VRAM vs CUDA core count?
paul_ae567ab9
Posts: 231
Not having much luck searching for answers.
I am building new machine.
Some REALLY expensive GPU's out there but all the hype is aimed at gamers not content creators. I do not need fast game performance just reasonably quick render times.
Should I be looking for more VRAM or more CUDA cores?
How do each play into various render engines?
Iray v OpenGL vs 3Delight
Daz 3D is part of 
Connect
DAZ Productions, Inc.
224 S 200 W, Suite #250
Salt Lake City, UT 84101
Licensing Agreement | Terms of Service | Privacy Policy | EULA
© 2024 Daz Productions Inc. All Rights Reserved.
Comments
VRAM every day of the week if you are going to render in Iray.
My basic understanding is that the scene must fit withing VRAM. If scene fits with spare what does more VRAM do for render speed vs more cuda cores?
Cuda cores don't matter if your scene doesn't fit in VRAM of your card, it will drop to CPU and the rendering speed will be 10, 20... times longer (think hours).
The RTX 3090 is of course the king if you still have a spare kidney, the new RTX 3080 Ti has 12GB's of VRAM but so does the RTX 3060 (non TI).
I'm using RTX 2070 Super that has 8 GB's. If I was on the market for a new one, I wouldn't buy one with 8GB's anymore and as an update to my current one, even 10GB versions wouldn't cut it (too little for an update)
There is an benchmark thread here in the forums, where you can see how the cards are performing.
Vram will determine how much you can fit in to a scene. Textures use up more vram than the models do, so large 4k textures on everything will eat up a lot of vram fast.
Cuda cores do matter as they are what Iray uses to do the rendering. More is better.
GPU clock speed also matters as it controls how fast those Cuda cores do their job.
I saw results of some benchmarks of the 3090 versus the 2080 and a few other nvidia cards recently on youtube and the 3090 did the Blender BMW benchmark in only a few minutes versus 12 or higher for the other cards. Think it was gamers nexus youtube channel.
I understand the fact that if scene does not fit it will note use cuda.
I will assume for sake of this discussion the scene will fit.
Back to the question. Will more VRAM or more cores most improve render speed.
If I add a replacement card (current card is 4GB which has never appeared to have problem holding scenes ) do I need to spend $$ for more VRAM or more cores. If I were wealthy I'd just buy a super computer! LOL
Iff the scene fits then cores, within a generation, matter - comparing coresd across generations (e.g. Turing to Ampere) is another matter.
If you are rendering in Iray, it sounds surprising if your 4GB card "never appeared to have a problem"... Are you sure the renders were done on GPU and not the CPU?
The benchmarking thread;
https://www.daz3d.com/forums/discussion/341041/daz-studio-iray-rendering-hardware-benchmarking/p1 ;
So how do we know if a scene "fits?"
Beforehand, your experience may give you an idea, but with the variance on creator skills and understanding of the different aspects, the only way to know for sure, is to render the scene - DS will then tell you if it fits or not.
3Delight rendering is done on the CPU, so that would be dependent on CPU core count and system RAM. The more CPU cores, the faster the render.
For Iray, it would be:
#1. VRAM and system RAM. Insufficient VRAM = render is kicked to the CPU; insufficient system RAM = crash the system. So getting an RTX 3090 with 24 Gb VRAM and only having 32 Gb RAM would be a big problem and a waste of a powerful GPU.
#2. CUDA Cores. The more you have working together, the faster the render, but again, if you're using multiple NVidia cards, you will need to have enough system RAM to support it which can be anywhere from 2 to 4 times(or greater still) the total VRAM being used to render.
As for what you should get, we couldn't tell you that since we have no idea what kind of scenes you work with. However, I will say that if you intend to work with large detailed environments & multiple HD characters, you'll need a beefy RTX GPU & the system RAM to support it(i.e. think TRX40 motherboard with a Threadripper CPU and at least 128 Gb RAM). If not, the only alternative would be to spend extra time on optimizing your scenes by eliminating/making textures smaller, lowering SubD on characters, etc. Given the way the current GPU market is, more are probably relying on post-work to add effects/details instead of buying overpriced hardware.
To answer your question directly ... As far as rendering speed, core count matters. Memory size has little effect on render speed for the most part, with the caveat regarding memory bus speed. I used to run three Titan RTX in my machine to get the CUDA count as high as possible, running in TCC mode using NVLink to get the VRAM to be additive, instead of just sharing. However, render speed in WDM mode or TCC mode (which made the 3 GPU's appear as one) didn't effect render times much. TCC mode just allowed me to load much bigger scenes. I now run single 3090's in my machines, which has more than enough VRAM. This is about 3,000 cores less than the three Titans at 13,824 cores. The main takeaway here for you to consider is even with less CUDA cores my render speeds are equivalent with just one 3090, due to the performance improvement of the 30xx series Ampere cores compared to the Titan RTX's earlier gen Turing cores. So more of the newest generation of CUDA cores is also important to consider, and part of this dicussion ... not something you put aside for another conversation.
Contrary to what others have said System RAM and GPU VRAM are not interdependent. Yes ... if you have insufficient system RAM (or any insufficient resources) you could crash, but that's the case with anything and is not exclusive to DAZ/Iray. However, system RAM has no effect on rendering performance once the GPU is loaded and running. The two are not mutually exclusive. System RAM plays a part in scene load times, nothing more. I suspect the differences people see drives an assumption based on improved load times, which is more a factor of PCI lanes and how they are connected directly to the CPU instead of through the chipset; like is the case with only a generation previous compared to the something like a newer TRX40 chipset. So people are assuming it is a RAM thing, instead of understanding the difference between newer and older generations of CPU/chipsets architectures.
With Iray once a scene is loaded into VRAM the system resources do very little in the render processing. This is proven by Iray Server, which is a separate machine rendering, independent from the host authoring machine. The scene is loaded across a network, and Iray renders on that machine using as many GPU's present. My Iray Server has run on as little as 16Gb with no hit on render performance. Yes ... It was unstable, but once I upgraded the RAM to 256Gb I saw no change in render time. In a normal setup you can have 32Gb of system RAM or 256Gb or 512Gb like I have, and you will see no difference in actual render times. Load time yes ... but not rendering; and we are only talking about a second or two in load time differences.
So in conclusion, if you are looking for speed in Iray the rule of thumb is, get the GPU you can afford with the most newest generation cores. All 30xx series will have enough RAM for all but the most complex scenes.
No point in getting into the whole GPU market situation, it has nothing to do with what the OP is asking about; a technical question. If he can afford or really wants the best render speed who are we to say otherwise about costs. I have three 3090's for my three workstations one each, I was able to get them, and can justify what I paid (which wasn't too bad). It's all relative. What I can say definitively is 30xx series is by far much faster than 20xx series given the same CUDA count as far as Iray is concerned. I ignore test result found online, basically useless information for our purposes with Iray.
My last comment ... the 3090 is a beast with Iray. Nuff said.
The more VRAM a scene uses, even more RAM it requires, just look at the results of my test.
Case A) was just one lightweight G8 character with lightweight clothing and hair
Case B) was four similar G8 characters with architecture
Case C and D) Started increasing SubD on the figures to see at which point the rendering would drop to CPU (= Case D)
Running an 8GB RTX 2070 Super with 64GB's of RAM
System RAM is also used in preparing a scene for the GPU, and the bigger the scene the more likely it is to run out - with scenes that demand a larger GPU then that may be a factor with lesser amounts of system RAM, hence the advice to aim at roughly double.
All I see there is memory size and memory used, not render times, which is what the OP is asking about. Of course if you increase the scene size it's going to use more resources. I would not need a spread sheet to know that. Now if you can show data on the difference of one scene with different system ram amounts and how that single change effects render time (minus load time) then we can talk. What you are talking about is not relative to the OP, where they clearly state " ... assuming the scene fits". So they are not concerned about scene size, where an 8Gb card was always enough. He probably only has 16Gb of system RAM as well, which is fine if his scenes are less than 8Gb loaded into VRAM.
The fact still remains, when a scene is loaded into VRAM and Iray begins rendering, system RAM is non sequitur.
If you look at the results, on a 8GB card, geometry+textures+working space 5.7GiB's on the last test that rendered on GPU, the rest is used by the base load of OS+DS+Scene+Something that I haven't identified yet. During that rendering RAM usage was at 37.8GB's = Quite a lot more than 16GB's
In practice about the maximum the geometry+textures can have (on 8GB GPU) is around 4.5-5GiB's of VRAM and that isn't hard to fill up, and the VRAM usage will start slowing down rendering the smaller Working Space the GPU is able to allocate (=1.75GiB's if there's space for it)
We have seen a number of people wondering why their computers were crashing when they tried rendering in Iray, and the reason was having 8 or 16GB's of RAM - I would not give false hope to anyone that later regrets believing it.
Pagefile at work doing what it is supposed to do, acting like ram, very slow ram.
I agree with you that the amount of RAM you have in your system(assuming that the amount is sufficient or exceeds it) has absolutely no bearing on render speeds. However, you're missing the point that if you do not have enough system RAM, nothing else will matter because the system will crash. Given the OP's question(s), I think it was an important bit of info to mention as they may not be aware of this. The scene may very well fit in VRAM, but it needs the necessary available system RAM to get there. For example, a system with 128 Gb RAM would not be able to fully utilize all 48 Gb VRAM of an RTX A6000 for renders because this will most likely need in excess of 128 Gb RAM just to load it onto the card.
Again that's all well and good, but not the question at hand. You seem pre-occupied with stability and CPU rollback, which is not the question of the OP. He wanted a simple answer to a simple question: Is speed related more to VRAM or core count. Stop confusing the issue. Anybody who has used DAZ for any lengtrh of time knows things get wonky if you don't have enough system RAM. Same for Adobe Premier, Maya, Davinci or any software that uses large image files. Your talking anout something unrelated to pure Iray render speed.
I gave the OP a link to the benchmark thread, which is the only benchmark existing for Iray rendering in DS, so that he/she can have an idea about the differencies in rendering speed with different cards, but as @magog_a4eb71ab pointed in his message, you are missing the point of us not talking to "Anybody who has used DAZ for any lengtrh of time", but people that do not know how DS or any other software that uses large images works and behaves.
They do not understand that it is of very little importance, whether the rendering takes 15 minutes or 14 minutes depending on the speed of the card, because they are used to rendering hours, and mislead by the gaming communities about what's good and what's not good. DS+Iray has different needs than gaming.