GPU Renderers and VRAM utilization

[Dimitris Tolios]

I have been trying to put together information on the VRAM requirements for iRay and VRay GPU. I am planning on building a home workstation soon, which will probably be used for personal work and occasional freelance visualization.

 

I would really appreciate if you could point me to real life end results and/or rendering setups (as far as proxies etc used) for me to get an idea of how much VRAM is utilized for each scene, and what for example is roughly "available" to a 1GB-1.5GB-2GB-3GB GPU user. I am already aware of the speed advantages, but I am more concerned with the "if it doesn't fit, it won't render" portion of GPU acceleration.

 

I am currently geared towards utilizing a GTX 580 3GB as my primary GPU, with the option for adding a second card in the future still open. I won't be making money on this machine initially, so I would really want to see the potential of GPU rendering without having to invest in Teslas and/or Quadros, which in their current fermi based versions are (in the long run) pretty much a GTX 580 with more VRAM.

 

I would also like to know if there is a utility that checks on current GPU VRAM utilization (edit: is it CPU-Z enough?), to know how much VRAM is actually available for the Renderer to take over, and maybe consider switching to a low-end GPU just for display puproses, while using the GTX as dedicated computation device (ala Tesla).

 

I am mostly interested in architectural visualizations, both interior and exterior, and if it wasn't for plants etc, I would not say that I use very complicated scenes.

I rarely render images bigger than 3-3500 x 2xxx, or equivalent for 200dpi printing on 11x17 or A3 sizes.

 

Hope it makes sense,

Thanks in advance for your input.

Edited April 10, 2012 by dtolios

[DavidRees]

I saw a calculation for this once, but can't find it now. In any case, nvidia said that Kepler cards will support unified memory, so your system ram will be added to the GPU calculation, although the 680 hasn't been tested this way, so may not support it. We'll see when Gk110 chip comes out this summer, I hope. I usually render ~6m polys with up to a gig of textures, so board memory is a big deal, and I can't take advantage of Vray GPU yet -looking fwd to unified mem/gk110 -hope it works!

[Dimitris Tolios]

Well, hoping to see Kepler receiving a (software/driver/re-compiler - w/e) boost in computation, as right now the only sample we have pales in comparison to the top Fermi offers. I would love the lower wattage (heat / noise and PSU requirement included, which is a big factor should you start adding 2 or 3 cards in a case), but i don't see any other advantage atm using a GTX 680, with the 2GB model being pricier and slower than the 3GB 580.

 

I don't know the price of the coming GK110 cards...should those be limited to Quadros costing 4digits a piece, would be a shame. I would still be looking into GTX 580s 3Gbs if that would be the case - hoping that those won't be phased out of the market completely by that time (some shortages are already there, with certain brands being out of stock completely and/or discontinued - usually those with the best air coolers like MSI - darn).

 

As far as textures, I think already VRay RT and iRay compress them to a 512px maximum size - part of the long initialization process for those renderers in GPU mode vs. CPU mode.

 

I doubt tho that nVidia will make "unified memory" or w/e available to gaming cards: unless it comes with a huge performance penalty, it would make Quadros and Teslas obsolete over GeForce cards with the same core, as the main selling point other than optimized drivers (irrelevant for anything but viewport "smoothness") was increased VRAM sizes.

[DavidRees]

http://www.xbitlabs.com/news/graphics/display/20120216125300_Nvidia_Kepler_Yields_Lower_Than_Expected_Chief_Executive_Officer.html

"...technologies that Nvidia promised to introduce in Kepler and Maxwell (the architecture that will succeed Kepler) include virtual memory space (which will allow CPUs and GPUs to use the "unified" virtual memory)..."