across all current-gen cards, but if your scene doesn't fit then nothing else matters.  If you can fit all of your scene in 8Gb of VRAM, the SUPER is fine.  Two of them is about 50% more fine.

If you need more than 8 and less than 11 GB to fit your scene on the card, than the TI is a better option than two SUPERs.

Its my initial thought, though, and your mileage may vary.  If your scenes all fit under 8 gigs, I'd say your fine.

Taoz

duckbomb said:

I'm not an expert, but I usually just go for the highest VRAM on-card.  The speed differences are negligible to me across all current-gen cards, but if your scene doesn't fit then nothing else matters.  If you can fit all of your scene in 8Gb of VRAM, the SUPER is fine.  Two of them is about 50% more fine.

If you need more than 8 and less than 11 GB to fit your scene on the card, than the TI is a better option than two SUPERs.

Its my initial thought, though, and your mileage may vary.  If your scenes all fit under 8 gigs, I'd say your fine.

+1

kenshaw011267

Nvlink/SLI does not improve performance in iRay over just having the two cards.

The 2080 super is not far off a 2080ti i performance. 2 should really be almost twice as fast as one. You need to check that both are selected in your render settings.

outrider42

If you are not seeing much of an improvement, something is wrong. There may be other factors at play, from your motherboard to power restrictions. You can try a hardware monitor to see that your GPUs are running. I advise one regardless so that you can control your temps better with fan curves.

Here, try this benchmark.

https://www.daz3d.com/forums/discussion/341041/daz-studio-iray-rendering-hardware-benchmarking/p1

I would like you to try the bench with a single 2080 Super, and then try it again with both.

xionis_2946116

Yes I have both cards selected for rendering along with my threadripper 3970x. I noticed that in task manager in my performance it's showing that 1 gpu is using about 80-100% while the other will show 0-2% usage but it says they're linked. I've tried rendering with sli enabled as well as having it disabled through the Nvidia control panel but in Daz I always made sure both cards are selected. I've even tried rendering without the cpu which seems to be slower. I'll try the benchmark tonight and report back with the results.

 

duckbomb said:

I'm not an expert, but I usually just go for the highest VRAM on-card.  The speed differences are negligible to me across all current-gen cards, but if your scene doesn't fit then nothing else matters.  If you can fit all of your scene in 8Gb of VRAM, the SUPER is fine.  Two of them is about 50% more fine.

If you need more than 8 and less than 11 GB to fit your scene on the card, than the TI is a better option than two SUPERs.

Its my initial thought, though, and your mileage may vary.  If your scenes all fit under 8 gigs, I'd say your fine.

How do I know how much memory my scene is using? 

xionis_2946116

Ok from looking at these benchmarks it appears that i don't need to have matched cards, because i'm seeing all sorts of dual card configs, so could that nvlink bridge be causing me performance issues and preventing daz from using both cards effeintly? Like it's trying to use them as 1 card instead of 2 seperate cards

tj_1ca9500b

If I understand correctly, NVLink allows you to pool the memory into a larger memory pool, but there is a performance hit when doing so, as the GPUs have to talk across the NVLink to pool said memory.  Also the software has to include support for said memory pooling.

I'm guessing here, but Daz may not have fully implemented said memory pooling, or the NVLink may otherwise be tripping Daz up, so that might be why the second card is essentially sitting idle.

Have you tried rendering with both cards with the NVLink removed? 

xionis_2946116

tj_1ca9500b said:

If I understand correctly, NVLink allows you to pool the memory into a larger memory pool, but there is a performance hit when doing so, as the GPUs have to talk across the NVLink to pool said memory.  Also the software has to include support for said memory pooling.

I'm guessing here, but Daz may not have fully implemented said memory pooling, or the NVLink may otherwise be tripping Daz up, so that might be why the second card is essentially sitting idle.

Have you tried rendering with both cards with the NVLink removed? 

From what i've read is that memory pooling is a feature in the new beta but i didn't notice much of a difference when I tried the beta, but it may be only available for Titans and Ti's because their NVlink is twice as fast from what i understand. I'm going to remove the nvlink tonight to see how well it performs. I'm not even sure if I need the memory pooling because i'm unsure how large my scenes really are, and I think I would most likely benefit from 2 cards rendering instead.

Richard Haseltine

I'm not sure about the 2080Ti, but it certainly isn't available on the lower cards. It is possible on Titans and Quadros, but only for textures - everything else must load completely onto each card

RayDAnt

xionis said:

I'm having a little buyers remorse and I'm wondering if I made a mistake. Lastnight I purchased 2 rtx 2080 super cards as well as a nvlink and I'm not noticing too much of a improvement over when I was running a single rtx 2080 super. I still have 30 days to return them to the store though... So should I take them back and buy a 2080 ti instead? Or are there additional settings that I need to check so that Daz will perform better using these 2 cards?

First and foremost, NVLink functionality between Nvidia GPUs in Iray will not enhance rendering performance/speed (it will actually hurt it slightly.) What it will do is increase the size (in terms of memory consumption) of the scenes you will be able to render at a CPU-excelerated rate on your GPUs. In terms of raw performance, testing so far indicates that two 2080s should get you at least 30% more rendering performance (speed) over a single 2080Ti. And the max theoretical memory capacity of two NVLinked 2080s is 16GB (likely more like 14GB) versus a single 2080Ti's 11GB. So assumnig you got a good deal on that dual 2080/NVLInk bridge combo - and NVLink can actually be made to work (more below) - I'd say it's more than worth it in comparison.

 

mikeheonis said:

Yes I have both cards selected for rendering along with my threadripper 3970x. I noticed that in task manager in my performance it's showing that 1 gpu is using about 80-100% while the other will show 0-2% usage but it says they're linked.

Don't trust what Windows Task Manager says about GPU activity in this instance. It is designed first and foremost to report back game engine and Windows display device driver processing activity. Not GPU compute activity - which is how Iray goes about doing its business. As far as your computer is concerned Iray isn't even officially a graphics program. It's an application that can run natively on your GPU which only coincidentally produces rendered graphics. 

The current best way to track what's actually going on with your GPUs with applications like Iray is to use Nvidia's commandline System Management Interface tool that is included with every Nvidia driver installation package (third-party apps like GPU-Z actually get their performance data via an API from this same tool.) Open a Command Prompt window and type "nvidia-smi". You should see a readout like the following:

C:\>nvidia-smiWed Jan 29 13:07:34 2020+-----------------------------------------------------------------------------+| NVIDIA-SMI 441.66       Driver Version: 441.66       CUDA Version: 10.2     ||-------------------------------+----------------------+----------------------+| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC || Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. ||===============================+======================+======================||   0  TITAN RTX          WDDM  | 00000000:01:00.0  On |                  N/A ||  0%   28C    P8     8W / 280W |    874MiB / 24576MiB |      1%      Default |+-------------------------------+----------------------+----------------------+

Except with two sets of numbered statistics (one for each of your 2080s) where you see the single set for my Titan RTX. Each card's GPU-Util (second from the right on the bottom) are the main stats you want to look for here.

 

mikeheonis said:

I've tried rendering with sli enabled as well as having it disabled through the Nvidia control panel but in Daz I always made sure both cards are selected. I've even tried rendering without the cpu which seems to be slower. I'll try the benchmark tonight and report back with the results.

SLI has never been required for multi-GPU use in Iray (because of that whole compute program vs official graphics program dinstinction.) In worst case situations (mismatched graphics cards) having SLI enabled while rendering in Iray can lead to signficant loss of performance siince SLI traditionally works by artificially downclocking the fastest card(s) to match the performance properties of the slowest. However in your case (identical cards linked together) this should be a non-issue. SLI on/off shouldn't matter in terms of performance/speed.

Where SLI on/off might matter for you is in getting NVLink to function for the additional usable memory capacity. Based on some preliminary testing done by Puget Systems in the months after RTX first came out (meaning that it is very possible to no longer be true) getting NVLink to work at all on either the 2080 or 2080Ti requires having SLI enabled in Graphics Settings as a workaround for getting P2P communication (the underlying technology behind both memory pooling and SLI on Turing cards) over the NVLink bridge to turn itself on. One simple way you might be able to verify whether this is still true is by running the command "nvidia-smi nvlink -s" (this will tell you whether your NVLink connection is "active") with your NVLink bridge installed - first with SLI enabled and then with it disabled - and see if the readout changes. I say might because Nvidia doesn't specify whether this command's "active" includes "active but idle". Meaning that getting "inactive" in both cases might not mean anything unless you are actively rendering a scene while running the command.

 

xionis said:

could that nvlink bridge be causing me performance issues and preventing daz from using both cards effeintly? Like it's trying to use them as 1 card instead of 2 seperate cards

It is very possible that this is indeed how it works. Quite simply, the hardware capabilities and (apparent) software support for memory pooling over NVLink in Iray on consumer GPUs is so new that no one really knows just yet how it works. Is texture data initially loaded redundantly to all member GPUs in an NVLink pool (as is done in all other Iray rendering situations) and then selectively cleared from/replaced by unique data on specific cards as the amount of scene data needing to be stored surpasses the memory capacity of the smallest card? Or is texture data uniquely loaded onto member GPUs in the NVLink pool from the start regardless of whether it could all fit redudantly on each and every card.

I highly suspect it is the latter case since the former would end up being extremely unwieldy (in terms of scene loading time) from a system performance standpoint. Meaning that it is very likely people will only be wanting to enable NVLink pooling when it is absolutely necessary for overcoming individual GPU memory constraints since it will always result in worse overall performance/speed than with it disabled.

tj_1ca9500b

Richard Haseltine said:

I'm not sure about the 2080Ti, but it certainly isn't available on the lower cards. It is possible on Titans and Quadros, but only for textures - everything else must load completely onto each card

At least not in Daz Studio apparently.  Some apps will support NVLink on RTX 2080 and 2080 Ti cards, but it depends on whether the application is coded for it.

https://www.pugetsystems.com/labs/articles/NVLink-on-NVIDIA-GeForce-RTX-2080-2080-Ti-in-Windows-10-1253/

Notably, according to the above article, Linux does support peer to peer NVLink on the 'lower end' cards...

Also, according to this article, the 2080 Ti supports 100 GB/Sec over NVLink, while the 2080 only supports 50 GB/Sec.

https://hothardware.com/reviews/nvidia-nvlink-review-taking-sli-to-the-next-level

Either way, the software needs to support it, and from what you've indicated, it would appear that so far the Titans and Quadro NVLinks are supported in the Daz Studio beta.

 

 

xionis_2946116

RayDAnt said:

xionis said:

I'm having a little buyers remorse and I'm wondering if I made a mistake. Lastnight I purchased 2 rtx 2080 super cards as well as a nvlink and I'm not noticing too much of a improvement over when I was running a single rtx 2080 super. I still have 30 days to return them to the store though... So should I take them back and buy a 2080 ti instead? Or are there additional settings that I need to check so that Daz will perform better using these 2 cards?

First and foremost, NVLink functionality between Nvidia GPUs in Iray will not enhance rendering performance/speed (it will actually hurt it slightly.) What it will do is increase the size (in terms of memory consumption) of the scenes you will be able to render at a CPU-excelerated rate on your GPUs. In terms of raw performance, testing so far indicates that two 2080s should get you at least 30% more rendering performance (speed) over a single 2080Ti. And the max theoretical memory capacity of two NVLinked 2080s is 16GB (likely more like 14GB) versus a single 2080Ti's 11GB. So assumnig you got a good deal on that dual 2080/NVLInk bridge combo - and NVLink can actually be made to work (more below) - I'd say it's more than worth it in comparison.

 

mikeheonis said:

Yes I have both cards selected for rendering along with my threadripper 3970x. I noticed that in task manager in my performance it's showing that 1 gpu is using about 80-100% while the other will show 0-2% usage but it says they're linked.

Don't trust what Windows Task Manager says about GPU activity in this instance. It is designed first and foremost to report back game engine and Windows display device driver processing activity. Not GPU compute activity - which is how Iray goes about doing its business. As far as your computer is concerned Iray isn't even officially a graphics program. It's an application that can run natively on your GPU which only coincidentally produces rendered graphics. 

The current best way to track what's actually going on with your GPUs with applications like Iray is to use Nvidia's commandline System Management Interface tool that is included with every Nvidia driver installation package (third-party apps like GPU-Z actually get their performance data via an API from this same tool.) Open a Command Prompt window and type "nvidia-smi". You should see a readout like the following:

C:\>nvidia-smiWed Jan 29 13:07:34 2020+-----------------------------------------------------------------------------+| NVIDIA-SMI 441.66       Driver Version: 441.66       CUDA Version: 10.2     ||-------------------------------+----------------------+----------------------+| GPU  Name            TCC/WDDM | Bus-Id        Disp.A | Volatile Uncorr. ECC || Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. ||===============================+======================+======================||   0  TITAN RTX          WDDM  | 00000000:01:00.0  On |                  N/A ||  0%   28C    P8     8W / 280W |    874MiB / 24576MiB |      1%      Default |+-------------------------------+----------------------+----------------------+

Except with two sets of numbered statistics (one for each of your 2080s) where you see the single set for my Titan RTX. Each card's GPU-Util (second from the right on the bottom) are the main stats you want to look for here.

 

mikeheonis said:

I've tried rendering with sli enabled as well as having it disabled through the Nvidia control panel but in Daz I always made sure both cards are selected. I've even tried rendering without the cpu which seems to be slower. I'll try the benchmark tonight and report back with the results.

SLI has never been required for multi-GPU use in Iray (because of that whole compute program vs official graphics program dinstinction.) In worst case situations (mismatched graphics cards) having SLI enabled while rendering in Iray can lead to signficant loss of performance siince SLI traditionally works by artificially downclocking the fastest card(s) to match the performance properties of the slowest. However in your case (identical cards linked together) this should be a non-issue. SLI on/off shouldn't matter in terms of performance/speed.

Where SLI on/off might matter for you is in getting NVLink to function for the additional usable memory capacity. Based on some preliminary testing done by Puget Systems in the months after RTX first came out (meaning that it is very possible to no longer be true) getting NVLink to work at all on either the 2080 or 2080Ti requires having SLI enabled in Graphics Settings as a workaround for getting P2P communication (the underlying technology behind both memory pooling and SLI on Turing cards) over the NVLink bridge to turn itself on. One simple way you might be able to verify whether this is still true is by running the command "nvidia-smi nvlink -s" (this will tell you whether your NVLink connection is "active") with your NVLink bridge installed - first with SLI enabled and then with it disabled - and see if the readout changes. I say might because Nvidia doesn't specify whether this command's "active" includes "active but idle". Meaning that getting "inactive" in both cases might not mean anything unless you are actively rendering a scene while running the command.

 

xionis said:

could that nvlink bridge be causing me performance issues and preventing daz from using both cards effeintly? Like it's trying to use them as 1 card instead of 2 seperate cards

It is very possible that this is indeed how it works. Quite simply, the hardware capabilities and (apparent) software support for memory pooling over NVLink in Iray on consumer GPUs is so new that no one really knows just yet how it works. Is texture data initially loaded redundantly to all member GPUs in an NVLink pool (as is done in all other Iray rendering situations) and then selectively cleared from/replaced by unique data on specific cards as the amount of scene data needing to be stored surpasses the memory capacity of the smallest card? Or is texture data uniquely loaded onto member GPUs in the NVLink pool from the start regardless of whether it could all fit redudantly on each and every card.

I highly suspect it is the latter case since the former would end up being extremely unwieldy (in terms of scene loading time) from a system performance standpoint. Meaning that it is very likely people will only be wanting to enable NVLink pooling when it is absolutely necessary for overcoming individual GPU memory constraints since it will always result in worse overall performance/speed than with it disabled.

Thanks so much for taking the time explaining all of this. I can't wait to get home and test the bench mark file with and without the nvlink installed or do I just disable nvlink/sli in the nvidia settings? I really wish Daz had some sort of scene memory usage meter so that I can monitor how large my scene is so I can better determine what size card I need. 

RayDAnt

@xionis here is how to (fingers crossed) definitively verify whether memory pooling over NVLink is actually working or not in Iray under any given testing situation:

1. Open Nvidia Control Panel, select "Desktop" from the top menu bar and make sure "Enable Developer Settings" is selected.
2. Still in Nvidia Control Panel, expand "Developer" in the task column on the left, click "Manage GPU Performance Counters"  and select "Allow access..." on the options page before clicking "Apply".
3. With your NVLink bridge installed, configure other drivers settings/windows settings/Daz Studio/Iray settings/etc to a state where you would like to test for memory pooling.
4. Open a Command Prompt window with administrator permissions (right click > "Run as admin...")
5. Type "nvidia-smi nvlink -sc 0bz". Assuming your NVLink bridge is active, this will set up a driver-level utilization counter for all data activity (measure in bytes) happening in real time over the NVLink connection between your GPUs.
6. Open Daz Studio and get a scene ready to render.
7. Type "nvidia-smi nvlink -r 0" in Command Prompt. This will reset to zero the previously confgured NVLink utilization counter.
8. Initiate a render in Daz Studio.
9. While the render is still going on, type "<strong>nvidia-smi nvlink -g 0</strong>" in Command Prompt multiple times over the course of the rendering process. This will tell you how much (if any) data is being passed from GPU to GPU (in bytes) over the NVLink connector as your scene renders. If these numbers are large and increase rapidly, that is an almost certain sign that memory pooling is taking place. If not, then that means that it isn't.

To find out more about the various Nvidia SMI tool commands listed above, type "nvidia-smi nvlink -h" in command Prompt. For more info on SMI tool functionality in general, type "nvidiai-smi -h".

xionis_2946116

One other question....Should I be using the nvidia studio drivers or the game ready drivers? I'm currently using the latest studio drivers which was released back in December but i noticed there is a newer game ready driver I could be using.

RayDAnt

xionis said:

One other question....Should I be using the nvidia studio drivers or the game ready drivers? I'm currently using the latest studio drivers which was released back in December but i noticed there is a newer game ready driver I could be using.

The only functional difference between the two for anything Iray related is the bug testing on Nvidia's end that went into them prior to release. Studio drivers are supposed to be more stringently back-tested/stable than Game Ready releases (hence why the release numbers of Studio drivers almost always lag behind Game Ready releases in terms of release dates.) But there have been times where even Studio releases had obviously Iray-affecting bugs to them. Personally I prefer Game Ready as I like being on the front line for feature improvements (and, as already mentioned, the "Studio" moniker has already proved to be fallible.) However, if/when in an active production process with deadlines I'd prefer the later.

namffuak

xionis said:

Ok from looking at these benchmarks it appears that i don't need to have matched cards, because i'm seeing all sorts of dual card configs, so could that nvlink bridge be causing me performance issues and preventing daz from using both cards effeintly? Like it's trying to use them as 1 card instead of 2 seperate cards

IIRC, NVLink is nly supported in the current public beta, so you won't see any benefit if you are not running it. But you should see both cards being used - check with a monitoring program like GPU-Z.

xionis_2946116

namffuak said:

xionis said:

Ok from looking at these benchmarks it appears that i don't need to have matched cards, because i'm seeing all sorts of dual card configs, so could that nvlink bridge be causing me performance issues and preventing daz from using both cards effeintly? Like it's trying to use them as 1 card instead of 2 seperate cards

IIRC, NVLink is nly supported in the current public beta, so you won't see any benefit if you are not running it. But you should see both cards being used - check with a monitoring program like GPU-Z.

Yeah i'm going to try monitoring with GPU-Z tonight instead of using task manager. I did try out the beta briefly lastnight but when trying to render a scene it kept crashing to desktop without any error messages, not even an error in the log file that I had noticed. So i'm going to stick with the public release for now. i'm exited to see if I can get this to perform better than it has been.

Chohole

Rob says

NVLink Peer Group Size defaults to "0" (effectively disabled). Without a way to set it in the General Release, you have no way to enable it... unless you use the latest Public Build, which exposes the option.
 

xionis_2946116

Well I have some interesting results with the benchmarks, which I will post the full results in the page linked above later. But I figured I'd share my results with the sli enabled/disabled and the nvlink removed as well. One thing is for sure its definitely using both video cards regardless of what sli or nvlink is set at, it all depends on if it's checked in Daz.

Here's my results of the benchmark file. I plan to update the driver to see if that changes anything but this is using the December studio driver in Daz studio 4.120.86

Sli enabled + nvlink adapter on - optix OFF: 2min 38.20secs
Sli enabled + nvlink adapter on - optix ON: 2min 38.22secs

Sli disabled + nvlink adapter on - optix OFF: 2min 38.4secs
Sli disabled + nvlink adapter on - optix ON: 2min 36. 2secs

Nvlink adapter removed - optix ON: 2min 37.48secs
Nvlink adapter removed - optix OFF: 2min 35.43

Single 2080 super - optix off: 4min 58.90secs
Single 2080 super - optix on: 4min 58.82secs

RayDAnt

xionis said:

Well I have some interesting results with the benchmarks, which I will post the full results in the page linked above later. But I figured I'd share my results with the sli enabled/disabled and the nvlink removed as well. One thing is for sure its definitely using both video cards regardless of what sli or nvlink is set at, it all depends on if it's checked in Daz.

Here's my results of the benchmark file. I plan to update the driver to see if that changes anything but this is using the December studio driver in Daz studio 4.120.86

Sli enabled + nvlink adapter on - optix OFF: 2min 38.20secs
Sli enabled + nvlink adapter on - optix ON: 2min 38.22secs

Sli disabled + nvlink adapter on - optix OFF: 2min 38.4secs
Sli disabled + nvlink adapter on - optix ON: 2min 36. 2secs

Nvlink adapter removed - optix ON: 2min 37.48secs
Nvlink adapter removed - optix OFF: 2min 35.43

Single 2080 super - optix off: 4min 58.90secs
Single 2080 super - optix on: 4min 58.82secs

For what it's worth, the reason why the OptiX Prime on/off times you got in each of your four test cases are so similar/virtually identical is because they ARE actually identical. When Nvidia re-tooled Iray to utilize RTX they switched it to using the pure OptiX API pipeline for all parts of the rendering process rather than just (optionally) raytracing acceleration. OptiX Prime acceleration serves no purpose on RTX cards when full OptiX is being used (since the RTX hardware already does what it does - except exponentially faster.) The OptiX Prime acceleration toggle in DS version 4.12.0.86 has no actual effect on whether OptiX Prime is used or not (as of this DS version, OptiX Prime is always active for non-RTX cards and never active for RTX cards with no capability for the user to change it.) The control for it in DS was only left there because of an oversight on Nvidia's part (they forgot to mention it's feature deprecation in their documentation of Iray RTX.)

xionis_2946116

So as of lastnight I noticed some of my scenes were getting close to my 8GB limit and I was thinking about taking my cards back in exchange for the 2080 Ti, if I understand correctly memory pooling in the beta only works with the 2080 Ti, Titan RTX, and quaddro correct? I just would like to not have to worry about the memory limit....I would get a RTX titan, but it wouldn't match the rest of my system and would slowly drive me insane. Plus I think having an additional 3GB would be plenty in most cases for what I use it for.

RayDAnt

xionis said:

So as of lastnight I noticed some of my scenes were getting close to my 8GB limit and I was thinking about taking my cards back in exchange for the 2080 Ti, if I understand correctly memory pooling in the beta only works with the 2080 Ti, Titan RTX, and quaddro correct? I just would like to not have to worry about the memory limit....I would get a RTX titan, but it wouldn't match the rest of my system and would slowly drive me insane. Plus I think having an additional 3GB would be plenty in most cases for what I use it for.

2080's have EXACTLY the same level of hardware/driver support for memory pooling as 2080Ti's. So if memory pooling currently works with one it will also work with the other. As of right now (to my knowledge) no one with the necessary hardware (two 2080's or 2080Ti's and an NVLink bridge) has managed to determine with a level of certainty whether memory pooling actually currently works and what driver/Daz Studio settings are required to make that so.

Even if you do decide to go for the single 2080Ti instead (which btw will be significantly slower at rendering than the dual 2080 SUPER setup you have right now), if you could do some testing into seeing whether memory pooling with what you have IS possible (whether you can get a scene with 8GB+ scene data in it to render on just GPU) you would be doing others considering a dual RTX card setup for Iray a great service. As far as I know, you are the first person to post on this forum with a dual RTX card + NVLink bridge setup with cards in a low enough gigabyte range to be realistically tested for memory pooling functionality (eg. so far the biggest scene I've ended up loading on my Titan RTX was about 15GB - well short of its 24GB limit.) And those lower capacity, better performance per dollar cards like the 2080 or 2070 SUPER are the ones where knowing whether memory pooling can be made to work is really potentially useful.

tj_1ca9500b

@xionis

Thank you for posting your 3970X results in the benchmarking thread, along with the Dual GPU only results along with your GPU + CPU results.  I'm responding here as I don't want to clutter the other thread unnecessarily, per request of the OP.

As someone that ended up moviing my rendering from dual 1080s to a single 1080 Ti, I can tell you that the extra 3 GB of memory comes into play often enough for me.  Not having to spend additional time optimizing scenes that are larger than the dual cards were able to handle is a timesaver.  Sure, optimizing scenes is probably a good idea anyways, and should help reduce render times a bit, but of course there are tradeoffs in some cases here, i.e. lower resolution texture maps and such.

I'm still hitting the 11 GB wall often enough though.  I have a thing for rendering a half dozen or more characters at once, so I still need to optimize in those cases, or just let the CPU only render run for a bit if I'm just looking to 'error check' my render, stopping the render once there's sufficient detail to spot pokethru, etc.  Plus, more complex scenes, at least from a raytracing standpoint, also come into play here when characters are added. But at least I'm hitting that wall less often.  The dual 1080s were still faster than my 1080 Ti though, all other things being equal.

The fact that I still hit the 11 GB wall often enough in my useage cases DOES have me seriously considering an RTX Titan though. 

Checking your results from the other thread, it looks like the 3970X Threadripper, with it's 32 cores is about as fast as a single GTX 1080, and a bit more than twice as fast as the 1950X Threadripper result (11 minutes 30.8 seconds  vs 27.minutes 28.35 seconds).  This could mean that the upcoming 3990X Threadrippers could clock in at around 6 minutes by comparison, but of course spending 4K on a CPU vs 2.5K on an RTX Titan, well the Titan completes the same benchmark in around 4 minutes... That's actually closer than I thought it would be...The 3990X may be able to handle larger scenes though, but I'm guessing you wouldn't hit the 24GB VRAM limit much at all.  Of course, if you are rendering using 3Delight, then that's a completely different story.

Anyways, back to your situation.  Comparing the dual 2080 Supers to a single 2080 Ti, your dual cards are renderring the scene nearly twice as fast.  So I guess you'll need to ask yourself how often you may be willing to optimize your larger scenes, or do said scenes in multiple passes, rendering a portion of your characters, etc. in each pass and merging stuff in photoshop afterwords and such.

There is a third option, that is picking up a 2080 Ti later, which would give you 3 cards to work witth, assuming that you aren't using a third PCIe x16 slot for something else.  Since you are using a TRX40 board, at minimum you should be able to fit 3 cards in there, possibly 4 depending on which motherboard you picked.

Anyways, you can think of that 3970X Threadripper as a 'rich man's' version of the GTX 1080/1660, if that helps, but with a much higher memory limit, for those times that you see the dreaded 'scene rendering on CPU only' message.

xionis_2946116

RayDAnt said:

xionis said:

So as of lastnight I noticed some of my scenes were getting close to my 8GB limit and I was thinking about taking my cards back in exchange for the 2080 Ti, if I understand correctly memory pooling in the beta only works with the 2080 Ti, Titan RTX, and quaddro correct? I just would like to not have to worry about the memory limit....I would get a RTX titan, but it wouldn't match the rest of my system and would slowly drive me insane. Plus I think having an additional 3GB would be plenty in most cases for what I use it for.

2080's have EXACTLY the same level of hardware/driver support for memory pooling as 2080Ti's. So if memory pooling currently works with one it will also work with the other. As of right now (to my knowledge) no one with the necessary hardware (two 2080's or 2080Ti's and an NVLink bridge) has managed to determine with a level of certainty whether memory pooling actually currently works and what driver/Daz Studio settings are required to make that so.

Even if you do decide to go for the single 2080Ti instead (which btw will be significantly slower at rendering than the dual 2080 SUPER setup you have right now), if you could do some testing into seeing whether memory pooling with what you have IS possible (whether you can get a scene with 8GB+ scene data in it to render on just GPU) you would be doing others considering a dual RTX card setup for Iray a great service. As far as I know, you are the first person to post on this forum with a dual RTX card + NVLink bridge setup with cards in a low enough gigabyte range to be realistically tested for memory pooling functionality (eg. so far the biggest scene I've ended up loading on my Titan RTX was about 15GB - well short of its 24GB limit.) And those lower capacity, better performance per dollar cards like the 2080 or 2070 SUPER are the ones where knowing whether memory pooling can be made to work is really potentially useful.

I was actually considering getting 2 2080 Ti's....I did test a scene that seems to exceed my memory limit in the beta, I think it's just over the 8GB because it kept forcing me to use CPU. I tried the pooling in the beta, by setting the NVlink group size to 1 and 2 (I wasn't sure what to put that there) and it completely would crash to desktop without any error messages once I started the render. I'd be happy to try to test it to get it to work, just let me know what settings to change or what to try.

RayDAnt

xionis said:

RayDAnt said:

xionis said:

So as of lastnight I noticed some of my scenes were getting close to my 8GB limit and I was thinking about taking my cards back in exchange for the 2080 Ti, if I understand correctly memory pooling in the beta only works with the 2080 Ti, Titan RTX, and quaddro correct? I just would like to not have to worry about the memory limit....I would get a RTX titan, but it wouldn't match the rest of my system and would slowly drive me insane. Plus I think having an additional 3GB would be plenty in most cases for what I use it for.

2080's have EXACTLY the same level of hardware/driver support for memory pooling as 2080Ti's. So if memory pooling currently works with one it will also work with the other. As of right now (to my knowledge) no one with the necessary hardware (two 2080's or 2080Ti's and an NVLink bridge) has managed to determine with a level of certainty whether memory pooling actually currently works and what driver/Daz Studio settings are required to make that so.

Even if you do decide to go for the single 2080Ti instead (which btw will be significantly slower at rendering than the dual 2080 SUPER setup you have right now), if you could do some testing into seeing whether memory pooling with what you have IS possible (whether you can get a scene with 8GB+ scene data in it to render on just GPU) you would be doing others considering a dual RTX card setup for Iray a great service. As far as I know, you are the first person to post on this forum with a dual RTX card + NVLink bridge setup with cards in a low enough gigabyte range to be realistically tested for memory pooling functionality (eg. so far the biggest scene I've ended up loading on my Titan RTX was about 15GB - well short of its 24GB limit.) And those lower capacity, better performance per dollar cards like the 2080 or 2070 SUPER are the ones where knowing whether memory pooling can be made to work is really potentially useful.

I was actually considering getting 2 2080 Ti's....I did test a scene that seems to exceed my memory limit in the beta, I think it's just over the 8GB because it kept forcing me to use CPU. I tried the pooling in the beta, by setting the NVlink group size to 1 and 2 (I wasn't sure what to put that there) and it completely would crash to desktop without any error messages once I started the render. I'd be happy to try to test it to get it to work, just let me know what settings to change or what to try.

That would be awesome! I'll see about putting together a test regimen for you and posting it in this thread tomorrow.

xionis_2946116

tj_1ca9500b said:

@xionis

Thank you for posting your 3970X results in the benchmarking thread, along with the Dual GPU only results along with your GPU + CPU results.  I'm responding here as I don't want to clutter the other thread unnecessarily, per request of the OP.

As someone that ended up moviing my rendering from dual 1080s to a single 1080 Ti, I can tell you that the extra 3 GB of memory comes into play often enough for me.  Not having to spend additional time optimizing scenes that are larger than the dual cards were able to handle is a timesaver.  Sure, optimizing scenes is probably a good idea anyways, and should help reduce render times a bit, but of course there are tradeoffs in some cases here, i.e. lower resolution texture maps and such.

I'm still hitting the 11 GB wall often enough though.  I have a thing for rendering a half dozen or more characters at once, so I still need to optimize in those cases, or just let the CPU only render run for a bit if I'm just looking to 'error check' my render, stopping the render once there's sufficient detail to spot pokethru, etc.  Plus, more complex scenes, at least from a raytracing standpoint, also come into play here when characters are added. But at least I'm hitting that wall less often.  The dual 1080s were still faster than my 1080 Ti though, all other things being equal.

The fact that I still hit the 11 GB wall often enough in my useage cases DOES have me seriously considering an RTX Titan though. 

Checking your results from the other thread, it looks like the 3970X Threadripper, with it's 32 cores is about as fast as a single GTX 1080, and a bit more than twice as fast as the 1950X Threadripper result (11 minutes 30.8 seconds  vs 27.minutes 28.35 seconds).  This could mean that the upcoming 3990X Threadrippers could clock in at around 6 minutes by comparison, but of course spending 4K on a CPU vs 2.5K on an RTX Titan, well the Titan completes the same benchmark in around 4 minutes... That's actually closer than I thought it would be...The 3990X may be able to handle larger scenes though, but I'm guessing you wouldn't hit the 24GB VRAM limit much at all.  Of course, if you are rendering using 3Delight, then that's a completely different story.

Anyways, back to your situation.  Comparing the dual 2080 Supers to a single 2080 Ti, your dual cards are renderring the scene nearly twice as fast.  So I guess you'll need to ask yourself how often you may be willing to optimize your larger scenes, or do said scenes in multiple passes, rendering a portion of your characters, etc. in each pass and merging stuff in photoshop afterwords and such.

There is a third option, that is picking up a 2080 Ti later, which would give you 3 cards to work witth, assuming that you aren't using a third PCIe x16 slot for something else.  Since you are using a TRX40 board, at minimum you should be able to fit 3 cards in there, possibly 4 depending on which motherboard you picked.

Anyways, you can think of that 3970X Threadripper as a 'rich man's' version of the GTX 1080/1660, if that helps, but with a much higher memory limit, for those times that you see the dreaded 'scene rendering on CPU only' message.

 

You're very welcome for posting the results. I was having alot of fun running them.....until I found out I had to do math at the end :facepalm: I think Daz should include a benchmark feature like some games do.

After running the benchmarks and seeing what a big improvement there was over my single card i'm no longer considering replacing these 2 cards with just 1 2080 Ti, because I can't sacrifice rendering speed. But like you I don't want to spend my time optimizing a scene, i already spend enough time posing. I did consider the RTX titan, but the fact that it comes in gold and doesn't have the RGB that my other cards do is a deal breaker for me because it just wouldn't look right with the rest of my system. I know it's a stupid reason to not get that instead, but to be honest i'm just looking to get something to hold me off til the 3080Ti comes out later this year, because I purchased a 2 year protection plan that will allow me to trade my cards in for the newer ones once they come out and I hope the rumor of the having at least 20GB is true, then I think it would be perfect...but until then I just want something that will allow me get past this 8GB barrier but still be just as fast as these 2 cards if not faster. So if I could pool the memory on these then I'll probably be happy to stick with these. 

tj_1ca9500b

xionis said:

tj_1ca9500b said:

@xionis

Thank you for posting your 3970X results in the benchmarking thread, along with the Dual GPU only results along with your GPU + CPU results.  I'm responding here as I don't want to clutter the other thread unnecessarily, per request of the OP.

As someone that ended up moviing my rendering from dual 1080s to a single 1080 Ti, I can tell you that the extra 3 GB of memory comes into play often enough for me.  Not having to spend additional time optimizing scenes that are larger than the dual cards were able to handle is a timesaver.  Sure, optimizing scenes is probably a good idea anyways, and should help reduce render times a bit, but of course there are tradeoffs in some cases here, i.e. lower resolution texture maps and such.

I'm still hitting the 11 GB wall often enough though.  I have a thing for rendering a half dozen or more characters at once, so I still need to optimize in those cases, or just let the CPU only render run for a bit if I'm just looking to 'error check' my render, stopping the render once there's sufficient detail to spot pokethru, etc.  Plus, more complex scenes, at least from a raytracing standpoint, also come into play here when characters are added. But at least I'm hitting that wall less often.  The dual 1080s were still faster than my 1080 Ti though, all other things being equal.

The fact that I still hit the 11 GB wall often enough in my useage cases DOES have me seriously considering an RTX Titan though. 

Checking your results from the other thread, it looks like the 3970X Threadripper, with it's 32 cores is about as fast as a single GTX 1080, and a bit more than twice as fast as the 1950X Threadripper result (11 minutes 30.8 seconds  vs 27.minutes 28.35 seconds).  This could mean that the upcoming 3990X Threadrippers could clock in at around 6 minutes by comparison, but of course spending 4K on a CPU vs 2.5K on an RTX Titan, well the Titan completes the same benchmark in around 4 minutes... That's actually closer than I thought it would be...The 3990X may be able to handle larger scenes though, but I'm guessing you wouldn't hit the 24GB VRAM limit much at all.  Of course, if you are rendering using 3Delight, then that's a completely different story.

Anyways, back to your situation.  Comparing the dual 2080 Supers to a single 2080 Ti, your dual cards are renderring the scene nearly twice as fast.  So I guess you'll need to ask yourself how often you may be willing to optimize your larger scenes, or do said scenes in multiple passes, rendering a portion of your characters, etc. in each pass and merging stuff in photoshop afterwords and such.

There is a third option, that is picking up a 2080 Ti later, which would give you 3 cards to work witth, assuming that you aren't using a third PCIe x16 slot for something else.  Since you are using a TRX40 board, at minimum you should be able to fit 3 cards in there, possibly 4 depending on which motherboard you picked.

Anyways, you can think of that 3970X Threadripper as a 'rich man's' version of the GTX 1080/1660, if that helps, but with a much higher memory limit, for those times that you see the dreaded 'scene rendering on CPU only' message.

 

You're very welcome for posting the results. I was having alot of fun running them.....until I found out I had to do math at the end :facepalm: I think Daz should include a benchmark feature like some games do.

After running the benchmarks and seeing what a big improvement there was over my single card i'm no longer considering replacing these 2 cards with just 1 2080 Ti, because I can't sacrifice rendering speed. But like you I don't want to spend my time optimizing a scene, i already spend enough time posing. I did consider the RTX titan, but the fact that it comes in gold and doesn't have the RGB that my other cards do is a deal breaker for me because it just wouldn't look right with the rest of my system. I know it's a stupid reason to not get that instead, but to be honest i'm just looking to get something to hold me off til the 3080Ti comes out later this year, because I purchased a 2 year protection plan that will allow me to trade my cards in for the newer ones once they come out and I hope the rumor of the having at least 20GB is true, then I think it would be perfect...but until then I just want something that will allow me get past this 8GB barrier but still be just as fast as these 2 cards if not faster. So if I could pool the memory on these then I'll probably be happy to stick with these. 

Keep in mind that you SHOULD be able to use all 3 GPUs when rendering, if you were to incorporate a third card.  Of course, if you can just get a refund on the 2080s and upgrade the cards to the 2080 Ti's, that's fine too.

From what I've read, Daz will simply 'drop' the cards with the smaller VRAM counts if a render is too large for the smaller cards, and just use the cards that the scenes can actually fit in.  So if you were to go with a dual 2080 + 2080 Ti setup, for smaller scenes you'd be rendering your scenes roughly 1.5x faster with 3 cards vs 2 cards, and for the larger scenes you'd still have a single card doing the render.  Another possibility here is to go with 2 2080 Ti's + a 2080 or smaller card, using the smaller card to drive your Daz viewport and desktop environment.  This way, you can continue to work using a separate Daz instance on your next scene without significant monitor lag using the third card. 

I actually do this wtih my integrated Ryzen APU, with the 1080 Ti rendering in the background while I work on something else, and the Ryzen APU also handles my non-Iray Daz viewport duties.  Of course, Threadrippers don't have integrated APUs, so not an option in your case, but I'm pretty happy with how this works out.  Admittedly, sometimes I'm just gaming while I'm waiting on a render, playing turn based games and such, but it's nice to be able to work on other stuff without the system slowing to a crawl while a render is baking.

I read recently where the more recent versions of Daz weren't allowing multiple Daz instances to run, and a bunch of people complained.  Hopefully this has been rectified, so that you can still work in a second Daz instance on your next scene while the first instance is busy baking a render or whatever.  I also use additional Daz instances to extract subsets of scenes so that I can just save the subset stuff I need, and then load that subset into the first instance, without having to load the rest of that other scene into the first instance or quit and then re-load said scene.  The multiple instances thing is a big deal for me.

I also have a theory that since Windows isn't using the 1080 Ti for anything, as the desktop runs on the Ryzen graphics, that I may be avoiding the 18% Windows 10 VRAM tax, giving me just a bit more memory to work with (i.e. the full 11GB instead of just 9.02), but I've never tried to verify that.  Windows 10 normally reserves 10% + 10% of the 90% remaining of each GPUs VRAM for no partiularly good reason.  This is something that people have complained about for years, but unless it was finally fixed recently, it's been an ongoing issue.  The Quadro cards and such supposedly have a software fix for this issue, at least from what I've read.  I'd have no problem if windows just reserved a set amont of ram for desktop duties, but the percentage thing is particularly aggregious, as the Windows desktop will run on relatively tiny cards with say 1-2 GB of VRAM, indicating that it doesn't NEED multiple GB of VRAM just to run the desktop.  But I digress...

My point here is that having a 3rd card that is 100% dedicated to desktop duties and maybe light gaming while you have renders baking works well with Daz.  You can even mix GPU brands for this!  Note that AMD GPUs won't help with your renders though, just your desktop environment in this situation.  An extra Nvidia card, though...

xionis_2946116

tj_1ca9500b said:

xionis said:

tj_1ca9500b said:

@xionis

Thank you for posting your 3970X results in the benchmarking thread, along with the Dual GPU only results along with your GPU + CPU results.  I'm responding here as I don't want to clutter the other thread unnecessarily, per request of the OP.

As someone that ended up moviing my rendering from dual 1080s to a single 1080 Ti, I can tell you that the extra 3 GB of memory comes into play often enough for me.  Not having to spend additional time optimizing scenes that are larger than the dual cards were able to handle is a timesaver.  Sure, optimizing scenes is probably a good idea anyways, and should help reduce render times a bit, but of course there are tradeoffs in some cases here, i.e. lower resolution texture maps and such.

I'm still hitting the 11 GB wall often enough though.  I have a thing for rendering a half dozen or more characters at once, so I still need to optimize in those cases, or just let the CPU only render run for a bit if I'm just looking to 'error check' my render, stopping the render once there's sufficient detail to spot pokethru, etc.  Plus, more complex scenes, at least from a raytracing standpoint, also come into play here when characters are added. But at least I'm hitting that wall less often.  The dual 1080s were still faster than my 1080 Ti though, all other things being equal.

The fact that I still hit the 11 GB wall often enough in my useage cases DOES have me seriously considering an RTX Titan though. 

Checking your results from the other thread, it looks like the 3970X Threadripper, with it's 32 cores is about as fast as a single GTX 1080, and a bit more than twice as fast as the 1950X Threadripper result (11 minutes 30.8 seconds  vs 27.minutes 28.35 seconds).  This could mean that the upcoming 3990X Threadrippers could clock in at around 6 minutes by comparison, but of course spending 4K on a CPU vs 2.5K on an RTX Titan, well the Titan completes the same benchmark in around 4 minutes... That's actually closer than I thought it would be...The 3990X may be able to handle larger scenes though, but I'm guessing you wouldn't hit the 24GB VRAM limit much at all.  Of course, if you are rendering using 3Delight, then that's a completely different story.

Anyways, back to your situation.  Comparing the dual 2080 Supers to a single 2080 Ti, your dual cards are renderring the scene nearly twice as fast.  So I guess you'll need to ask yourself how often you may be willing to optimize your larger scenes, or do said scenes in multiple passes, rendering a portion of your characters, etc. in each pass and merging stuff in photoshop afterwords and such.

There is a third option, that is picking up a 2080 Ti later, which would give you 3 cards to work witth, assuming that you aren't using a third PCIe x16 slot for something else.  Since you are using a TRX40 board, at minimum you should be able to fit 3 cards in there, possibly 4 depending on which motherboard you picked.

Anyways, you can think of that 3970X Threadripper as a 'rich man's' version of the GTX 1080/1660, if that helps, but with a much higher memory limit, for those times that you see the dreaded 'scene rendering on CPU only' message.

 

You're very welcome for posting the results. I was having alot of fun running them.....until I found out I had to do math at the end :facepalm: I think Daz should include a benchmark feature like some games do.

After running the benchmarks and seeing what a big improvement there was over my single card i'm no longer considering replacing these 2 cards with just 1 2080 Ti, because I can't sacrifice rendering speed. But like you I don't want to spend my time optimizing a scene, i already spend enough time posing. I did consider the RTX titan, but the fact that it comes in gold and doesn't have the RGB that my other cards do is a deal breaker for me because it just wouldn't look right with the rest of my system. I know it's a stupid reason to not get that instead, but to be honest i'm just looking to get something to hold me off til the 3080Ti comes out later this year, because I purchased a 2 year protection plan that will allow me to trade my cards in for the newer ones once they come out and I hope the rumor of the having at least 20GB is true, then I think it would be perfect...but until then I just want something that will allow me get past this 8GB barrier but still be just as fast as these 2 cards if not faster. So if I could pool the memory on these then I'll probably be happy to stick with these. 

Keep in mind that you SHOULD be able to use all 3 GPUs when rendering, if you were to incorporate a third card.  Of course, if you can just get a refund on the 2080s and upgrade the cards to the 2080 Ti's, that's fine too.

From what I've read, Daz will simply 'drop' the cards with the smaller VRAM counts if a render is too large for the smaller cards, and just use the cards that the scenes can actually fit in.  So if you were to go with a dual 2080 + 2080 Ti setup, for smaller scenes you'd be rendering your scenes roughly 1.5x faster with 3 cards vs 2 cards, and for the larger scenes you'd still have a single card doing the render.  Another possibility here is to go with 2 2080 Ti's + a 2080 or smaller card, using the smaller card to drive your Daz viewport and desktop environment.  This way, you can continue to work using a separate Daz instance on your next scene without significant monitor lag using the third card. 

I actually do this wtih my integrated Ryzen APU, with the 1080 Ti rendering in the background while I work on something else, and the Ryzen APU also handles my non-Iray Daz viewport duties.  Of course, Threadrippers don't have integrated APUs, so not an option in your case, but I'm pretty happy with how this works out.  Admittedly, sometimes I'm just gaming while I'm waiting on a render, playing turn based games and such, but it's nice to be able to work on other stuff without the system slowing to a crawl while a render is baking.

I read recently where the more recent versions of Daz weren't allowing multiple Daz instances to run, and a bunch of people complained.  Hopefully this has been rectified, so that you can still work in a second Daz instance on your next scene while the first instance is busy baking a render or whatever.  I also use additional Daz instances to extract subsets of scenes so that I can just save the subset stuff I need, and then load that subset into the first instance, without having to load the rest of that other scene into the first instance or quit and then re-load said scene.  The multiple instances thing is a big deal for me.

I also have a theory that since Windows isn't using the 1080 Ti for anything, as the desktop runs on the Ryzen graphics, that I may be avoiding the 18% Windows 10 VRAM tax, giving me just a bit more memory to work with (i.e. the full 11GB instead of just 9.02), but I've never tried to verify that.  Windows 10 normally reserves 10% + 10% of the 90% remaining of each GPUs VRAM for no partiularly good reason.  This is something that people have complained about for years, but unless it was finally fixed recently, it's been an ongoing issue.  The Quadro cards and such supposedly have a software fix for this issue, at least from what I've read.  I'd have no problem if windows just reserved a set amont of ram for desktop duties, but the percentage thing is particularly aggregious, as the Windows desktop will run on relatively tiny cards with say 1-2 GB of VRAM, indicating that it doesn't NEED multiple GB of VRAM just to run the desktop.  But I digress...

My point here is that having a 3rd card that is 100% dedicated to desktop duties and maybe light gaming while you have renders baking works well with Daz.  You can even mix GPU brands for this!  Note that AMD GPUs won't help with your renders though, just your desktop environment in this situation.  An extra Nvidia card, though...

Sorry I forgot to mention in my past post that I looked and I definitely don't have room for a 3rd card, because the cards I have take up 3 slots each ... Unless it's a single slot card, but I'd also be worried about heat soaking that top card I already have 10 fans in this case (lian li O11 Dynamic XL) trying to keep this threadripper and these 2 cards as cool as possible.... I could liquid cool the cards and then easily be able to fit 3 of them but the reason I don't do that is because it would void my replacement warranty I purchased. I don't really do too much else on my computer since I started using Daz last year, I've just been trying to learn as much as possible and keep my system rendering while I'm at work or sleep. So I don't do much of gaming anymore until cyberpunk 2077 finally decides to release... then I'll probably render alot less lol

xionis_2946116

RayDAnt said:

xionis said:

RayDAnt said:

xionis said:

So as of lastnight I noticed some of my scenes were getting close to my 8GB limit and I was thinking about taking my cards back in exchange for the 2080 Ti, if I understand correctly memory pooling in the beta only works with the 2080 Ti, Titan RTX, and quaddro correct? I just would like to not have to worry about the memory limit....I would get a RTX titan, but it wouldn't match the rest of my system and would slowly drive me insane. Plus I think having an additional 3GB would be plenty in most cases for what I use it for.

2080's have EXACTLY the same level of hardware/driver support for memory pooling as 2080Ti's. So if memory pooling currently works with one it will also work with the other. As of right now (to my knowledge) no one with the necessary hardware (two 2080's or 2080Ti's and an NVLink bridge) has managed to determine with a level of certainty whether memory pooling actually currently works and what driver/Daz Studio settings are required to make that so.

Even if you do decide to go for the single 2080Ti instead (which btw will be significantly slower at rendering than the dual 2080 SUPER setup you have right now), if you could do some testing into seeing whether memory pooling with what you have IS possible (whether you can get a scene with 8GB+ scene data in it to render on just GPU) you would be doing others considering a dual RTX card setup for Iray a great service. As far as I know, you are the first person to post on this forum with a dual RTX card + NVLink bridge setup with cards in a low enough gigabyte range to be realistically tested for memory pooling functionality (eg. so far the biggest scene I've ended up loading on my Titan RTX was about 15GB - well short of its 24GB limit.) And those lower capacity, better performance per dollar cards like the 2080 or 2070 SUPER are the ones where knowing whether memory pooling can be made to work is really potentially useful.

I was actually considering getting 2 2080 Ti's....I did test a scene that seems to exceed my memory limit in the beta, I think it's just over the 8GB because it kept forcing me to use CPU. I tried the pooling in the beta, by setting the NVlink group size to 1 and 2 (I wasn't sure what to put that there) and it completely would crash to desktop without any error messages once I started the render. I'd be happy to try to test it to get it to work, just let me know what settings to change or what to try.

That would be awesome! I'll see about putting together a test regimen for you and posting it in this thread tomorrow.

I'll be happy to do it this weekend, If it doesn't work then I'll exchange them for ti's and we can try it again.

tj_1ca9500b

xionis said:

tj_1ca9500b said:

xionis said:

tj_1ca9500b said:

@xionis

Thank you for posting your 3970X results in the benchmarking thread, along with the Dual GPU only results along with your GPU + CPU results.  I'm responding here as I don't want to clutter the other thread unnecessarily, per request of the OP.

As someone that ended up moviing my rendering from dual 1080s to a single 1080 Ti, I can tell you that the extra 3 GB of memory comes into play often enough for me.  Not having to spend additional time optimizing scenes that are larger than the dual cards were able to handle is a timesaver.  Sure, optimizing scenes is probably a good idea anyways, and should help reduce render times a bit, but of course there are tradeoffs in some cases here, i.e. lower resolution texture maps and such.

I'm still hitting the 11 GB wall often enough though.  I have a thing for rendering a half dozen or more characters at once, so I still need to optimize in those cases, or just let the CPU only render run for a bit if I'm just looking to 'error check' my render, stopping the render once there's sufficient detail to spot pokethru, etc.  Plus, more complex scenes, at least from a raytracing standpoint, also come into play here when characters are added. But at least I'm hitting that wall less often.  The dual 1080s were still faster than my 1080 Ti though, all other things being equal.

The fact that I still hit the 11 GB wall often enough in my useage cases DOES have me seriously considering an RTX Titan though. 

Checking your results from the other thread, it looks like the 3970X Threadripper, with it's 32 cores is about as fast as a single GTX 1080, and a bit more than twice as fast as the 1950X Threadripper result (11 minutes 30.8 seconds  vs 27.minutes 28.35 seconds).  This could mean that the upcoming 3990X Threadrippers could clock in at around 6 minutes by comparison, but of course spending 4K on a CPU vs 2.5K on an RTX Titan, well the Titan completes the same benchmark in around 4 minutes... That's actually closer than I thought it would be...The 3990X may be able to handle larger scenes though, but I'm guessing you wouldn't hit the 24GB VRAM limit much at all.  Of course, if you are rendering using 3Delight, then that's a completely different story.

Anyways, back to your situation.  Comparing the dual 2080 Supers to a single 2080 Ti, your dual cards are renderring the scene nearly twice as fast.  So I guess you'll need to ask yourself how often you may be willing to optimize your larger scenes, or do said scenes in multiple passes, rendering a portion of your characters, etc. in each pass and merging stuff in photoshop afterwords and such.

There is a third option, that is picking up a 2080 Ti later, which would give you 3 cards to work witth, assuming that you aren't using a third PCIe x16 slot for something else.  Since you are using a TRX40 board, at minimum you should be able to fit 3 cards in there, possibly 4 depending on which motherboard you picked.

Anyways, you can think of that 3970X Threadripper as a 'rich man's' version of the GTX 1080/1660, if that helps, but with a much higher memory limit, for those times that you see the dreaded 'scene rendering on CPU only' message.

 

You're very welcome for posting the results. I was having alot of fun running them.....until I found out I had to do math at the end :facepalm: I think Daz should include a benchmark feature like some games do.

After running the benchmarks and seeing what a big improvement there was over my single card i'm no longer considering replacing these 2 cards with just 1 2080 Ti, because I can't sacrifice rendering speed. But like you I don't want to spend my time optimizing a scene, i already spend enough time posing. I did consider the RTX titan, but the fact that it comes in gold and doesn't have the RGB that my other cards do is a deal breaker for me because it just wouldn't look right with the rest of my system. I know it's a stupid reason to not get that instead, but to be honest i'm just looking to get something to hold me off til the 3080Ti comes out later this year, because I purchased a 2 year protection plan that will allow me to trade my cards in for the newer ones once they come out and I hope the rumor of the having at least 20GB is true, then I think it would be perfect...but until then I just want something that will allow me get past this 8GB barrier but still be just as fast as these 2 cards if not faster. So if I could pool the memory on these then I'll probably be happy to stick with these. 

Keep in mind that you SHOULD be able to use all 3 GPUs when rendering, if you were to incorporate a third card.  Of course, if you can just get a refund on the 2080s and upgrade the cards to the 2080 Ti's, that's fine too.

From what I've read, Daz will simply 'drop' the cards with the smaller VRAM counts if a render is too large for the smaller cards, and just use the cards that the scenes can actually fit in.  So if you were to go with a dual 2080 + 2080 Ti setup, for smaller scenes you'd be rendering your scenes roughly 1.5x faster with 3 cards vs 2 cards, and for the larger scenes you'd still have a single card doing the render.  Another possibility here is to go with 2 2080 Ti's + a 2080 or smaller card, using the smaller card to drive your Daz viewport and desktop environment.  This way, you can continue to work using a separate Daz instance on your next scene without significant monitor lag using the third card. 

I actually do this wtih my integrated Ryzen APU, with the 1080 Ti rendering in the background while I work on something else, and the Ryzen APU also handles my non-Iray Daz viewport duties.  Of course, Threadrippers don't have integrated APUs, so not an option in your case, but I'm pretty happy with how this works out.  Admittedly, sometimes I'm just gaming while I'm waiting on a render, playing turn based games and such, but it's nice to be able to work on other stuff without the system slowing to a crawl while a render is baking.

I read recently where the more recent versions of Daz weren't allowing multiple Daz instances to run, and a bunch of people complained.  Hopefully this has been rectified, so that you can still work in a second Daz instance on your next scene while the first instance is busy baking a render or whatever.  I also use additional Daz instances to extract subsets of scenes so that I can just save the subset stuff I need, and then load that subset into the first instance, without having to load the rest of that other scene into the first instance or quit and then re-load said scene.  The multiple instances thing is a big deal for me.

I also have a theory that since Windows isn't using the 1080 Ti for anything, as the desktop runs on the Ryzen graphics, that I may be avoiding the 18% Windows 10 VRAM tax, giving me just a bit more memory to work with (i.e. the full 11GB instead of just 9.02), but I've never tried to verify that.  Windows 10 normally reserves 10% + 10% of the 90% remaining of each GPUs VRAM for no partiularly good reason.  This is something that people have complained about for years, but unless it was finally fixed recently, it's been an ongoing issue.  The Quadro cards and such supposedly have a software fix for this issue, at least from what I've read.  I'd have no problem if windows just reserved a set amont of ram for desktop duties, but the percentage thing is particularly aggregious, as the Windows desktop will run on relatively tiny cards with say 1-2 GB of VRAM, indicating that it doesn't NEED multiple GB of VRAM just to run the desktop.  But I digress...

My point here is that having a 3rd card that is 100% dedicated to desktop duties and maybe light gaming while you have renders baking works well with Daz.  You can even mix GPU brands for this!  Note that AMD GPUs won't help with your renders though, just your desktop environment in this situation.  An extra Nvidia card, though...

 

Sorry I forgot to mention in my past post that I looked and I definitely don't have room for a 3rd card, because the cards I have take up 3 slots each ... Unless it's a single slot card, but I'd also be worried about heat soaking that top card I already have 10 fans in this case (lian li O11 Dynamic XL) trying to keep this threadripper and these 2 cards as cool as possible.... I could liquid cool the cards and then easily be able to fit 3 of them but the reason I don't do that is because it would void my replacement warranty I purchased. I don't really do too much else on my computer since I started using Daz last year, I've just been trying to learn as much as possible and keep my system rendering while I'm at work or sleep. So I don't do much of gaming anymore until cyberpunk 2077 finally decides to release... then I'll probably render alot less lol

Ahh well then.  It's something to consider though, if you end up trading in the 2080 Max Q's, i.e. looking for 2 slot 2080 TI's, or perhaps a Waterforce card or some other 'hybrid' water cooled retail card.