Iray Starter Scene: Post Your Benchmarks!

1373840424349

Comments

  • ebergerlyebergerly Posts: 3,255
    kameneko said:

     

    ebergerly said:

    But RayDAnt, Iray/Studio hasn't implemented the RTX stuff yet, correct? How are present metrics meaningful, other than as an interim set of data that will change in a few months? 

    Who will implement it? Nvidia or Daz developers themselves?

    Again, it's complicated... laugh

    Mostly though it's a bunch of groups in NVIDIA, developing a bunch of different software components. And it's another group in NVIDIA who develop Iray, and they need to integrate with all the RTX-related software the other groups are developing. That's a HUGE task. And then once they produce that the DAZ folks need to modify their user interface and some underlying software so that the user can access all this cool new stuff. 

  • outrider42outrider42 Posts: 3,679
    ebergerly said:
    kameneko said:

     

    ebergerly said:

    But RayDAnt, Iray/Studio hasn't implemented the RTX stuff yet, correct? How are present metrics meaningful, other than as an interim set of data that will change in a few months? 

    Who will implement it? Nvidia or Daz developers themselves?

    Again, it's complicated... laugh

    Mostly though it's a bunch of groups in NVIDIA, developing a bunch of different software components. And it's another group in NVIDIA who develop Iray, and they need to integrate with all the RTX-related software the other groups are developing. That's a HUGE task. And then once they produce that the DAZ folks need to modify their user interface and some underlying software so that the user can access all this cool new stuff. 

    There is nothing complicated about it. Daz does not develop Iray, so they must depend on Nvidia to supply them with any updates to Iray.

  • bluejauntebluejaunte Posts: 1,859

    How do you define what the average user renders? That is both impossible and illogical.

    Not really IMO. Look at the gallery. The vast majority of renders have an environment with a character in it.

    Now for a little speculation: Daz is "due" for a new Genesis figure this year if they hold to their 2 year cycle every past Genesis has had. That means a Summer release for Genesis 9.

    Daz have confirmed a few days ago in another thread that there will be no G9 this year.

  • LenioTGLenioTG Posts: 2,116
    ebergerly said:
    kameneko said:

     

    ebergerly said:

    But RayDAnt, Iray/Studio hasn't implemented the RTX stuff yet, correct? How are present metrics meaningful, other than as an interim set of data that will change in a few months? 

    Who will implement it? Nvidia or Daz developers themselves?

    Again, it's complicated... laugh

    Mostly though it's a bunch of groups in NVIDIA, developing a bunch of different software components. And it's another group in NVIDIA who develop Iray, and they need to integrate with all the RTX-related software the other groups are developing. That's a HUGE task. And then once they produce that the DAZ folks need to modify their user interface and some underlying software so that the user can access all this cool new stuff. 

    Uh...so how many months do you think will pass until then?

    I should swap my GPU this summer if everythings goes well! :)

  • outrider42outrider42 Posts: 3,679

    Now for a little speculation: Daz is "due" for a new Genesis figure this year if they hold to their 2 year cycle every past Genesis has had. That means a Summer release for Genesis 9.

    Daz have confirmed a few days ago in another thread that there will be no G9 this year.

    Really? I am in a state of shock right now, LOL. Seriously. And relieved to be honest. I was very much against the 2 year cycle to begin with. Actually this news makes me very happy! That's another topic, though. (Genesis 9, not my happiness....)

    Ok, so scratch that then. But RTX is coming to Daz Studio in 2019 according to Nvidia. The weird thing is that 4.11 is STILL in beta itself, so what will they do? Will 4.11 get a full release and 4.12 (or whatever) release as beta? It feels like the 4.11 beta has been out a really long time.

  • ebergerlyebergerly Posts: 3,255
    edited April 2019
    kameneko said:

     

    Uh...so how many months do you think will pass until then?

    I should swap my GPU this summer if everythings goes well! :)

    My hunch is maybe 6 months until all this gets ironed out. If you look at the link I posted on RTX, there are an incredible number of what are called SDK's and API's that need to be developed and updated for tons of RTX-related software. And they all have to work together, and they all have to work with the latest CUDA and GPU drivers and not explode with existing GPU's. And there are different versions in different stages of development, and different revisions coming out that have to be integrated. And all of that needs to be integrated into Iray. And then DAZ (or any other rendering software company) has to integrate that Iray software into their rendering UI, and that takes time and resources. It's not like you can just get something from NVIDIA, load it into Studio, and update everyone's version. It's far more complicated, and it also has to be integrated with existing planned work and prioritized. 

    BTW, I also wouldn't be surprised if different portions of the RTX technologies get implemented in Studio in different stages. Which could push this out even longer until it all gets implemented. So they may say "we will get RTX in 2019", but that might mean only a portion gets actually implemented. That's totally normal for software development. I'm also guessing that some RTX features/functionality will never get implemented in Studio. 

    Post edited by ebergerly on
  • Takeo.KenseiTakeo.Kensei Posts: 1,303
    edited March 2019

    Not all usage cases are going to match yours, Takeo.  You are talking about an ideal rendering config, which is all fine and good, but a lot of people use their computers to do other things as well.  Not everyone is going to need HEDT, or a Titan RTX for that matter, and there is no such thing as 'one size fits all'.

    I don't know where you get the "one size fits them all" idea in my posts or where you saw me post about my usage. Nor did I talk about ideal rendering config. The only thing I did is put a perspective from a price/performance standpoint of a 2990W vs RTX Titan in the restricted field of DS Iray render when dealing with heavy assets that don't fit a gtx 1080 ti

     

    Don't overcomplicate things. It's not that hard. Create a scene that represents what the average user renders every day. Probably an environment with one or more characters in it, some HDRI light and some spot lights. There is no reason to create a benchmark that tests specific features of the hardware. It's irrelevant if those features are unable to influence the render times of an average real world render. Artificial benchmarks serve no purpose.

    Actually, I think there is no need for a new benchmark for the moment. There is sufficient data to differentiate the RTX line and compare to Pascal

    If real time raytracing ever comes to DS the question then will be if there is a need for a new benchmark

    The point where we diverge is that better built benchmark (understand focused, not artificial) would reduce variance and would be more precise. That would have prevented many people to conclude that Optix Prime acceleration doesn't bring anything to RTX cards for example

     

    ebergerly said:

    Ah, right. They are calling the whole platform RTX. I was thinking just RT cores when we say RTX.

    Yeah, unfortunately that's one of those very common tech enthusiast myths resulting from a lot of people watching youtube videos that don't come close to really explaining the complicated parts of various tech subjects. So they grab on to something simple like "tensor cores" thinking that's all there is (and not really knowing what they are). Unfortunately, "complicated" doesn't sell in the youtube world. 

    But yeah, the RTX cards and related software have potential to do a lot of cool stuff, but unfortunately you never hear about a lot of it because nobody has posted attention-grabbing videos about them laugh

    Other features are for gaming or machine learning and have little impact on Iray render. That's why

    ebergerly said:

    BTW, I just did a quick check and it looks like someone posted a $2,500 Titan RTX render time that gives it an insanely overpriced Price/Performance ratio of twice what you'd expect: instead of the 15 range it's more like 32. I assume this will improve once RTX gets fully implemented.

    And since the GTX 1080ti has been priced out of the market, it has a Price/Performance of 23. 

    So yeah, if the RTX cards get their price/performance down in the 15 range or better that's when I'll get interested. Although it looks like the RTX 2060 has a ratio of 8 unless my calculator is broken. Sweet.  

    I don't know why you still say RTX is not ready for prime

    From a strict performance/price stand point the RTX card are already better than Pascal even without raytrace tech

    From my POV, for a fixed price you get superior performance but less VRAM. Nvidia is not making you pay more for performance. They make you pay to get more VRAM. The lineup should be seen as follow

    2060 > 1070 minus 2GB VRAM

    2070 > 1080

    2080 > 1080 ti minus 3GB VRAM

    2080 Ti > Titan Pascal minus 1GB VRAM

    RTX Titan is a new class of it's own.

    People who will buy the Titan will do it because of it's 24GB VRAM. Speaking of perf/$ for this one has no meaning unless you compare to quadro

     

    Post edited by Takeo.Kensei on
  • outrider42outrider42 Posts: 3,679
    edited April 2019
    kameneko said:
    ebergerly said:
    kameneko said:

     

    ebergerly said:

    But RayDAnt, Iray/Studio hasn't implemented the RTX stuff yet, correct? How are present metrics meaningful, other than as an interim set of data that will change in a few months? 

    Who will implement it? Nvidia or Daz developers themselves?

    Again, it's complicated... laugh

    Mostly though it's a bunch of groups in NVIDIA, developing a bunch of different software components. And it's another group in NVIDIA who develop Iray, and they need to integrate with all the RTX-related software the other groups are developing. That's a HUGE task. And then once they produce that the DAZ folks need to modify their user interface and some underlying software so that the user can access all this cool new stuff. 

    Uh...so how many months do you think will pass until then?

    I should swap my GPU this summer if everythings goes well! :)

    You really have to make that decision yourself, because only you can decide what you want more. My advice is to always get the absolute best GPU you afford, but that is my advice.

    If it takes a few months for RTX to come to Daz, then you have extra time to save up for a better GPU. But again, this all depends on how much you value your time spent rendering, and how large your scenes are. Plus...normally GPU prices come down over time. Or a sale happens. Right NOW there is a sale at OverClockersUK on all RTX cards if you live in the UK.

    https://wccftech.com/nvidia-geforce-rtx-20-series-price-drop-rtx-2080-ti-rtx-2070-rtx-2060/

     

    ebergerly said:
    kameneko said:

     

    Uh...so how many months do you think will pass until then?

    I should swap my GPU this summer if everythings goes well! :)

    My hunch is maybe 6 months until all this gets ironed out. If you look at the link I posted on RTX, there are an incredible number of what are called SDK's and API's that need to be developed and updated for tons of RTX-related software. And they all have to work together, and they all have to work with the latest CUDA and GPU drivers and not explode with existing GPU's. And there are different versions in different stages of development, and different revisions coming out that have to be integrated. And DAZ (or any other rendering software company) has to integrate that software into their rendering UI, and that takes time and resources. It's not like you can just get something from NVIDIA, load it into Studio, and update everyone's version. It's far more complicated, and it also has to be integrated with existing planned work and prioritized. 

    BTW, I also wouldn't be surprised if different portions of the RTX technologies get implemented in Studio in different stages. Which could push this out even longer until it all gets implemented. So they may say "we will get RTX in 2019", but that might mean only a portion gets actually implemented. That's totally normal for software development. 

    OptiX Prime is the main driver for Iray's ray tracing speed. OptiX (non Prime) already released 6.0 which has RTX. One would assume that the people working on Iray would have had access to OptiX 6.0 long before it released to the public. This may be speculation, but it only makes sense for Iray to get the full OptiX plugin, and the people at OptiX have stated that Prime will NOT be updated for RTX. Unless of course, we are getting an all new render plugin.

    Post edited by outrider42 on
  • ebergerlyebergerly Posts: 3,255

    Optix is one small part of the RTX technologies, which include:

    • Ray Tracing (OptiX, Microsoft DXR, Vulkan)
    • AI-Accelerated Features (NGX)
    • Rasterization (Advanced Shaders)
    • Simulation (CUDA 10, PhysX, Flex)
    • Asset Interchange Formats (USD, MDL)

    Developing all of this and getting it to work together and getting it implemented in Iray and Studio is a HUGE task. I'm sure most software developers would look at this and just shake their heads. I'm sure this has been in the works for many years before we even heard about it, and still it's in the development phase. 

  • bluejauntebluejaunte Posts: 1,859
    ebergerly said:

    Optix is one small part of the RTX technologies, which include:

    • Ray Tracing (OptiX, Microsoft DXR, Vulkan)
    • AI-Accelerated Features (NGX)
    • Rasterization (Advanced Shaders)
    • Simulation (CUDA 10, PhysX, Flex)
    • Asset Interchange Formats (USD, MDL)

    Developing all of this and getting it to work together and getting it implemented in Iray and Studio is a HUGE task. I'm sure most software developers would look at this and just shake their heads. I'm sure this has been in the works for many years before we even heard about it, and still it's in the development phase. 

    Don't think most of that stuff has to be in Iray. Rasterization again sounds like stuff for game engines. Tensor cores are already working for denoising according to this:

    https://www.irayplugins.com/forum/showthread.php?27969-When-is-iray-of-a-new-version-for-RTX

    Still, 6 months sounds fairly reasonable. Later this year is all we really know.

  • bluejauntebluejaunte Posts: 1,859
    edited April 2019

    Now for a little speculation: Daz is "due" for a new Genesis figure this year if they hold to their 2 year cycle every past Genesis has had. That means a Summer release for Genesis 9.

    Daz have confirmed a few days ago in another thread that there will be no G9 this year.

    Really? I am in a state of shock right now, LOL. Seriously. And relieved to be honest. I was very much against the 2 year cycle to begin with. Actually this news makes me very happy! That's another topic, though. (Genesis 9, not my happiness....)

    Yeah I think it's good news too. They should make this more official than buried somewhere in a forum thread though smiley

    Thread here, scroll down a bit: https://www.daz3d.com/forums/discussion/266736/genesis-9-coming-next-year-and-if-so-will-you-buy-it/p12

    Post edited by bluejaunte on
  • outrider42outrider42 Posts: 3,679
    edited April 2019

    Which of those does Iray actually use, though? You are basically looking at CUDA and OptiX. CUDA is already working with Turing, and in fact, it is working beautifully. Iray has NO physics simulation. Daz Studio uses dforce for that, which is a separate plugin, and dforce does not use PhysX.

    ebergerly said:

    Optix is one small part of the RTX technologies, which include:

    • Ray Tracing (OptiX, Microsoft DXR, Vulkan)
    • AI-Accelerated Features (NGX)
    • Rasterization (Advanced Shaders)
    • Simulation (CUDA 10, PhysX, Flex)
    • Asset Interchange Formats (USD, MDL)

    Developing all of this and getting it to work together and getting it implemented in Iray and Studio is a HUGE task. I'm sure most software developers would look at this and just shake their heads. I'm sure this has been in the works for many years before we even heard about it, and still it's in the development phase. 

    Don't think most of that stuff has to be in Iray. Rasterization again sounds like stuff for game engines. Tensor cores are already working for denoising according to this:

    https://www.irayplugins.com/forum/showthread.php?27969-When-is-iray-of-a-new-version-for-RTX

    Still, 6 months sounds fairly reasonable. Later this year is all we really know.

    Of note is that the Daz Studio Iray may differ from other implementations of Iray. There are some things not available in Daz's Iray, like motion blur. But Iray is capable of motion blur.

    Since you have a 2080ti now, have you tried the denoiser with it? Can you see any difference with it, or is it noticeably faster than the 1080ti's were? This may be hard to test, because the 2080ti is so much faster as it is, it may be hard to tell if the denoiser is using Tensor cores or not. Do any GPU monitors track Tensor use?

    **Trying to edit my post to avoid double posting, so this may look weird.

    bluejaunte said:

    Now for a little speculation: Daz is "due" for a new Genesis figure this year if they hold to their 2 year cycle every past Genesis has had. That means a Summer release for Genesis 9.

    Daz have confirmed a few days ago in another thread that there will be no G9 this year.

    Really? I am in a state of shock right now, LOL. Seriously. And relieved to be honest. I was very much against the 2 year cycle to begin with. Actually this news makes me very happy! That's another topic, though. (Genesis 9, not my happiness....)

    Yeah I think it's good news too. They should make this more official than buried somewhere in a forum thread though smiley

    Thread here, scroll down a bit: https://www.daz3d.com/forums/discussion/266736/genesis-9-coming-next-year-and-if-so-will-you-buy-it/p12

    Thanks for that link! I spy something very interesting in that post:

    "We do however, have a lot of other cool stuff that will be releasing in Q2-Q4.  Watch, you'll start to see a lot of new functionality in content and Studio itself over the next several months that will let you do a lot of things with Genesis 8 that you haven't been able to.  "

    Note how he says Q2-Q4 for "cool stuff". Its odd he wont talk about RTX, but this statement could be referring to that. It looks like a new version of DS is coming for sure. Of course, Q2-Q4 is really open ended. Quarter 2 starts tomorrow! So it could be almost anytime. I think 3-4 months sounds logical.

    Post edited by outrider42 on
  • bluejauntebluejaunte Posts: 1,859

     

    Since you have a 2080ti now, have you tried the denoiser with it? Can you see any difference with it, or is it noticeably faster than the 1080ti's were? This may be hard to test, because the 2080ti is so much faster as it is, it may be hard to tell if the denoiser is using Tensor cores or not. Do any GPU monitors track Tensor use?

    Yeah, but frankly I don't even know what's supposed to be faster. Denoising seems like an instant process to me. You tell it at what iteration it should kick in, and it does. There's no wait of any kind that I can see. And it still looks crappy to me, washing out details like lashes for example. It just kinda blurs everything, except maybe outlines. For the type of stuff that I do, which is endlessly test-rendering characters I'm working on, it's pretty much unuable. Even if I let it kick at high iteration, like almost where I would normally stop the render anyway, it still blurs out stuff. Not a fan.

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019

    Benchmarking the Benchmarks Part 1

    Author's note: Unless otherwise stated, all numbers/statistics in the tables/graphs below were either taken directly from Daz Studio log files spanning multiple test runs or read from one of Daz Studio's numerous configuration options menus.

     

    Introduction

    The following is a no-frills summary of descriptive data (with occasional notes/explanations of things) on the existing DS/Iray benchmarking scenes assembled by me as a preliminary step for creating a new RTX-friendly benchmarking scene (coming soon...) Feel free to question/comment away on the info/conclusions it contains. Also, expect to see a part 2 continuation of this post soon, featuring Iteration Time Graphs for all the scenes studied (as well as explanations of what those even are, what they are useful for, and how they can be constructed by you from just the data in your log files.)

     

    Descriptives

    1. Metadata

    File Name: Self-explanatory.

    Title: Self-explanatory.

    Source: Self-explanatory.

    Creator: Self-explanatory.

    Creation Date: Creation (technically last modification) dates manually harvested from inside each publicly shared .duf file. Of particular note here is the date for SickleYield's original benchmarking scene (March 11th, 2015) emphasizing its potential for outdatedness given the appearance of many newer hardware and software iterations in the 4+ years since.

     

    2. Scene Content

    Description: Self-explanatory. Take special note of the SickleYield's use of a G2F character (more evidence of its relative age) and the total lack of figures in either variations of the Aala (due to it being a Daz Studio non-specific cornell box by design), as well as the variety of environmental scene lighting configurations.

    Nodes: Each and every logical object present in each scene further broken down by type (as listed under the Node tab in the Scene pane.) Keep in mind that Lights does not include light emissing objects or environmental scene lighting configurations.

    Memory Consumption: Take special note of how radically different Texture memory consumption is when switiching between the current production and beta releases of Daz Studio using the same scene. I have seen someone mention the theory that Iray in the current beta utilizes some form of data compression when processing/transferring scene assets to a GPU that the current production version lacks, and these results would seem to back this up. Also take note of the lack of Texture memory consumption on both Aala variations - this is again by virtue of it being a cornell box.

     

    3. Render Settings

    General: Self-explanatory.

    Render Mode: Not presently relevant.

    Progressive Rendering: Self-explanatory. Take special note of SickleYield, JamesJAB (all variations), outrider42 and to a lesser extent Aala (both variations) here, since the way they are configured (with multiple active render stop conditions) can lead to unexpected benchmarking results on differing hardware/software cnfigurations.  

    Alpha: Not presently relevant.

    Optimization: Self-explanatory.

    Filtering: Self-explanatory.

    Spectral Rendering: Not presently relevant.

    Tone Mapping: Not presently relevant.

    Environment: Self-explanatory. Take special note of how these settings relate to each scenes' Description and Node breakdown already covered earlier.

     

    4. Render Results

    Best TRT: The shortest Total Rendering Time reported by each listed hardware/software configuration out of several runs. Take special note of the lack of times for all but the first JamesJAB variation in both GTX 1050 2GB configurations (for obvious reasons...) Also take note of the lack of times for Aala 2 min (since it is - by design - a time limited test.) 

    OptiX Performance Boost: Based on multiple runs. Whether rendering each scene with OptiX Prime enabled lead to a better (shorter) Total Rendering Time than without it for that partcular hardware/software configuration. Take special note of the "No" answers for all the post 2017 scenes with the Titan RTX configuration. 

    Peak VRAM Use: Based on multiple runs. The maximum amount of video memory (as reported by GPU-Z) used by the system while rendering with OptiX Prime enabled vs. disabled for each scene. Take special note of the extremely inconsistent and sometimes even negative values for OptiX Cost (how much additional video memory OptiX Prime required) for both GTX 1050 2GB configurations. This is because of the GTX 1050 2GB's limited memory capacity triggering Iray's built-in memory balancing algorithms in order to avoid CPU fallback (something which never happened during any test run.) Also take note of the uniform postive values of OptiX Cost for the Titan RTX configuration where limited memory capacity was most definitely not an issue. These results not only back up the claim that OptiX Prime leads to greater VRAM usage in a GPU rendering scenario, but they also indicate the exact amount varies scen-by-scene.

    Stop Trigger: The event named by Iray in the log file that resulted in the scene "finishing" being rendered. Take special note of how the explanation for SickleYield when going between hardware/software configurations isn't always the same. This is a side effect of having multiple stop conditions active in a scenes' Progressive Rendering configuration. And forms a potentially fatal flaw in a scene's usefulness as a performance benchmark.

    Post edited by RayDAnt on
  • RayDAntRayDAnt Posts: 1,120
    ebergerly said:

    I'm thinking what might be most helpful in this benchmark thread is a list of standard config stuff to make sure everyone is comparing apples to apples. Such as making sure you specify what scene you're rendering (now that there are multiple scenes out there), where you're reading your render time numbers, whether you're doing an intial render or the second render (I recall the second one is significantly faster than the 1st since the scene is already loaded or something like that...), making sure you have the correct render settings, etc. 

    I already have all the pertinent details figured out and drafted into the first post of a new benchmarking thread. All I have left to do at this point finalize the design of the RTX-friendly benchmarking scene I'm already mostly done with (something I'm hesitant to release into the wild until after the RTCore component of RTX fully makes its way into the Das Studio/Irau ecosystem for the very same reasons you've already mentioned.)

    Introducing yet another benchmarking scene into the mix isn't something I take lightly. However there really isn't any other effective way to up the ante on useful benchmarking data for prospective hardware/software users in the future without introducing something identifiably new.

     

    ebergerly said:
    RayDAnt said:
    ebergerly said:

    But RayDAnt, Iray/Studio hasn't implemented the RTX stuff yet, correct? How are present metrics meaningful, other than as an interim set of data that will change in a few months? 

    Because in order to know best where you're going you've gotta understand first where you've been.

    ETA: Also, to be clear, I'm mostly talking about the questions people have been bringing up in this thread as of the last 20 posts or so - not just the last 5 (ie. not RTX related.)

    RayDAnt, one thing you might consider if you haven't already is including something like I did in a spreadsheet I made last year, where I included Price/Performance ratios to see what GPU's give you the most bang for the buck (copy attached).

    Oh, I've definitely considered it. The problem with the whole Cost Per Frame (or, in this case, Cost Per Iteration) way of evaluating device performance is that it fails to take into account power consumption (the price you keep paying for a particular component every day you use it after you buy it.) And without some way of numerically accounting for that, the whole statistic is kind of disengenous. Plus it makes people prone to drawing less than fully comprehensive value comparisons between distinctly different classes of hardware like 2080ti's vs Titan RTX's (the former of which is a limo; the latter a stretch limo - they both get you and your closest friends to your destination in style. The strech just does it for TWICE AS MANY of your friends...)

     

  • RayDAntRayDAnt Posts: 1,120
    ebergerly said:

    It's not complicated for me. I don't need to know technical details, I need to know how fast stuff is and that's it. I already know that a 2080 TI renders roughly twice as fast as a 1080 TI and costs roughly twice as much.

    Well, actually they cost about the same, $1,200. 

    And BTW, I'm impressed that you found an RTX 2080ti render time in this thread. Honestly it's gotten so complicated with all the new benchmark scenes and extraneous discussions I've given up trying to add stuff to my summary spreadsheet. In fact I'm very impressed that RayDAnt has taken on the task of summarizing all of this. laugh

    I used to work in a research lab where I did stuff like write grant proposals, design computer-based research tools (college professor speak for benchmarks), collect/clean/summarize data, and then write studies based on it. So perusing threads like this one is just a chance to relive my glory days... I suppose...

  • outrider42outrider42 Posts: 3,679

    That's a pretty comprehensive breakdown there. I would like to point out that while my scene has no photometric lights, it has 3 light sources as 3 of the spheres producing light. Maybe 4 if you count the unseen HDRI. I didn't bother turning the dome off as I didn't see a need. And though I left 95% convergence as a stop condition, I am very confident nobody will ever hit that before hitting 5,000 iterations. When I tested maxxing out the iteration cap, my scene took 13,871 iterations to reach 95% convergence, which took 15 minutes (and 0.83 seconds) for two 1080ti's.

    There are 6 lights in the SY scene? I didn't know that.

    Out of curiosity, have the Cornell Boxes been uploaded here? I'd like to throw my 1080ti's at them and see how they match up vs your Titan RTX, and if the difference between them is statistically similar to the difference between the other benches. It doesn't have to be in this thread, to prevent confusion.

    Oh, and what driver are you on? Some drivers might impact performance. Recently people were complaining about losing performance from a certain driver, which was mostly fixed, I think, with another driver.

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019

     

    Out of curiosity, have the Cornell Boxes been uploaded here? I'd like to throw my 1080ti's at them and see how they match up vs your Titan RTX, and if the difference between them is statistically similar to the difference between the other benches. It doesn't have to be in this thread, to prevent confusion.

    Considering they originated as part of this thread, I wouldn't worry about it.

    Oh, and what driver are you on? Some drivers might impact performance. Recently people were complaining about losing performance from a certain driver, which was mostly fixed, I think, with another driver.

    The TRT times in the spreadsheets I posted are a bit outdated (I orginally put this all together 1-2 months ago - the only recent info there are the stats about vram consumption regarding OptiX which I only finished compiling today.)  Give me about 20 minutes and I'll update that graphic with current driver numbers for the Ttian RTX.

     

    EDIT: Updated the Render Results TRT times with more recent driver results and added driver versions to descriptions.

    Post edited by RayDAnt on
  • outrider42outrider42 Posts: 3,679
    I remember seeing them, but I didn't remember if they were posted. I took a long time out from Daz Studio and skimmed through a lot of this thread when I got back. I'm still taking a time out from the store. <.<
  • LenioTGLenioTG Posts: 2,116

    @RayDAnt Great job with your sheets! :D I love statistics too, but I'm bad with them xD Do you have a spreadsheet based on the GPU models as well?

    @outrider42 Yes, I'll but what I'll be able to, but it will mostly depend by my Patreon...on my own I'll barely be able to afford a 2060, or a 1660Ti xD

    @ebergerly 6 months don't look that bad I guess! As long as the main features are brought to Daz, we can live without the others! It's weird to think that our experience could change that much in just a GPU generation, but if we take a look at the best renders of some years ago we can see that there's been a huge improvement already!

  • Takeo.KenseiTakeo.Kensei Posts: 1,303
    edited April 2019

    @Raydant

    Nice effort. I don't know what your goal is with this but I think CPU and motherboard information could be a good optional information. Because there are still people who use the CPU as additionnal render ressource and we know all CPU are not equal on that matter

     

    Few remarks about your stats :

     

    RayDAnt said:
     

    Memory Consumption: Take special note of how radically different Texture memory consumption is when switiching between the current production and beta releases of Daz Studio using the same scene. I have seen someone mention the theory that Iray in the current beta utilizes some form of data compression when processing/transferring scene assets to a GPU that the current production version lacks, and these results would seem to back this up. Also take note of the lack of Texture memory consumption on both Aala variations - this is again by virtue of it being a cornell box.

    There is a Texture compression setting since forever in the render settings->advanced

    Are you sure the parameters are the same for both DS versions ?

    RayDAnt said:

    OptiX Performance Boost: Based on multiple runs. Whether rendering each scene with OptiX Prime enabled lead to a better (shorter) Total Rendering Time than without it for that partcular hardware/software configuration. Take special note of the "No" answers for all the post 2017 scenes with the Titan RTX configuration. 

    It would be better to have a percentage or time gained with Optix Prime rather than a Yes/No

    The fact that the RTX Titan doesn't benefit from Optix Prime for the scenes after 2017 whereas the GTX 1050 does, also raises few questions :

    1*/ Why ? Obvious answer would be that the raytracing performance is a bit better with the RTX Titan than with the GTX 1050

    2°/ Does it expand to all RTX cards or is it just for the quickest (need somebody who has a 2060 to test to get a beginning of an answer) ?

    I guess that the answer will be that it works for all cards and that it is a particularity of the Turing architecture but can't confirm without test

    RayDAnt said:

    Peak VRAM Use: Based on multiple runs. The maximum amount of video memory (as reported by GPU-Z) used by the system while rendering with OptiX Prime enabled vs. disabled for each scene. Take special note of the extremely inconsistent and sometimes even negative values for OptiX Cost (how much additional video memory OptiX Prime required) for both GTX 1050 2GB configurations. This is because of the GTX 1050 2GB's limited memory capacity triggering Iray's built-in memory balancing algorithms in order to avoid CPU fallback (something which never happened during any test run.) Also take note of the uniform postive values of OptiX Cost for the Titan RTX configuration where limited memory capacity was most definitely not an issue. These results not only back up the claim that OptiX Prime leads to greater VRAM usage in a GPU rendering scenario, but they also indicate the exact amount varies scen-by-scene.

    If you had different settings for Texture compression, some tests may have to be done again (sorry)

    Also, we don't really know how Optix Prime works concerning Vram usage, but is it important ? I mean, since that won't trigger a drop to the CPU, I don't see the information as meaningfull

     

    A rather interresting statistic would be the reported available Vram, and also if the card was used to plug one or many monitors and the used framebuffer if available as well.  These one would be more important to know in my POV. For me, once you get real time raytracing, the render speed will not be your primary concern. Your main consideration will be, if you have enough VRAM to make the scene as you imagine it

    Post edited by Takeo.Kensei on
  • ebergerlyebergerly Posts: 3,255
    edited April 2019

    Wow, RayDAnt, I’m very impressed with all the work you put in to make this summary. And since you asked for comments, here are mine:

    • I'm really not sure the point of the exercise. I think the average user here just wants to know “what GPU should I buy?”. And most of that comes down to “how fast will it render my scenes, and how much will it cost?” (aka, “what’s the best bang for my buck?”). It looks like your summary is comparing a $150 GTX 1050 with a $2,500 Titan RTX for 6 different benchmark scenes in 20+ different configurations (if I’m reading it correctly). So I guess I’m not really sure of the goal of your exercise. I’m not sure how many users are wondering whether to spend $150 on a 1050 or  $2,500 on a Titan RTX.
    • You mentioned that taking cost into account (“Cost per frame” or “cost per iteration”) doesn’t take into consideration power consumption costs. While I agree those two metrics are virtually meaningless for most of us, I think that whole power consumption cost issue has already been addressed, and for most users the cost of power consumption is negligible and doesn’t need to be considered. Even if you are rendering flat out for 8 hours a day, 365 days per year (which I doubt many or any of us are actually doing), the incremental cost of using a GPU that draws an additional 100 watts (which few GPU’s do) is only around $35 per year. For most of us, an unused bathroom light or running the aircon a few minutes longer will cost more than that. And for the average user who only is actually rendering for, say, 2 hours a day a few times a week, the cost is only in the $1 or $2 per year range, which is totally negligible. Or if it’s that big a concern, just add a few bucks a year for the higher wattage GPU’s in the analysis of whether a GPU is worth it. Anyway, with RTX prices being so high, I think most of us are wondering “are they worth the cost?”. So I think that a simple “sticker price/render time performance” ratio is far more appropriate for most users.
    • While I appreciate the expertise behind this analysis, I worry that the original intent of this thread is being lost, and it has become more like a massive and speculative research project rather than a simple reference for most users who are wondering how fast the different GPU’s render.  And honestly, after scanning the last 10 or 20 pages of this thread, I can’t even figure out what the posted results mean, since there are so many benchmark scenes and variables and inconsistencies. I’d update my spreadsheet with price/performance values, but I just can’t figure out what the answers are. So like I say, I’ll probably re-visit this thread in 6 months or so when (and if) Iray has implemented the RTX stuff and see where we stand. Maybe at that point we can see exactly what RTX features have been implemented in Iray and Studio, and develop an appropriate benchmark scene based on that.    
    Post edited by ebergerly on
  • outrider42outrider42 Posts: 3,679
    edited April 2019

    Prices flunctuate too much for a cost analysys to be much use. Prices also vary wildly by region. Like in Australia, prices are insane. There are many Daz users not in the US. And right now the UK has a big RTX sale going one where nearly all the RTX cards have decent discounts. Its just not practical.

    Maybe it would be a good idea to have a hardware discussion thread, or a thread for the "new" benchmarks. I aksed for a hardware discussion sub forum some time ago, but obviously that isn't happening. I still think it would be a great idea. The "Commons" is just too saturated in general with every kind of thread under sun.

    I'm doing some new benchmarks to double check. All are 4.11 beta.

    SY Scene Nvidia driver 417.35

    two 1080ti's  1 minute 2.35 seconds OptiX ON        ( 2.55 seconds faster than 1 Titan RTX  4%) 

    two 1080ti's  1 minute 29.55 seconds OptiX OFF

    There is a pretty big drop in speed just turning OptiX off on the SY scene. I don't think I ever benched it off. My best time on this scene is 58 seconds a number of months ago. I cannot recall which driver that was.

    Aala Cornell Box 1K Iterations

    two 1080ti's  1 minute 20.6 seconds  OptiX ON     (9.3 seconds faster than 1 Titan RTX   11.5%)

    two 1080ti's  1 minute 26.67 seconds  OptiX OFF

    So the SY scene benefits evens more from OptiX than a Cornell Box scene? I noticed something else, but this could be my imagination. It seemed like my GPU fans kicked in faster running the Cornel Box renders. But I was not not monitoring how fast the fans ramped up, so I can't really confirm that. Maybe another day I'll look into that.

    Aala Cornell Box Timed (I'll list each GPU's count and add them)

    two 1080ti's  2 minutes 0.97 seconds   779+791 iterations= 1570    OptiX ON         (I don't see the Titan's numbers to compare)

    two 1080ti's  2 minutes 1.83 seconds  718+723 iterations= 1441    OptiX OFF

    I see a problem with timed tests. The render does not stop at the exact time. This might only throw the data off by a tiny percentile, but it is noteworthy. That's a real bummer, because I was thinking that a timed bench would be a great idea. A timed bench would mean anybody could test it and not worry about it taking an hour if they have just a CPU. As long as the iteration count is low, since the max is 15000, we have to make sure that 15000 is not attainable, at least for a few generations of GPUs.

    Outrider42 scene

    two 1080ti's  4 minutes 47.6 seconds  OptiX ON         (18 seconds faster than 1 Titan RTX  6.3%)

    two 1080ti's  5 minutes 16.59 seconds  OptiX OFF

    ---------------------------------------------------------------------------------------------------------------------------------------

    Now look at the percentage differences. The SY scene has the smallest gap here, with only a 4% difference. The Cornell Box, as expected, shows the largest gap with a 11.5% difference. My scene cuts right down the middle with a 6.3% difference. These are all based on best times, which while the 1080ti(x2) achieved better performance every time with OptiX on, this is not always true for the RTX. That fact is noteworthy. Because my OptiX OFF times are slower than the Titan RTX OFF times in every case.

    In effect, you can see the differences right here. And this is what I was getting at before. Iray does not scale perfectly from scene to scene, it truly does matter what conent the scene has. The differences might not be huge, but going from 4% to 11.5% can mean a lot over time. The SY scene is making cards look much closer together in performance than they really are. I believe this demonstrates the best reason yet as to why multiple benchmarks are a logical step. Even though RTX ray tracing cores are not working yet, we can still have these benchmarks in place. And I expect these gaps to get bigger once RTX is enabled, after all, OptiX ON is SLOWER for several scenes on RTX. This could change completely once RTX gets enabled.

    Post edited by outrider42 on
  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    kameneko said:

    @RayDAnt Great job with your sheets! :D I love statistics too, but I'm bad with them xD Do you have a spreadsheet based on the GPU models as well?

    No (at least, not yet.) I do have a 800+ item spreadsheet of ALL the benchmark results so far posted in this thread. But I'm sorry to say that - with the exception of approximately 8 of them - the numbers they give are essentially useless due to a lack of system configuration details directly effecting performance (really basic things like what Windows/Mac OS version or even which specific release of Daz Studio/Iray the scene was rendered under) that are really really important to know when dealing with a dataset that's been steadily growing for 4+ years now.

     

    Nice effort. I don't know what your goal is with this

    This is all part of research I'm doing into what kinds of scene configurations lead to the most consistent rendering results since I am in the process of creating a new RTX-oriented (but still previous gen GPU/high performance CPU friendly) Daz Studio Iray benchmarking scene with much more comprehensive reporting instructions (to be released soon on this forum - in its own thread to avoid further confusion.)

     

    Few remarks about your stats :

     

    RayDAnt said:
     

    Memory Consumption: Take special note of how radically different Texture memory consumption is when switiching between the current production and beta releases of Daz Studio using the same scene. I have seen someone mention the theory that Iray in the current beta utilizes some form of data compression when processing/transferring scene assets to a GPU that the current production version lacks, and these results would seem to back this up. Also take note of the lack of Texture memory consumption on both Aala variations - this is again by virtue of it being a cornell box.

    There is a Texture compression setting since forever in the render settings->advanced

    Are you sure the parameters are the same for both DS versions ?

    Yes, because all tests were purposefully done with all parameter settings in each version of Daz Studio tested left at their defaults. And both 4.10.0.123 and 4.11.0.236 default to the same values for both Texture Compression thresholds (512 and 1024 for Medium and High respectively.)

     

    RayDAnt said:

    OptiX Performance Boost: Based on multiple runs. Whether rendering each scene with OptiX Prime enabled lead to a better (shorter) Total Rendering Time than without it for that partcular hardware/software configuration. Take special note of the "No" answers for all the post 2017 scenes with the Titan RTX configuration. 

    The fact that the RTX Titan doesn't benefit from Optix Prime for the scenes after 2017 whereas the GTX 1050 does, also raises few questions :

    1*/ Why ? Obvious answer would be that the raytracing performance is a bit better with the RTX Titan than with the GTX 1050

    2°/ Does it expand to all RTX cards or is it just for the quickest (need somebody who has a 2060 to test to get a beginning of an answer) ?

    I guess that the answer will be that it works for all cards and that it is a particularity of the Turing architecture but can't confirm without test

    Fwiw the answer to this question is already evident.

     

    RayDAnt said:

    Also, we don't really know how Optix Prime works concerning Vram usage, but is it important ? I mean, since that won't trigger a drop to the CPU, I don't see the information as meaningfull

    If 20 (or 200) MB more memory usage is what it takes for a scene to trigger CPU fallback on your system, then it is very important. And the reality is that low to mid capacity VRAM GPUs are far more common than high capacity behemoths like the Titan RTX.

     

     

    ebergerly said:

    Wow, RayDAnt, I’m very impressed with all the work you put in to make this summary. And since you asked for comments, here are mine:

    • I'm really not sure the point of the exercise.

    The purpose behind this particular effort was to collect background information on the setup and execution patterns of as many existing Daz Studio/Iray benchmarking scenes as possible on as many computing systems with diverse rendering capabilities as possible, in order to better understand what makes for a well-designed scene from a benchmarking perspective. Nothing more, nothing less. 

     

    ebergerly said:
    • I think the average user here just wants to know “what GPU should I buy?”

    That's what I think too. Beleive it or not, this is actually all just pre-requisite information gathering for creating a brand new benchmarking thread (featuring a much more robustly designed benchmarking scene as well as thorough instructions on how to run it/report results with sufficient system configuration info to make those results meaningful to others) where the first thing in the OP is going to be a routinely updated straightforward summary of crowdsourced performance statistics on different GPU/CPU rendering devices. I'm just being very thorough about how I'm setting it up in terms of the R&D for it since I have specialized knowledge/skills in just this sort of thing.

    Plus I firmly believe in either doing something right, or not at all.

    Post edited by RayDAnt on
  • tj_1ca9500btj_1ca9500b Posts: 2,047
    edited April 2019

    ebergerly said:

    • I think the average user here just wants to know “what GPU should I buy?”

    The quickest answer to that is whichever Iray capable Nvidia card you can afford that has the most VRAM.  When I'm asked this question, I usually start with the 6GB 1060 (not the 3GB one) as a minimum, with the 8 GB 1070 above that, then the 1080, and then recommend the 1080 Ti as the preferred option if they can afford it.

    The 20xx and the newly released 16xx cards are also relevant choices.  Assuming that the 20xx cards are usable outside of the beta version of Daz Studio, yeah those are faster choices of course.  I'm guessing that Nvidia has resolved the 'higher failure rate' issues for the 2080 Ti's by now?  And if it's still a beta only thing, well that's why I'm not recommending 2080 Ti's by default as of yet.

    The main point I'm making here is that the amount of available VRAM is the primary consideration, with how fast they are being a secondary consideration.  It's pretty easy to bump up against 6 GB or 8 GB limitations (especially if you are losing 18% of that VRAM to the Windows 10 VRAM tax), and it's not that hard to exceed the 11GB of the 1080 and 2080 Ti's for that matter.  I'm stating the obvious of course, but yeah, the more VRAM you have, the less often you may run into CPU only rendering situations, where you may need to optimize a scene to squeeze it into your graphics card, or render parts of your scene in multiple passes.

    That's what makes the Titan RTX look like such an attractive value. 24GB of VRAM should cover most scene situations.  I was seriously looking at the P6000 at one point, which also has 24GB of VRAM, but it retails for nearly twice as much as the newer Titan RTX, and is slower...

    Outrider's observation of 2 1080 Ti's being slightly faster than a single Titan RTX is also a consideration, though.  For the price of that Titan RTX, yeah you should be able to afford the 2 1080 Ti's.  Two 2080 Ti's should, by extension, be signifcantly faster than that single Titan RTX.  Of course, you'll need to have two available PCIex16 slots (running at x8 or x16) for the dual card config.  So if you aren't building big scenes very often, yeah the two cheaper cards might be the better option, if your motherboard can accomodate both cards.  Or three cards, or four cards...

    The multiple cheaper cards approach is a bit more budget friendly as well, if your motherboard can accomodate them, as you can buy the additional cards as budget allows.  The Titan RTX is simply out of reach for most budgets at any given point, but grabbing a second 1080 Ti later down the road, or even a faster card (20xx or a future gen card), well then you can at least make due with the single cheaper card for now.  Again, it'll all come down to a person's available budget.

    That's incidentally why we have people combining older cards newer ones and sharing their benchmarks in this thread.  Those cheaper/older/less capable cards are probably what they started with.  Fortunately, Daz Studio allows you to combine older cards with newer ones, to a point, i.e. the 'scene needs to be able to fit in the smallest card thing.

    And THAT is what makes even the lowly 8 core Threadripper look so attractive.  The 64 (60 + 4 to chipset) PCIe lanes can comfortably accomodate 4 GPUs, assuming you pick the right motherboard with appropriate PCIe 16 slot spacing.  It'll probably be a x16/x8/x16/x8 lane setup, but x8 is perfectly fine for rendering.  You WILL need to get a suitably large power supply of course if you are looking at that route.  But that's a subject for another thread I think.

    Here's a fun rumor/leak.  It's been noted (in a leak provided to Adored TV on Youtube) that the 7nm EPYC chips may have 160 PCIe or more lanes available in 2 socket configs.  The current 7xxx EPYCs have 128 available PCIe lanes in 2 socket configs, due to the other 128 (64 per chip) being needed for the infinity fabric cross-connect between CPUs.  The 7nm UMA (not NUMA) architecture in Rome apparently needs significantly less than 128 lanes for said cross-connect.  While this won't matter to 99% of the people here (who probably can't afford a 2 socket EPYC system in the first place), well for you high end types, that could mean 8 GPUs with 16 dedicated lanes apiece, easy peasy.  Of course, you'd also need a suitable motherboard/rack system to accomodate those 8 GPUs.  The interesting question will then be, how many of those slots will be PCIe4 as opposed to PCIe 3?  PCIe4 probably won't help much with render times, but we won't really know until the benchmarks of suitable PCIe4 GPUs and systems start showing up...

    Post edited by tj_1ca9500b on
  • Takeo.KenseiTakeo.Kensei Posts: 1,303
    RayDAnt said:
     
    RayDAnt said:

    OptiX Performance Boost: Based on multiple runs. Whether rendering each scene with OptiX Prime enabled lead to a better (shorter) Total Rendering Time than without it for that partcular hardware/software configuration. Take special note of the "No" answers for all the post 2017 scenes with the Titan RTX configuration. 

    The fact that the RTX Titan doesn't benefit from Optix Prime for the scenes after 2017 whereas the GTX 1050 does, also raises few questions :

    1*/ Why ? Obvious answer would be that the raytracing performance is a bit better with the RTX Titan than with the GTX 1050

    2°/ Does it expand to all RTX cards or is it just for the quickest (need somebody who has a 2060 to test to get a beginning of an answer) ?

    I guess that the answer will be that it works for all cards and that it is a particularity of the Turing architecture but can't confirm without test

    Fwiw the answer to this question is already evident.

    Not at all.  You don't have access to internal code to prove that

    1°/ There's a difference between math theory and implementation. Especially when transposing to GPU. And raytraycing is not just a bvh algo

    2°/ Pascal cards had 1/3 INT throughput vs FP32 and that was the case on most previous generation cards. From a performance standpoint, using FP32 is better

    3°/ Even when you have an algorithm performed on integers doesn't imply you have a data structure with INT.

    4°/ Luxmark and Vray benchmarks on the GTX 1660 which has dedicated INT make me think that datas are stored as FP32 for these renderers https://www.anandtech.com/show/13973/nvidia-gtx-1660-ti-review-feat-evga-xc-gaming/14

    Your huge INT gain is possibly what is seen with geekbench

    You may be right if Iray was not optimized (which is a possibility) and only a 1660 bench can tell that

    But as I said, with RTX cards, that is not relevant if the whole raytracing is taken care by RT core (and not just the BVH algo). Your dedicated INT will just sit there doing almost nothing as seen in the third graph from Nvidia

    And the question of Optix Prime on other RTX cards for scenes after 2017 still stands for me

    RayDAnt said:
    RayDAnt said:

    Also, we don't really know how Optix Prime works concerning Vram usage, but is it important ? I mean, since that won't trigger a drop to the CPU, I don't see the information as meaningfull

    If 20 (or 200) MB more memory usage is what it takes for a scene to trigger CPU fallback on your system, then it is very important. And the reality is that low to mid capacity VRAM GPUs are far more common than high capacity behemoths like the Titan RTX.

    You talk about memory load balancing to prevent CPU fallback and consider optix prime to possibly trigger a CPU fallback ? I rather consider it will not and that if there is not enough memory, Optix Prime will run partially or not at all (so is it pooly programmed?). Did anybody report that Optix prime was making the render to fall to CPU?

     

    And THAT is what makes even the lowly 8 core Threadripper look so attractive.  The 64 (60 + 4 to chipset) PCIe lanes can comfortably accomodate 4 GPUs, assuming you pick the right motherboard with appropriate PCIe 16 slot spacing.  It'll probably be a x16/x8/x16/x8 lane setup, but x8 is perfectly fine for rendering.  You WILL need to get a suitably large power supply of course if you are looking at that route.  But that's a subject for another thread I think.

    Here's a fun rumor/leak.  It's been noted (in a leak provided to Adored TV on Youtube) that the 7nm EPYC cards may have 160 PCIe or more lanes available in 2 socket configs.  The current 7xxx EPYCs have 128 available PCIe lanes in 2 socket configs, due to the other 128 (64 per chip) being needed for the infinity fabric cross-connect between CPUs.  The 7nm UMA (not NUMA) architecture in Rome apparently needs significantly less than 128 lanes for said cross-connect.  While this won't matter to 99% of the people here (who probably can't afford a 2 socket EPYC system in the first place), well for you high end types, that could mean 8 GPUs with 16 dedicated lanes apiece, easy peasy.  Of course, you'd also need a suitable motherboard/rack system to accomodate those 8 GPUs.  The interesting question will then be, how many of those slots will be PCIe4 as opposed to PCIe 3?  PCIe4 probably won't help much with render times, but we won't really know until the benchmarks of suitable PCIe4 GPUs and systems start showing up...

    There is also the PCIe riser solution which is way cheaper. Seems some people got it working even with a 1x link (I would have thought 4x was the minimum required) https://www.deviantart.com/3drcomics/art/External-GPU-Benchmark-751664498

    So the mining Rig is probably an other possible route. I guess you will get a very slow transfer from CPU memory to GPU, which may prevent real time render to occur

    If you search the Octane forum, many people built such thing

    For your question about PCIe4, it will probably bring marginal gain for classical GPU render. For real time raytrace, that could be an other story

  • tj_1ca9500btj_1ca9500b Posts: 2,047

    EPYC cards, what the heck is that, a card game?

    Just edited my last post to say EPYC chips, not cards.  I'm sure you guys knew what I meant, but I like to be accurate about such things...

    A low end 2P EPYC system may actually be within striking range for a fair number of budgets, assuming you don't mind low core counts.  So I'll be curious to see what the lowest core count for 7nm full size EPYC will be, how much they cost, and if they retain all 128 PCIe lanes and 8 memory channels (you should populate at least four of those).  The EPYC 3000's are basically repurposed Ryzens (that's an oversimplification, but they have significanlty less PCI lanes than the 7xxx series), so just to be clear, I'm referring to the lower end SKUs for the FULL SIZE EPYCs.

    The EPYC 7251 is only 8 cores, and currently retails for less than $500.  If there's a 7nm ROME EPYC with that low of a core count, but with the same 128 PCIe lanes and 8 memory channels, and 160+ lanes in a 2P config, well that could mean a core 2p 7nm system for less than 2-3K, not counting memory and storage, and a 1P system for not much more than a regular desktop system.  But of course we won't know much more about the 7nm EPYC Rome lineup until the official launch, so it's purely speculation at this point.  a 1P or 2P 7251 system has 128 PCIe lanes, which makes 8 graphics cards running at x8 easy peasy with leftover lanes for storage, etc., IF you can find the right rack setup for that at a decent price.  For the 7000 series, a 1P 16 core EPYC is probably the more economical choice over 2 8 core 7251's, but a single 7251 may be plenty for Daz if you are mainly doing multi GPU renders.

    BTW, here's some benches for the 7251, from back when it initially launched:

    https://www.servethehome.com/dual-amd-epyc-7251-linux-benchmarks-least-expensive-2p-epyc/

    Threadrippers aren't really ideal up for such large PCIe lane setups that accomodate more than 4 cards, although I suppose you could find some PCIe splitters (risers, ribbons, etc.) and run a bunch of cards at x4.  At x4 though, there's a significant performance hit, unlike the x8 vs x16.  Plus I've seen a few mentions of regular windows systems running into issues with memory addressing with that many GPUs in play, at least on the Threadripper systems. 

    I'm specifically referring to this thread, which is a bit dated and refers to a 1950x, so I'm not sure how the 29xx Threadrippers are faring with this issue:  Someone in that thread did mention that a corresponding Intel system at that time seemed to be handling 12 GPUs just fine.

    https://community.amd.com/thread/228930

    In my mind, an 8 card system would be ideal for setting up two rendering chanels with 4 cards dedicated to one Daz instance, and the other 4 cards dedicated to a second Daz instance, alternating between two Daz instances as you set up separate renders for each instance.  Or dedicating the first instance to rendering an animation sequence while you crank out individual renders with the second instance.  This could also be a strategy with 4 cards (2 cards per instance).

    But, to be fair, most people won't see much benefit in having more than 4 GPUs in the first place.  Really, 2 or 3 is probably plenty for most people, so this is pie in the sky for the typical Daz Studio user.  And it doesn't really relate to benchmarking, although it'd be fun to see some Daz Studio benchmarks with say 8 or 12 GPU cards.  Along with some 2990WX and 2970 WX benchmarks.

  • RayDAntRayDAnt Posts: 1,120
    edited April 2019
    RayDAnt said:
     
    RayDAnt said:

    OptiX Performance Boost: Based on multiple runs. Whether rendering each scene with OptiX Prime enabled lead to a better (shorter) Total Rendering Time than without it for that partcular hardware/software configuration. Take special note of the "No" answers for all the post 2017 scenes with the Titan RTX configuration. 

    The fact that the RTX Titan doesn't benefit from Optix Prime for the scenes after 2017 whereas the GTX 1050 does, also raises few questions :

    1*/ Why ? Obvious answer would be that the raytracing performance is a bit better with the RTX Titan than with the GTX 1050

    2°/ Does it expand to all RTX cards or is it just for the quickest (need somebody who has a 2060 to test to get a beginning of an answer) ?

    I guess that the answer will be that it works for all cards and that it is a particularity of the Turing architecture but can't confirm without test

    Fwiw the answer to this question is already evident.

    Not at all.  You don't have access to internal code to prove that

    The illustrations that post is an analysis of came directly from Nvidia and were generated by their developers with the luxury of having access to internal code. Hence why I bothered posting an analysis of it - because it was an authoritative source publishing proprietary information.

     

     

     

    You talk about memory load balancing to prevent CPU fallback and consider optix prime to possibly trigger a CPU fallback ? I rather consider it will not and that if there is not enough memory, Optix Prime will run partially or not at all (so is it pooly programmed?). Did anybody report that Optix prime was making the render to fall to CPU?

    Nvidia's own documentation on how exactly it is Iray - in conjunction with OptiX Prime - works explicitly states that it leads to increased vram usage to the point that it will cause some borderline too-large scenes to fallback to CPU versus without it. The point of the statistics I compiled was to discover how much of a difference it makes - not question whether there is a difference or not, because honetly I can see no reaon why Iray's developers would lie about that.

    Post edited by RayDAnt on
  • ArtiniArtini Posts: 8,528
    edited April 2019

    SickleYield starter scene:
    https://www.daz3d.com/forums/discussion/53771/iray-starter-scene-post-your-benchmarks/p1

    Optix off

    Rendering Time: 1 minute 50.60 seconds
    Device statistics:
    CUDA device 0 (Quadro P5000):      1486 iterations, 14.520s init, 94.336s render
    CUDA device 1 (Quadro P5000):      1458 iterations, 14.592s init, 93.811s render
    CUDA device 2 (Quadro P5000):      1477 iterations, 14.559s init, 93.772s render
    CPU:      579 iterations, 14.301s init, 94.300s render
    2 x Xeon Gold 6140 @ 2.3 GHz
    36 cores, 72 threads
    128 GB RAM

    Daz Studio 4.10, Nvidia driver 391.25

    Optix on
    Rendering Time: 1 minute 14.75 seconds
    CUDA device 0 (Quadro P5000):      1528 iterations, 14.784s init, 58.333s render
    CUDA device 1 (Quadro P5000):      1503 iterations, 14.747s init, 57.874s render
    CUDA device 2 (Quadro P5000):      1527 iterations, 14.746s init, 58.214s render
    CPU:      442 iterations, 14.808s init, 57.881s render

    image

    sy01.jpg
    400 x 520 - 29K
    Post edited by Artini on
  • ArtiniArtini Posts: 8,528
    edited April 2019

    Outrider42 test scene:
    https://www.daz3d.com/gallery/#images/526361/
    Daz Studio 4.10
    Optix on
    Total Rendering Time: 2 minutes 47.26 seconds
    CUDA device 0 (Quadro P5000):      1510 iterations, 0.698s init, 162.595s render
    CUDA device 1 (Quadro P5000):      1494 iterations, 0.707s init, 161.926s render
    CUDA device 2 (Quadro P5000):      1506 iterations, 0.671s init, 161.793s render
    CPU:      490 iterations, 0.797s init, 161.671s render
    Using OptiX Prime ray tracing (3.9.1)
    2 x Xeon Gold 6140 @ 2.3 GHz
    36 cores, 72 threads
    128 GB RAM
    3 x Nvidia Quadro P5000
    Driver version:        391.25
    Core clock:        1607 MHz
    Memory data rate:    9026 MHz
    Memory interface:    256-bit
    Memory bandwidth:    288.83 GB/s
    Dedicated video memory:    16384 MB GDDR5X

    image

    or42pic01.jpg
    720 x 520 - 75K
    Post edited by Artini on
Sign In or Register to comment.