Hardware for 3d softwares

[__Xav__]

Hi all,

I open a new thread to discuss everything related to hardware for 3d softwares.

I personnaly use Daz3d and I will upgrade my PC. I just want to know how many RAM I have to increase to match with my future GPU (probably a GTX 1070).

For the moment, I have 2x2Gb + 2x4Gb = 12Gb. My mother card have the space for only 4 strips.
So 2 options come to me :

1. Replace my strips of 2Gb by strips of 8Gb ; in this case, I will have 24Gb in total
   
2. Replace all my RAM by strips of 8Gb ; for a total of 32Gb
   

Knowing that I want to operate larges scenes, with high quality HDRI or large infrastructure and a few characters, will 24Gb be sufficient or not ?

When the Cuda technology of the new generation cards of Nvidia (GTX 1080 & 1070) will be recognized by Daz3d for the rendering in Iray (this is not the case in 09/07/2016), I will have 8Gb on the GPU. Which one of the RAM will be a limit while rendering in Iray ? The 32Gb or the 8Gb of the GPU ? In other words, is it important to have more than 24Gb when we don't have a quad sli of Titan X ?

Thanks for your answers. And sorry for my english, I try to do my best.

[sikotikxiii]

Both types of RAM are used for different things in Daz Studio, and the more you have of both, the more headroom you will have overall, so I would suggest getting as much as you can afford/fit, regardless.

In detail: If you exceed the VRAM (Video RAM - the RAM on GPUs), Daz Studio will essentially stop using it for rendering (might as well, anyway, as the processor at that point is trying to render and distribute, something that taxes any CPU for everything it's worth). If you exceed RAM... Well, assuming Daz Studio doesn't crash first (it probably will), your computer will crash. Everything in your scene and Daz Studio processes themselves will use RAM. You will use far more of this than you will VRAM in most scenes/cases. VRAM will have to store, at various points of concentration, geometry and textures. Your processor will be shuttling everything back and forth between the card(s), which also uses RAM. As an example, I can (and have, repeatedly), with careful optimization of textures (not in Iray, but the theory should still apply, I think), gotten a scene that uses ~40GB of RAM to render in less than 5GB of VRAM. Textures matter a lot, and HDRIs, in particular, can be either a huge factor or almost inconsequential depending on far too many factors to quantify here succinctly. VRAM will always be your limiting factor when rendering on a GPU, but not having enough RAM will either bring your system to a crawl or crash it, no matter what you're doing. Daz Studio can use a lot of both, frequently at the same time. Large does not equal high resource when it comes to 3d modeling either. It's all about the construction of the geometry, and the textures in the scene. There are cars made in 3DSMAX and other modeling programs that would, by themselves, blow any Stonemason set out of the water, in terms of resources, if just loaded into Daz Studio and used as is (I tried it, and it wasn't fun, I assure you).

[De4lt]

sikotikxiii explained it very well. The memory of your computer (RAM) is basically there for constructing the scene, and rendering using the Processor (CPU). 

If you are rendering with Iray, the faster option is to render with the video card memory (VRAM). 

The more you have of both, the better. 

e.g. My laptop has 16Gb of RAM, but only 2Gb of VRAM. My PC, on the other hand, has 32Gb RAM and a GTX 980Ti with 6Gb VRAM. I use my laptop to construct my scenes, but I save the scenes and then render them on the PC. 

You are right that, currently, the newer nVidia cards aren't supported for Iray yet, but it's only a matter of time before that gets fixed. 

There is talk on the Daz website, however, that having one powerful (and expensive) card is not necessarily the best thing for 3D rendering software, especially Iray, and that two less powerful/less expensive cards working together can actually be better. Iray works better with more memory in order to pack more into the scene, but it also renders faster with more CUDA cores. I also have a GTX 970 in my PC (4Gb VRAM). If I keep my scene size within 4Gb I can use both graphics cards to render the scene. I benchmarked a scene recently, and one that took 4hrs 15min with just the GTX 970 took just 1.5hrs with both cards working together. 

[__Xav__]

Thank you very much of you two for the explanations. I confirm that 12Gb regularly leads to crash my computer. I will up my RAM to 32Gb of DDR3, until my new config in a few years with DDR4.

You are right that, currently, the newer nVidia cards aren't supported for Iray yet, but it's only a matter of time before that gets fixed. 

Yes, I cannot wait !
There is already a comparison in Blender between the top cards and the new Pascal GPU architecture with its GTX 1080, whereas the drivers are not yet developed. We can see that the GTX 1080 has a very big potential ! It is sometimes a lot faster than the GTX 908Ti.
https://lionrender.com/2016/06/16/top-gp...-gtx980ti/
About the poor performances of the GTX Titan X, it seems to be due to Blender, which is just its weakness.

[1e177491968]

My brother raced old sports cars and the saying was that if you want to go faster....spend more money. The same is true with computer graphics. While the post is a bit tongue-n-cheek, it really does hold true. You can't have enough ram or a fast enough video card or enough hard drive space and so on. When I build a computer I read all the computer gaming magazines to bone up on the available hardware, I subtract the cost of my mortgage, gas for the car, cat food for my pussy then spend whatever is left over on as much hardware as I can get. Technology quickly dates whatever I own but I don't get concerned with that. No matter what you build, it won't be fast enough.
Oh, and don't buy on a promise. I bought a GTX 960 because it works NOW. I make money with my computers and can't afford to wait for developers to trickle down to my requirements.

[Trety]

Question..
If you could have 2 GTX 1080's or 1 Titan X, which would you choose and why?
I will need to make a decision on this by end of Sept.

[De4lt]

Question..
If you could have 2 GTX 1080's or 1 Titan X, which would you choose and why?
I will need to make a decision on this by end of Sept.

For me the only deciding factor would be whether I need 12Gb VRAM or whether I could get away with 8Gb (bearing in mind two x 8Gb cards will still only render an 8Gb scene becasue the entire scene has to fit in BOTH cards).


Titan X - 12Gb VRAM, 3584 CUDA cores. 
nVidia's price for Titan X is £1,099 (may be cheaper elsewhere).

GTX 1080 - 8Gb VRAM, 2560 CUDA cores. 
nVidia's price for GTX 1080 is £619 (may be cheaper elsewhere). 

When it's supported by Iray and Daz Studio, for me, it would be a no-brainer. Far more graphic processing power in 2 x 1080's than one Titan X. 

(source of my info is nVidia's own Geforce website).

[sikotikxiii]

I would add a caveat to De4lt above:

The 980ti would be a viable and cost-effective candidate in that price range (~$600 USD each for the aftermarket ones), with 2816 CUDA cores, 6 GB VRAM, at higher speeds (in general, watercooling changes things slightly, but not too much). You can get most things below the 6GB VRAM threshold, if you optimize properly (I know, I'm using 4 of them). 2 of those would be roughly equal in price to one Titan X, while giving a surplus of 1448 CUDA cores. Not top of the line, no. Nothing stays that way anyway, and they also work now. I would also suggest that learning the optimization with a lower limit now would pay off more in the long run, as when you do upgrade in the future (when the new stuff actually works in anything and will be cheaper anyway), you'll be able to get far more out of the upgrade.

My two cents: you shouldn't need 12 GB of VRAM unless you are working for a studio with incredibly exacting specs in what is being used how, and you can't really call the new models an upgrade until they actually, y'know, work 'n stuff. Given the two options, I'd pick the two over the one any day (assuming, of course, that either case would actually function as needed at that moment), as performance scales almost directly with how many cores/cards you throw at a render.

[ravenpost]

Just a remark about the 6-8/12GB  VRAM cap.
My experience is that a fully clad girl is about 2GB....
So if you want 5 actors in a pretty well lit indoor scene (the most difficult in Iray) with some props, 8 is not enough.

[sikotikxiii]

There are a lot of variables that go into how much VRAM any given thing takes, honestly. Displacement and subdivision being the biggest on the geometry side, and texture resolution on the texture side, as well as the resolution you are rendering. Plus lights, as mentioned, but only to the number of passes and depths needed, and that amount is typically miniscule in comparison the the rest. Unless you are rendering a 4k image of the 5 figures close up somehow, you can turn down subdivision, use lower resolution maps, and turn off displacement to lower the VRAM requirements significantly (well below 2gb per figure, properly done and depending on your render resolution, you could get that scene to fit on a 4gb card). Alterernate render engines do this much, much better and more transparently than Iray, but the options still exist. The lighting only affects calculation time, not VRAM required, beyond the textures that said light makes visible. How many passes the lights have to make and reflections work the same, they only matter if they can be seen, and rarely add to VRAM needs. As I said, learning how and when to do those optimizations effectively is far more useful (and much, much cheaper, since knowledge is free aside from time invested) in the long term than constantly throwing money at components to compensate for the lack of that knowledge. But of you have the money to throw, I suppose there's nothing stopping you. Seems silly to be here torrenting in that case, though, in my humble opinion, for whatever that's worth.