Oct 31
iray and GPU FAQ
Some iray and GPU questions keep appearing throughout the blog comments so I thought it might be best to create one post to answer those most commonly asked questions. Read on for the list.
Question 1: How can I combine the memory of multiple GTX type cards? For example I have two GTX 580′s with 3GB of memory, how (or will) iray see that as 6GB?
Answer: As of today, it’s not possible to combine the memory of GPUs for rendering. No, attaching that SLI cable won’t make it work either. In fact if anything attaching the SLI cable will slightly decrease your performance with iray and rendering. If you want or need 6GB you have to use a Tesla 2070 or Quadro 6000.
With GPU rendering in mind I think it’s helpful to view the GPU’s as render nodes on a CPU based system. If you were to add a new render node to your ‘farm’ that doesn’t increase the RAM in your workstation. It just gives you another node to render with. That’s what happens when you add a GPU to your workstation and use a GPU rendering application.
NOTEs:
a. SLI is not required for applications like iray to recognize multiple GPUs. By default iray will use whatever compatible GPUs you have installed in your workstation. You can also manually assign GPUs via scripts and/or the hardware controls available in newer versions of iray.
b. If your scene exceeds the memory of your GPU, then iray defaults to using your CPU which will often be far slower than your GPU (depending on the GPU and CPU types).
Question 2: Those Quadro/Tesla cards are a ripoff. My GTX 580 renders so much faster than my Quadro FX1800. With that in mind I don’t understand why would anyone waste the money on these Tesla/Quadro cards?
Answer: Go to the nvidia site and do some research before buying a GPU. You’ll want to look at the available memory, number of CUDA cores, and their speed. With that in mind you’ll find the GTX/gaming type GPUs typically have more CUDA cores than a similar series Quadro GPU. Plus the GTX cards are usually clocked faster than similar Quadro/Tesla cards. More CUDA cores + faster clock speeds = faster renders. However, that also means more heat generated.
It’s been said that the GTX type cards aren’t designed with rendering in mind and may fail when pushed to the extreme like that. For me, the jury is still out on that one. I know heat = bad when it comes to electronics in general. But I think if a person were to keep the GTX type cards cool while rendering (85c or lower), they would last longer. Time will tell on this one as more people adopt GPU type rendering with various types of GPUs.
So, why would a person need to spend money on a Quadro/Tesla GPU for rendering? The primary answer is simple, memory. While the GTX cards may be faster, they are still only limited to 3GB of memory (as of the date I wrote this article). I’m using the 6GB Quadro/Tesla models because I need the extra memory. I could use more memory (8 to 12 gb would be nice), but that animal doesn’t exist yet. Bottom line, you don’t buy a Quadro/Tesla card because it’s blistering fast. IMHO you’d buy a Quadro or Tesla card when you need more than 3GB of memory for rendering, or if you’re burning through GTX cards, or perhaps for better viewport performance (that’s up for debate).
Question 3: Which GPU should I buy?
Answer: That’s up to you to decide. I say that because I don’t know your budget, what type of scenes you work on or their size (scenes + textures have to fit onto each GPU), whether or not you have ample cooling for GTX type cards, and/or how important (or good) your viewport performance will be, etc..
As a rule of thumb here’s what I’ve been telling others that asked the same question: Start with something like a 2gb or 3gb GTX card (a 3gb 580 is a good choice) to get a feel for iray and/or GPU rendering in general to see if it even fits into your pipeline. If it doesn’t, well at least you have a nice gaming card to use or sell. If you’ve tried it and discovered that GPU rendering will work well for you and your GPU temps aren’t terribly bad, then you may want to get another GTX type card or two for faster renders.
After testing, if you discover your scenes require a bit more than 3gb then you have to decide whether or not it’s feasible for you to invest in the higher end Quadro and/or Tesla cards. Typically you’ll use a Quadro series GPU to drive your viewport and Tesla for rendering. You can render with the Quadro GPU’s, but do NOT buy a Quadro for rendering only. I say this because the Tesla computing GPUs are less expensive. For example, a 6GB Quadro 6000 cost around $4,000.00 while a 6GB Tesla c2075 will cost around half that at $2,000.00.
As of July 2012, 6gb is currently the maximum amount of memory available on an nvidia GPU. If your scenes require more than 6gb, then you’ll have to get creative , like split render them, etc.. At that point you’ll also need to decide whether or not that extra work negates the speed of GPU rendering all together.
Also keep in mind that the dual GPU cards, like the GTX 590/690, share their memory. Therefore while it may have 3GB it’s split into 1.5gb x2. Therefore your 3d scenes will be restricted to 1.5gb of memory as well.
Question 4: iray is not a production ready render yet.
Answer: While it may not work well for all situations, I know many people that are using it in production work (myself included). No, iray doesn’t have all the features other CPU based rendering applications have (yet). But that doesn’t mean it’s not a production ready render. GPU rendering applications are certainly not a one size fit’s all type rendering solution. It seems to work best for product vis and some architectural vis scenarios. Bottom line, try GPU rendering yourself to see if it should be incorporated in your pipeline or not. IMHO that holds true for all rendering applications, GPU based or not.
Question 5: There’s a lot of noise/grain in my iray renders…
Answer: iray will refine an image over time by removing noise/grain to acceptable levels. Some scenes are more efficient at this than others. For example, an architectural interior scene lit by a small light source (or a few small light sources) will take a very long time to clear as opposed to an architectural interior scene lit by very large windows (daylight) and/or large area light sources. Your hardware plays a big role in how efficient/quickly a scene renders as well, just as it does with CPU rendering. You can also find scene optimization tips on the iray website HERE.
Question 6: Can iray render animations?
Answer: Yes.
Question 7: iray isn’t using my GPU.
Answer: It may be something as simple as you need to update your GPU drivers. Or perhaps you’re using an older GPU that doesn’t have CUDA cores? You can update your GPU drivers from the nvidia website. You can also find a list of CUDA-enabled GPUs here.
Question 8: Will my ATI card work with iray?
Answer: No, you’ll need an nvidia card for iray. At least that’s how it is as of the time of writing this article.
Question 9: How well does GPU “X” work with my 3ds Max viewport?
Answer: I personally have no idea. I’ve seen people using similar GPUs and their 3ds Max viewport experiences were completely different because they worked on different types of scenes or rigs, etc..
Question 10: What if I want to assign a specific GPU to use?
Answer: In the 3ds Max/Design 2012 Subscription Advantage Pack there’s an update to iray that allows you to specify hardware used in the settings. Prior to that you could use string commands or better yet, the iray manager script.
Question 11: I can’t use iray because it’s far too slow.
Answer: Path tracing applications aren’t right for every scene. For example, architectural interior scene illuminated with a few small light sources will typically take a long time to render with path tracing type applications. The same interior scene lit with large windows and/or light sources won’t take as long to render. So if your main job is illuminating interiors with a candles, path tracing type applications will probably not be your best choice in rendering software. However, it should work fairly well for things like architectural exteriors and product/automotive visualization.
Of course hardware also plays a huge role in how efficient iray and similar GPU rendering applications are…but that shouldn’t be surprising since it’s true for all rendering applications (CPU and GPU based).
Question 12: How can I optimize my scenes for GPU rendering?
Answer: Keep geometry limited to the basic amount of subdivisions needed. You’ll also want to keep in mind that geometry/proxy instancing on a GPU doesn’t work as efficiently as it does with more typical CPU based rendering applications (at least that’s how it is at the time of writing this).
Keeping texture maps as light as possible, and as few as you can get away with will be helpful with memory as well.
In iray ver.1 the frame buffer itself would consume memory but that has been addressed in newer versions so frame buffer size shouldn’t matter going forward.
You can also find some performance tips as well as tips for preparing content for iray at the irayrender.com website.
Question 13: What are your system specs?
My current GPU based workstation is built using these parts:
Mainboard: ASUS P6T7
CPU: Intel i7 960 3.2GHz
Ram: OCZ 12gb (3 x 4GB)
GPU01: nVidia Quadro 6000
GPU02: nVidia Tesla c2075
GPU03: nVidia Tesla c2075
Power Supply: Antec 1200 watt modular
Case: Corsair Obsidian 800D Full Tower
O/S: Windows 7 (64-bit)
Question 14: How much memory will I need for my scene?
Answer: Depends on your scene. Shane Griffith provided the following guidelines in is iray FAQ at the AREA site: For estimating memory usage, budget about 1 GB per 8 million triangles, to which you must also add 3 bytes/pixel for any referenced bitmaps.
A real world example would be THIS scene used 908mb of GPU memory. The scene stats were: 3,093,471 poly’s / 1,781,655 verts with several texture maps used throughout the scene.
Question 15: iray never seems to stop rendering, even when the results look clean.
Answer: Sounds like you have set the render time to unlimited. Which of course means the render will continue until you manually stop it..
Question 16: Can I render final images with iray or is it just an active shade preview?
Answer: Yes. You can render/save final images with iray and use it as an activeshade option.
Question 17: Using the primary GPU for rendering isn’t as efficient as having a dedicated secondary GPU for rendering because driving the monitor/display consumes some of the available memory. What can I do to help free up some of this memory?
Answer: You can turn off all the bells/whistles/pretty stuff in windows (like the AERO options, etc.). Shut down as many graphic consuming resources as possible. For example, Oguz Birgoren reported that switching 3dsmax over to OpenGL frees up more GPU memory than using Nitrous or DirectX.
Question 18: iray is not using my GPU to render / I see some GPU error messages
Answer: Make sure you’re using a version of 3ds Max that supports rendering with your GPU(s). For example, if you’re using a kepler based GPU (6XX series GTX cards for example), then you must to be running a version of 3ds Max that supports that GPU. That means you must be running 3ds Max 2013 with PU06 or higher to use a kepler based GPU.
You’ll also want to make sure that you’re running the latest drivers for your GPU. Also ensure your power supply (and motherboard) can handle the GPU(s) that you have installed.
51 comments
Trackback e pingback
-
GTX/Quadro/Tesla – my opinion on today’s GPU selections for rendering | jeffpatton.net
[...] NOTE: You may also check my iray/GPU F.A.Q. HERE. [...]
thanks your the best !!
ps: have you used octane render ? what is your opinion ?
Octane is nice. Seems to be evolving pretty well. I’m looking forward to their upcoming 3ds Max plugin that they are currently developing.
Hi Jeff – A while back I decided to go watercooling with the EVGA hydro copper 580 gtx 3gb. My temps don’t top 47C during iRay renders and the speed is pretty good. The memory is definitely limiting, especially with exterior scenes + foliage, but it’s not TOO bad. I’ll let you know if my card fails
Anyway – just thought I’d throw in what data I had. I would hesitate to run GTX cards without watercooling. Appreciate your posts!
Hi,
does the system (WINDOWS/IRAY) recognize gtx 580/590 with full 3072MB RAM capacity(like it was one card). or those cards act like 2x1536MB in SLI.
In regards to rendering with iray a 590 is seen as 2 x 1536mb. SLI doesn’t make any difference to that whatsoever.
On work we use 4 Tesla C2075…. And for our project they are too slow. I hope for a lot of more memory on gpu´s….
Hi Jeff! Great website. I like the idea of seeing GPUs as render nodes. Based on what you wrote, I take it you don’t SLI your cards. So does this mean if I have I have 2 Quadro 2000 cards (which don’t have SLI capabilities) I could still use each card as a rendering node and actually assign a set number of rendering frames to each video card?
i.e. frames 1-1000 renders on Quadro 2000 #1 and frames 1001-2000 renders on Quadro 2000 #2
Is this really possible?
“So does this mean if I have I have 2 Quadro 2000 cards (which don’t have SLI capabilities) I could still use each card as a rendering node and actually assign a set number of rendering frames to each video card? i.e. frames 1-1000 renders on Quadro 2000 #1 and frames 1001-2000 renders on Quadro 2000 #2 Is this really possible?” - No, I don’t know of anyway you can assign specific frames to multiple GPUs in one machine. The only way I can imagine something like that working is if you took one of those GPUs and placed it in a separate workstation and used backburner to assign specific frames to each machine. However, I don’t think that would be as efficient as using both the GPUs in one workstation and letting them calculate every frame.
When you use your GPUs to render, do you experience any lock ups with your system/workstation?
I’d like to be able to render using the GPUs and still be able to do my 3d modeling work at the same time and check emails/surf the web. I really appreciate your input Jeff.
It should be more responsive if you don’t use your primary GPU and CPU(s). That being said, I’ve also noticed that in iray V2 (3ds Max 2012 SAP) it seems less responsive at times than V1 even when the primary GPU isn’t used in the render process. Maybe it would be better to install the extra GPUs into a render only machine and assign the render job to that machine while you continue to work on your primary workstation.
hi! Thank you Jeff for your very cool blog.
i’m looking for a new video card for I-Ray rendering with Maya 2012. Actually i use a 460gtx but often i have “out of memory” messages.
Looking around the web, i found this card with interesting specs (memory and cooling) and a low price:
GAINWARD GTX560 TI PHANTOM 2048MB GDDR5
Anyone knows that card? Any feedback?
Thanks
That computer is Beastly! Either nVidia is handing you cards as some sort of sponsor deal or something, or you dropped 80% of your cost into graphics cards. Not so bad when you call that 4 render nodes though.
Still, I’d love a twenty thousand dollar machine sitting on my desk. I’m curious about 2 more things with that setup. Hard drives? Did you do SSD and SATA mix? I’m using a PCI drive now, but it doesn’t seem you have any room left.
And cooling: Did you add any? I know quadro/tesla runs a bit cooler, but that is a Lot of heat to be contained in one box. I’m running two GTX 570′s in one box, but wouldn’t do more than that without some heavy modification.
Very nice article, detailed, and I appreciate the comments about which bits are not necessarily true in the future. And I’m another you can add to the list that uses iRay in production. I used MR on my first job here, and every job after that was iRay. Most designers here have decent video cards, so running backburner over night I can chug through quite a bit.
“Still, I’d love a twenty thousand dollar machine sitting on my desk.” – I hear that a lot for some reason. It’s a common misconception that a setup like this would be horribly expensive when in fact a similar GPU configuration to what I have can be built for about $7000.00. For example: You can purchase the tesla c2075′s (6GB models) for around 2k each. Then use something like a $700.00 Quadro 4000 to drive your display(s) instead of the $3500.00 Quadro 6000.
“Hard drives? Did you do SSD and SATA mix? “ – I’m still using all SATA drives.
“And cooling: Did you add any? I know quadro/tesla runs a bit cooler, but that is a Lot of heat” – Nothing special other than I typically leave the side door off my case. Currently the temps run like this:
GPU0 (Quadro 6000) – 71c idle & 89c rendering
GPU1 (Tesla c2070) – 79c idle & 91c rendering
GPU2 (Tesla c2075) – 58c idle & 75c rendering
GPU3 (Tesla c2075) – 45c idle & 70c rendering
That’s with the GPUs stacked right against each other. The newer c2075′s certainly seem more efficient than the other two cards. At any rate I hope to find a motherboard that will run 4 GPUs and leave a little breathing room in between the PCI-e slots as I’m sure that would help bring the temps down further.
Hi Jeff,
do you find that your Quadro and Tesla are similar, as far as render times?
So if you used only the Quadro to render to a certain quality, this would be the same for the Tesla?
And then do you see a gain of approximately half render times when stacking together?
Cheers,
Steve
Yes, for rendering similar spec’ed tesla & quadro GPUs perform about the same. They usually have similar clock speeds & number of cores.
Outstanding info
! Thank you, it’s sincerely appreciated!
Tesla newbie
I just bought a Dual 6core 48 gigs of ram Quadro 2000 and a Tesla 1050. I have yet to see any information in 3Dsmax or other apps that I am using the Telsa card. Do I need to setup the Tesla card up some how or is it working with the Quadro. Seems like this should be smoken fast. Cuda cores are showing 192 in the system info
Thanks
You will need to use a GPU rendering application like iray, VRay R/T, Octane, or Arion in order for the GPUs to have any impact on rendering. You can monitor the GPU usage in many ways, either by unique applications (nvidiainspector, GPUz, etc.) or by watching the render message window in 3ds Max.
Thanks
I am using Iray and I have looked at Octane. Is it best to use A&D shaders with Photometric lights? Most of the stuff I a rendering look like Final gather renders. Better yet any good links to getting started with Iray.
“any good links to getting started with Iray.” – As with anything one of the best places to start is the actual documentation that ships with the application. For iray, the 3ds Max help file explains what materials/lights/etc. are supported. Here’s a direct link to the online help file.
There’s the irayrender.com website with information. Users can ask questions at sites like the AREA website or the nvidia advanced rendering forum.
Of course there’s also quite a bit of iray information on this very website.
Sorry, that was a lame question. Thanks for the info. I have gotten a lot of info already from this webpage. Your stuff is awesome btw.
Thanks
Hi i have a question?
if i use quadro 6000 for performance display and 6gb Ram for big scene
and use Gtx 590 x 3 for many cuda and faster clock rate gpu
what do you think this plan ?
“(CUDA Cores#) x (Clock Speed) x (GPU Architecture) = Relative Performance.” http://area.autodesk.com/blogs/shane/the_iray_faq
Hi i have a question?
if i use quadro 6000 for performance display and 6gb Ram for big scene
and use Gtx 590 x 3 for many cuda and faster clock rate gpu
what do you think this plan ?
“(CUDA Cores#) x (Clock Speed) x (GPU Architecture) = Relative Performance.” http://area.autodesk.com/blogs/shane/the_iray_faq
“use Gtx 590 x 3 for many cuda and faster clock rate gpu” – It always goes back to memory. As long as your scenes can fit onto the 1.5gb of the GTX 590′s (remember 590′s are dual GPU setup, so they split the GPU memory) then 3x 590′s should be fine. If your scenes are larger than 1.5gb, then the 590′s won’t be used for rendering.
Thanks for posting this, it’s brilliant <3
Greetings Jeff…Gtx 680 came out and it has 1536 Cuda cores compared to Gtx 580 which have 512 and Tesla and Quadro which have 448…by this it should be faster cca double…however no reviews yet…just thought it might be interesting…as well the speed is over 1000mhz and as we know gtx is 750-850 so lets see…the downside they are making gtx 680 only with 2GB…
…update… seams i was wrong http://www.tweaktown.com/reviews/4665/palit_jetstream_geforce_gtx_680_4gb_video_card_review/index.html…keep up the good work…
..new update …for the gtx 680 Nvidia still did not as it seams put out drivers for support in iray …
Hope Jeff you don!t mind that I post this news here; …thought it might be interested for you too.
I don’t mind at all. Thanks for helping spread the info on the Keplers as many people are asking about their performance in rendering.
I’m running the newest version for Vray with Vray-RT. On my work station I have a quadro 4000 2gb – Question is this – If I install another quadro 4000 in another PCI-E slot I do not need to SLI them together because Vray-rt will see all available cuda cores, correct?
Correct, you do not need to SLI the GPUs in order for VRay r/t to use them. It will use all available GPUs and/or you can use the included tool to set which GPUs are used in the OpenCL mode.
PS – If you already have an extra Quadro 4000 then I understand why. But don’t buy an extra Quadro specifically for secondary GPU rendering. You can get a similarly equipped Tesla for the second/third/etc. render GPU for significantly less money.
Thanks Jeff for your reply – just wanted to make sure – have a great weekend
Hey Jeff, your work and your studies are really very impressive. well i m working on Iray since few months, i do more on interior scenes, as Light entering from window and illuminating the whole scene. Well it takes hell lot of time to render, even after rendering to 12- 15 hrs it has many noise. the image is not very clear.
And Regarding Config of my System, i use Core i7 950. GTX 560 Ti. & 16GB ram.
whats the solution for this, if u want i can upload the image too.
There’s THIS, but other than that there’s no tweaks to make with iray…other than adding more GPUs to speed the rendering process up.
Hi Jeff,
I wanted to know, that does iray in 3ds Max work with “Kepler GPU Architecture” cards?
i.e., GeForce GTX 680 and 690
http://blog.irayrender.com/post/20003193613/iray-on-kepler
Hi Jeff,
I would like to share one thing that I found out while trying to render at 4K resolutions in in iray with GTX card with lesser memory than decent Quadro or Tesla.
I try to use CPU-z to monitor the memory usage while I am rendering a scene with iray. From what I see since Nitrous stores the model data on the GPU memory itself. It does leave very less memory to render. If you maximize the viewport, than GPU tends to use more of the gpu memory. This is also the same for D3D viewport.
At first I thought when we press the make pretty button, max would clear GPU mem and we would be good to go.
But it’s not! So while rendering GPU memory is filled by two versions of the model. I’m guessing thats because to make UI flawless. But this creates a real bottle neck.
To avoid that I tried to change the viewport to software rendered viewport, but I think that option is no more viable and disabpled to be selected. So I chose OpenGL and now the car model I’m working on does not allocate any memory on GPU. It stands still at 90MB. So all the free memory can be used to render. I can now render at higher resolutions.
I tested this with GTX 400 series card. I’m not sure if it works the same with quadro or with quadro performance drivers.
I’m not sure this has been covered somewhere else. I thought it would be right to share it here, as your blog became the bible for iray.
Thanks,
“So I chose OpenGL and now the car model I’m working on does not allocate any memory on GPU. It stands still at 90MB. So all the free memory can be used to render. I can now render at higher resolutions.” – Indeed, when you’re using the primary GPU for rendering it will have to share it’s memory with whatever is happening on your screen/monitor. I would imagine all the AO/DOF/Exposure/etc. that happens with Nitrous would consume/hold more memory than other modes. Perhaps they (Autodesk) will be able to optimize that (or freeze/release it during rendering) in the future but that’s just a guess as I don’t know that to be possible for certain.
Either way this is one reason for nvidia’s “Maximus” configuration where you have multiple GPUs in your workstation. The secondary GPU(s) handle the rendering process without the display device(s) eating into the available memory and your primary GPU can still be used for it’s main tasks if needed.
Thanks for bringing this point up. I’m sure there are many others out here that use the main GPU and will benefit from knowing OpenGL mode will free up some memory. I’ve added your comments to the FAQ directly.
Oh. That’s cool.
I was able to free up 700 MB GPU memory driven by Nitrous simply by switching to OpenGL.
It’s a real deal when you’d want to use main GPU along with your 2nd card.
Thanks.
This might come out as a very odd question, but still having uncertainty with. Say you have 2 video cards, and your monitor is hooked up to your primary video card, if I assign my render to render off of my secondary video card (which has no monitor hooked up to it), will I still be able to preview my render progress on screen (which is hooked up to my primary video card)?
Thank you so much for your time Jeff!
“if I assign my render to render off of my secondary video card (which has no monitor hooked up to it), will I still be able to preview my render progress on screen (which is hooked up to my primary video card)?” – I think that’s a great question. The answer is absolutely yes. You will still be able to preview your render progress on screen, activeshade, etc. just as you normally would regardless of which GPU you use. Using a second and/or third GPU to render in lieu of the primary GPU won’t prevent you from viewing the render as you normally would. The benefit is that you can work on other things without any speed hit while rendering on the non-primary GPU(s).
Outstanding!
Seems like the best budget option would be to go with a Tesla c1060, which has 4gb (meets the 3GB minimum), has a decent number of cores, and can be had for around $500 online. Rather than messing with GTX 5xx or making a bad buy and getting GTX 6xx.
Just a quick question, can Maya (software/hardware/MR) take advantage of gpu rendering in any way?
Keeping in mind everything you said here and considering to assembly a WS for production rendering, would be better use a n.1 or a n.2 CPU mobo?
In any case would be better a X79 or a Z77 chipset mobo?
Thanx!
hi jeff,
i would like to use quadro fx 1700 for my viewport performance in studio max 2013.
as for the rendering/computing part, i would like to use 2x gtx 680.
do you think this will work flawlessly without any conflict?
please advise if there is any bottleneck too.
rdgs,
mel
I’ve been reading and comparing GPU’s for the past 3 hours, from desktop to workstation GPU’s, from nvidia to amd and I came across this beautiful post and you sir just saved my day. Thank you.
I have two questions…. whats your opinion on using many small GPUs (like for example say 6-7 Nvidia 9800GT singleslot) vs 3x GTX280 or similar?
And second, do you know how important PCIe-lanes are? I mean, getting 5 times PCIe x4 is no problem, but more than 3 PCIe x16 are a pain in the ass.
Do I really need 16 or 8?
Jeff. Thanks for sharing your extensive knowledge, It’s much appreciated.
Could you please let me know how long the restaurant renders took in regards to set up time and also rendering..
Question no. 2, would you be interested in a trip to Australia to teach us your Iray skills?
Otherwise I’d be happy to come to you if you’re interested in doing that? Happy to pay for your time.