Highest consumed amount of VRAM in a single render

Hello!

I know this might sound stupid, but I'm batting around the idea of an Nvidia Quadro RTX 8000 with 48 GB of VRAM.  You might call me crazy, but that's a different topic.

I'd like to know if any of you have run renders, either in DAZ or elsewhere, that demanded more than the 11 or 12 GB of VRAM available on high-end consumer cards.

Does anybody with a 2080 or RTX Titan ever see it drop to CPU?

Thank you in advance!

Comments

  • Can you link to the thread about you being crazy ????
  • duckbombduckbomb Posts: 585

    Haha I was just being sarcastic...  the truth of the matter is I have a situation at work where we're parting out some cluster workstations from a Business Unit that's been sold, and I have an opportunity to pick one up for a pretty fantastic price.  It still isn't cheap, though, so I wanted to pose the question.

    Unfortunately, I really don't have a good reference point as my current GPU is an AMD so I always CPU render as it is. I know just about any current-gen Nvidia card will be better than this, but if I skip this opportunity because of a (relatively) small price difference between this and a new RTX Titan I'm gonna kick myself if I ever run into a situation where I'm strapped for VRAM.

    I figure it can't hurt to ask...  ever run a scene that needed <24 Gigs?

  • I have a 1080ti and I've had a few scenes drop to CPU but with scene optimizer that was fixable. The worst offender is a certain outdoor set with a random tree set for some just nuts subD (4? higher maybe?).

    The RTX titan has 24Gb of VRAM. If the price is close for Quadro, within a couple of hundred, I'd get it. It's the same basic GPU chip as the RTX Titan but clocked a little lower so it should last longer. Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

  • duckbomb said:

    Haha I was just being sarcastic...

    Me too. I'm the last person to call anyone for OTT extravagant purchases. Over the last 3 years I've spent more than the cost of your quadro on a certain musical. Regular trips from the North of England to London, 2 trips to Toronto and one to New York. If it works for you and your happy that's all that matters.
  • kyoto kidkyoto kid Posts: 41,847

    ...crikey if you can get a good deal on that Quadro RTX 8000 I'd go for it even if it means eating canned chili for several weeks. 48 GB of VRAM, that's more than either of my systems have io the MB and only 8 GB less than both combined.

  • Matt_CastleMatt_Castle Posts: 3,004

    It'd be easily possible to get into double-digit GBs in terms of usage if you've got multiple characters in the scene, a set with high resolution textures, lots of high subdivision modifiers, fibre hair, etc. And that's only going to increase as content continues to try to push the boundaries.

    In this case, 48GB of VRAM would be a very impressive reserve that would mean that even the most adventurous scenes should remain on the GPU.

    If you can do it without it being a burden on your finances, then I'd say go for it. If nothing else, it'll mean you won't need to upgrade again quite so soon. (That's always my brother's policy on buying CPU parts - if you get the best stuff, then you won't have to upgrade it nearly as soon).

  • joseftjoseft Posts: 310

    Inside of Daz Studio it is probably quite rare for most users to need more than double digit GB of VRAM.

    But i can say that outside of Daz Studio, it is quite common to need more than that. Inside software like Maya that is designed to handle far greater poly counts than DS, and scenes that are far more complex than what 99% of people do inside DS, its not uncommon to need the kind of VRAM found on the big quadro cards, or even more. 

    Another thing to remember is that other programs like Maya/C4D/3DSM etc have multiple other render engines available, like Redshift/Octane (yes i know Octane is avaialble for DS), where even if your scene does not fit inside your GPU VRAM, it can still render using the GPU (out-of-core), so its not the end of the world if the scenes in question do require so much RAM. 

     I could link a few articles/images of projects done that had geometry in the billions of polygons, and textures on top of that, the ram requirements for something like that are in the hundreds of gigabytes.

    On a side note, i was using an experimental/beta build of another renderer recently, and i was running some tests that resulted in a crash. On investigation of the log file, it turned out that the renderer had tried to allocate 22 petabytes of vram. For those that dont know what a petabyte is, its 1 million gigabytes = 1000 terabytes. So apparently my render needed 22 million gigabytes of ram. It was ,of course, a bug, but i did find it quite funny.

     

  • Serious question - if you have any more available, hit me up.

  • HavosHavos Posts: 5,576

    Keep in mind if you want to use the full 48GB of VRAM, you will need a PC with 96GB or more to support it.

    In my opinion you can blow the VRAM on any card if you refuse to think at all about optimising. I think any scene, even those with many characters should be able to fit into top of the range consumer cards (ie 8-11GB), if you are willing to spend at least some time optimising your scene. The only reason to need a huge amount of VRAM is if you want to render at very large resolution, ie 4Kx4K or more.

  • AsariAsari Posts: 703
    I had one scene with a large background and many props. 2 G8 figures with HD morphs dialed in at subdivision 4. Both figures had strand-based-hair. Each hair was at tessellation 5 and consisted of 5 different pieces. Both figures were clothed with multiple parts. Some clothing featured 4k textures. I had visible mesh lights in the scene and also a 16k background HDRI. If I wanted to render in 8k resolution it exceeded my 11 GB vram. Even in 4k it dropped to CPU immediately. I reduced the resolution of many textures and rendered the image separately in different steps and added some missing shadows in postwork.
  • duckbombduckbomb Posts: 585

    Wow everybody thank you so much for your responses!  Honestly, I have to say I was expecting a sarcastic response to my question, so thank you so much for your help here!

    I suppose I have to agree with kyoto kid and just eat my Ramen Noodle soup for the next few weeks to make up for this, but considering what you've all said it looks to me like I'd regret  not just going full Montey on this thing.

     

    Havos said:

    Keep in mind if you want to use the full 48GB of VRAM, you will need a PC with 96GB or more to support it.

    In my opinion you can blow the VRAM on any card if you refuse to think at all about optimising. I think any scene, even those with many characters should be able to fit into top of the range consumer cards (ie 8-11GB), if you are willing to spend at least some time optimising your scene. The only reason to need a huge amount of VRAM is if you want to render at very large resolution, ie 4Kx4K or more.

    Do you mean I'd need 96 GB or more of regular RAM to support this?  I currently only have 64... I suppose that's something worth considering.  If I'm being honest, I have to admit I vary rarely render an optimized scene.  I tend to build up giant setpieces and pre-light them all, and then hide them in chunks to render out panels of a cohesive story.  I used to delete them outright, but I've too many times come back after a render and hit "save" without saving off a copy, first.  In addition, I do like to throw in some new mats from Substance on the fly, and since I don't really care for dialing things in at that level I tend to output everything at 8K and just apply it without optimizing.  I know it's a bad habit to work in, but I suppose I'm a victim of laziness and complacency here.  I don't suppose a Quadro is going to make that any better, you're probably right in that I should consider learning how to optimize and apply that to my workflow.  I do render in about 8K, which I know is crazy again, but it really helps with my post-processing and I use a lot of that natural compression to get rid of that plastic look.

     

    joseft said:

    Inside of Daz Studio it is probably quite rare for most users to need more than double digit GB of VRAM.

    But i can say that outside of Daz Studio, it is quite common to need more than that. Inside software like Maya that is designed to handle far greater poly counts than DS, and scenes that are far more complex than what 99% of people do inside DS, its not uncommon to need the kind of VRAM found on the big quadro cards, or even more. 

    Another thing to remember is that other programs like Maya/C4D/3DSM etc have multiple other render engines available, like Redshift/Octane (yes i know Octane is avaialble for DS), where even if your scene does not fit inside your GPU VRAM, it can still render using the GPU (out-of-core), so its not the end of the world if the scenes in question do require so much RAM. 

     I could link a few articles/images of projects done that had geometry in the billions of polygons, and textures on top of that, the ram requirements for something like that are in the hundreds of gigabytes.

    On a side note, i was using an experimental/beta build of another renderer recently, and i was running some tests that resulted in a crash. On investigation of the log file, it turned out that the renderer had tried to allocate 22 petabytes of vram. For those that dont know what a petabyte is, its 1 million gigabytes = 1000 terabytes. So apparently my render needed 22 million gigabytes of ram. It was ,of course, a bug, but i did find it quite funny.

     

    This is a really good point, and one I keep neglecting.  I do like to keep ZBrush and Substance running in the background, and while I don't use Maya I should probably consider my usage "potential", rather than current documented needs.  ALSO... YOU WIN!  My question was who had a single render take the most VRAM, and I think I can just go ahead and call this one...  22 Million is... a lot.

     

    Serious question - if you have any more available, hit me up.

    Hey!  I'm sorry, I absolutely would if I could, but these are all being auctioned off internally within my company to recoup the loss of the Business Unit.  I mentioned we sold them, but I didn't mention tha we sold them at a pretty hefty loss, so this is just an extreme measure to try and balance that out a bit.  They were doing some sort of autonomous car... thing... in cooperation with Jaguar, so there's even an F-Type up for grabs, but I suppose I'll just settle for this card.  I'm sorry, I did ask about letting external people bid, and they were pretty explicit about the policy.  I only have one set aside because of my role in IT, and I'm the one physically parting all this stuff out.

     

    droidy001 said:
    Me too. I'm the last person to call anyone for OTT extravagant purchases. Over the last 3 years I've spent more than the cost of your quadro on a certain musical. Regular trips from the North of England to London, 2 trips to Toronto and one to New York. If it works for you and your happy that's all that matters.

    Ha!  That's awesome!  I completely agree... life is way too short and often mundane to not indulge in the things that make you happy.  If you can balance that side of you along with the responsible side that takes care of your family and gives unto others, well then I'd say you pretty much have that sweet spot dialed right in :)

     

    Alright, guys, I'm gonna do it!  Thank you so much for your honest responses!  I promise, when I get it installed I'll do my best to fill it up and render out a scene with as many things as it takes and then I'll share it here so we can all balk at the absurdity!

  • duckbomb said:
     

    Hey!  I'm sorry, I absolutely would if I could, but these are all being auctioned off internally within my company to recoup the loss of the Business Unit.  I mentioned we sold them, but I didn't mention tha we sold them at a pretty hefty loss, so this is just an extreme measure to try and balance that out a bit.  They were doing some sort of autonomous car... thing... in cooperation with Jaguar, so there's even an F-Type up for grabs, but I suppose I'll just settle for this card.  I'm sorry, I did ask about letting external people bid, and they were pretty explicit about the policy.  I only have one set aside because of my role in IT, and I'm the one physically parting all this stuff out.

    No worries, if they fail to sell or dont meet any reserve price, the offer is still there :)

  • HavosHavos Posts: 5,576
    edited November 2019
    duckbomb said:

    Wow everybody thank you so much for your responses!  Honestly, I have to say I was expecting a sarcastic response to my question, so thank you so much for your help here!

    I suppose I have to agree with kyoto kid and just eat my Ramen Noodle soup for the next few weeks to make up for this, but considering what you've all said it looks to me like I'd regret  not just going full Montey on this thing.

     

    Havos said:

    Keep in mind if you want to use the full 48GB of VRAM, you will need a PC with 96GB or more to support it.

    In my opinion you can blow the VRAM on any card if you refuse to think at all about optimising. I think any scene, even those with many characters should be able to fit into top of the range consumer cards (ie 8-11GB), if you are willing to spend at least some time optimising your scene. The only reason to need a huge amount of VRAM is if you want to render at very large resolution, ie 4Kx4K or more.

    Do you mean I'd need 96 GB or more of regular RAM to support this?  I currently only have 64... I suppose that's something worth considering.  If I'm being honest, I have to admit I vary rarely render an optimized scene.  I tend to build up giant setpieces and pre-light them all, and then hide them in chunks to render out panels of a cohesive story.  I used to delete them outright, but I've too many times come back after a render and hit "save" without saving off a copy, first.  In addition, I do like to throw in some new mats from Substance on the fly, and since I don't really care for dialing things in at that level I tend to output everything at 8K and just apply it without optimizing.  I know it's a bad habit to work in, but I suppose I'm a victim of laziness and complacency here.  I don't suppose a Quadro is going to make that any better, you're probably right in that I should consider learning how to optimize and apply that to my workflow.  I do render in about 8K, which I know is crazy again, but it really helps with my post-processing and I use a lot of that natural compression to get rid of that plastic look.

    Well I have never rendered using a 48GB VRAM card, so I can only assume that is how much you will need. In all the testing I have done, my observation is that Daz Studio uses roughly double the RAM, as the card uses VRAM. If that is still true for one of these quadros, I honestly don't know.

     

     

    Post edited by Havos on
  • alexhcowleyalexhcowley Posts: 2,403
    edited November 2019

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    Post edited by alexhcowley on
  •  Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

  • HavosHavos Posts: 5,576

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

    From what I have read, the space being reserved for controlling the screens is not the main issue. Even a high resolution screen does not take up that much space in relative terms, ie tens of MB, which is nothing for modern graphics cards. The bigger problem is that Windows 10 stops any app from using more that 90% (might be 80%, I am not sure) of the total VRAM, assumably to stop rogue applications from hogging all the resources. On a large card like a 1080Ti, this 10% comes to more than a gig.

    If Quadros are affected by this "issue", I am not sure.

  • RayDAntRayDAnt Posts: 1,154
    Havos said:

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

    From what I have read, the space being reserved for controlling the screens is not the main issue. Even a high resolution screen does not take up that much space in relative terms, ie tens of MB, which is nothing for modern graphics cards. The bigger problem is that Windows 10 stops any app from using more that 90% (might be 80%, I am not sure) of the total VRAM, assumably to stop rogue applications from hogging all the resources. On a large card like a 1080Ti, this 10% comes to more than a gig.

    If Quadros are affected by this "issue", I am not sure.

    VRAM usage in relation to display driving in Windows 10 is in the neighborhood of HUNDREDS of megabytes per display in the case of HD resolutions - not tens (eg. dropping a native 4K Display to 1080p will instantly shave 300-400MB off your OS reserved total.)

  • Havos said:

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

    From what I have read, the space being reserved for controlling the screens is not the main issue. Even a high resolution screen does not take up that much space in relative terms, ie tens of MB, which is nothing for modern graphics cards. The bigger problem is that Windows 10 stops any app from using more that 90% (might be 80%, I am not sure) of the total VRAM, assumably to stop rogue applications from hogging all the resources. On a large card like a 1080Ti, this 10% comes to more than a gig.

    If Quadros are affected by this "issue", I am not sure.

    It's a video buffer not strictly a frame buffer. From what I've been able to figure out Windows assumes you might want to start playing a game with on the fly rendered animations, or the like, while still watching a 60fps video.

  • HavosHavos Posts: 5,576
    Havos said:

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

    From what I have read, the space being reserved for controlling the screens is not the main issue. Even a high resolution screen does not take up that much space in relative terms, ie tens of MB, which is nothing for modern graphics cards. The bigger problem is that Windows 10 stops any app from using more that 90% (might be 80%, I am not sure) of the total VRAM, assumably to stop rogue applications from hogging all the resources. On a large card like a 1080Ti, this 10% comes to more than a gig.

    If Quadros are affected by this "issue", I am not sure.

    It's a video buffer not strictly a frame buffer. From what I've been able to figure out Windows assumes you might want to start playing a game with on the fly rendered animations, or the like, while still watching a 60fps video.

    Well a video buffer would certainly take up a lot more space than a single frame.

    However regardless of how much space Windows grabs for the screen(s), it is still grabbing a percentage as well, as I have seen the higher the VRAM, the more that is "lost", even when the screen resolution and number of screens supported by the card is the same as cards with less VRAM.

  • QuasarQuasar Posts: 679

    I would love a card with even double the VRAM of the 1080ti. I wish Iray could use the system RAM as well. Until they get more RAM onto their more affordable gaming cards, I'll continue to optimize as best I can.

  • joseftjoseft Posts: 310
    Quasar said:

    I would love a card with even double the VRAM of the 1080ti. I wish Iray could use the system RAM as well. Until they get more RAM onto their more affordable gaming cards, I'll continue to optimize as best I can.

    Not going to happen until they significantly increase the VRAM on the quadro/professional cards. The added VRAM on the pro cards is one of a few things that enables them to charge such high prices for them and have people pay it.

    Theres really nothing else stopping them from putting that much VRAM on the gaming grade cards. 

  • RayDAntRayDAnt Posts: 1,154
    edited November 2019
    joseft said:

    The added VRAM on the pro cards is one of a few things that enables them to charge such high prices for them and have people pay it.

    The price gap between Nvidia's professional and prosumer/consumer product lines is about much more than just raw VRAM capacity. Professional grade Quadro/Tesla GPUs feature much higher cost ECC memory and much more durable, higher cost assortments of assisting circuit board level components than their Titan/GeForce counterparts.

    Post edited by RayDAnt on
  • joseftjoseft Posts: 310
    RayDAnt said:
    joseft said:

    The added VRAM on the pro cards is one of a few things that enables them to charge such high prices for them and have people pay it.

    The price gap between Nvidia's professional and prosumer/consumer product lines is about much more than just raw VRAM capacity. Professional grade Quadro/Tesla GPUs feature much higher cost ECC memory and much more durable, higher cost assortments of assisting circuit board level components than their Titan/GeForce counterparts.

    I agree, which is why i said it was one of a few reasons - not that it was the only reason :)

    My point was more that the amount of VRAM on geforce is a choice, one that is made in order to avoid having the geforce cards in closer competition to the professional cards. They could easily put more VRAM on them, it just doesnt make financial sense for them to do so.

  • outrider42outrider42 Posts: 3,679
    edited November 2019
    When you say that "just about any current generation RTX would be better than this", you would be wrong. The Titan RTX, 2080ti, and Quadro RTX 8000 all use the exact same chip. The 2080ti is very slightly cut down. The Titan and Quadro are both using the full chip, and therefore should render at mostly the same speed. I think the Titan is clocked slightly higher, so it will render slightly faster, but again, only slightly. The Quadro RTX 6000 also uses he same full chip, but it has half the VRAM, like the TITAN RTX.

    Puget did a benchmark on these. It is for Octane, not Iray, but the performance comparison should hold true for Iray. https://www.pugetsystems.com/labs/articles/OctaneBench-2019-Preview-Quadro-RTX-Performance-Boost-1387/

    Here is a benchmark with Iray from Migenius. https://www.migenius.com/products/nvidia-iray/iray-rtx-2019-1-1-benchmarks
    Post edited by outrider42 on
  • Quasar said:

    I would love a card with even double the VRAM of the 1080ti. I wish Iray could use the system RAM as well. Until they get more RAM onto their more affordable gaming cards, I'll continue to optimize as best I can.

    The next generation of cards, due next year, is rumored to have more VRAM at each level.

  • alexhcowleyalexhcowley Posts: 2,403
    RayDAnt said:
    Havos said:

     Alsoif You don't use it to drive a monitor you can put it into TCC and not rerserve a frame buffer on it.

    TCC?  Can you elaborate on this, please?  I'm  buying a 2080ti in a few months time and I'd rather Windows didn't grab any of its memory.

    Cheers,

    Alex.

    First only pro cards and Titans can do this so a 2080ti will always reserve a frame buffer in Windows.

    TCC, Tesla Compuet Cluster mode, disables all video outputs on Quadros and Titans, its also always on for Tesla's because they have no video outs at all. With no active video outs Windows frees up the video buffer it would otherwise reserve on the card. It cannot be enabled at all if there is a monitor connected to the card as well and plugging one in while the card is in TCC crashes the computer. TCC also does some other things but that is what matters in this regard.

    From what I have read, the space being reserved for controlling the screens is not the main issue. Even a high resolution screen does not take up that much space in relative terms, ie tens of MB, which is nothing for modern graphics cards. The bigger problem is that Windows 10 stops any app from using more that 90% (might be 80%, I am not sure) of the total VRAM, assumably to stop rogue applications from hogging all the resources. On a large card like a 1080Ti, this 10% comes to more than a gig.

    If Quadros are affected by this "issue", I am not sure.

    VRAM usage in relation to display driving in Windows 10 is in the neighborhood of HUNDREDS of megabytes per display in the case of HD resolutions - not tens (eg. dropping a native 4K Display to 1080p will instantly shave 300-400MB off your OS reserved total.)

    Thanks for all the comments, guys.

    Cheers,

    Alex.

  • nicsttnicstt Posts: 11,715

    Well, I had 4 figures in a scene, 4 dynamic hair and Studio was using 54GB on CPU. Think that is the most RAM I've seen used.

    If you're thinking about a card like that, I'm tempted to say, Studio is not the software for you. :) But, why not, Iray seems to me to be a resource hog.

Sign In or Register to comment.