There’s a growing amount of information on video cards for rendering these days. That being said after spending some quality time with a Quadro 5000, Tesla 2050, and GTX 470 I thought I’d take a minute to share some of my findings/opinions on how these cards stack up in my own daily use with 3ds Max and iray. Perhaps this info will benefit anyone that’s trying to figure out which video card(s) to purchase.
I recently built a new workstation based on the EVGA x58 FTW3 motherboard. I chose it because it seemed to provide a decent amount of newer features like SATA 3.0, USB 3.0 and of course plenty of PCIe slots for multiple video cards. For the video card I chose an EVGA GTX 470 “gaming card”. I chose this card because it is a FERMI based card and at the time seemed to offer the best bang-for-the-buck as it was on sale.
A few months later I added an NVIDIA Quadro 5000 and Tesla 2050 to the mix. Now, the first thing to note is that a combination of GTX470, Quadro 5000, and Tesla 2050 is an odd combination. I say “odd combination” because the GTX 470 with it’s 1280mb of on-board memory can bottleneck the rendering process since the 3ds Max scenes (including texture maps) must fit into the memory footprint of EACH video card. The on-board memory of the video cards are NOT considered cumulative amounts. So, while the Quadro 5000 has 2.5GB of on-board memory and the Tesla 2050 has 3GB of on-board memory, the scenes can not exceed the 1280MB on the GTX 470 if I wish to use all three video cards at the same time.
As you can imagine since the scenes have to fit onto each GPU, the amount of GPU on-board memory is critical. The more memory your GPU has, the more complex your scenes can be. Of course you can render with iray and use on the CPU and system RAM…but it’s going to be painfully slow in comparison to using GPUs, especially the new FERMI based cards.
So we’ve established that on-board memory is critical but the number of CUDA cores and their speed is also critical. The more CUDA cores (Stream Processors) your GPU has, the better. The faster they are, the faster your scenes will render. However, on the flip side the faster they are the more heat they generate. Here’s a comparison of CUDA cores & clock speeds on these three cards:
GTX 470 = 448 CUDA cores (core clock: 1215 MHz)
Quadro 5000 = 352 CUDA cores (unsure of clock information)
Tesla 2050 = 448 CUDA cores (core clock: 1150 MHz)
As you can see the “Gaming card” has the same amount of CUDA cores as the Tesla and they are clocked at a higher speed. This is reflected in the performance as well. When rendering with iray the GTX is noticeably faster than the Quadro 5000 and a little faster than the Tesla 2050 (provided I can fit the scene onto the GTX card). The next consideration and what might be the most important to many of us is cost.
Here’s a cost break down between these three video cards:
GTX 470: $259.99
Quadro 5000: $1,779.99
Tesla 2050: $2,499.99
In looking at the prices and with the knowledge that the gaming card is faster to render a scene than the Quadro 5000 or the Tesla 2050 the obvious question then becomes: How can anyone justify the huge cost difference of the Tesla & Quadro cards vs. the 400 series FERMI based “gaming cards”?
Answer: The main consideration (for me) is memory limitations of the GTX type cards. The Quadro & Tesla series house more on-board memory therefore I can render larger/more complex scenes. Of course I do have to say that it’s still difficult to fit complex scenes on the 3GB of RAM of the Tesla 2050 but it’s a lot easier than trying to fit a scene onto a 1gb GPU like my GTX 470. On the upside you can find some GTX cards with up to 2gb of on-board memory so I’d definitely look into those over the 1gb models if you plan on rendering with the card.
The other issue to consider is heat. I can tell you that my GTX 470 generates WAAAAY more heat when rendering than the Tesla 2050 or Quadro 5000. I do worry about that damaging the GTX card as I’m sure it wasn’t designed for this kind of abuse. I could water cool it, but to be perfectly honest since these gaming cards are common, somewhat inexpensive, and updated/redesigned frequently I would probably replace it within a year or so anyway. So if it lasts that long I’m good to go. The Quadro & Tesla cards are built to last longer under the stress of rendering. So your investment in a Tesla or Quadro card should last longer in that aspect.
NOTE: Seriously, if you render with GTX/gaming cards please make sure you have ample cooling and/or the workstation isn’t near anything flamable! I’m not saying they will instantly burst into flames when you hit the “make pretty button” but these things get CRAZY hot when rendering for long periods so please be aware of that.
The final consideration for me is 3ds Max viewport performance. In this aspect the GTX 470 pales in comparison to either the Quadro 5000 or the Tesla 2050. Don’t get me wrong, my 3ds max viewport performance isn’t bad with my GTX 470 but it’s noticeably more responsive/smooth with either the Quadro or Tesla cards.
Bottom Line: Which GPU(s) should you purchase? All I can offer on that question is advice on what I’d do. I do like the Quadro 5000 and Tesla 2050 because they are quieter & run cooler than the GTX card and of course I can render larger scenes with the Quadro/Tesla cards. That being said if I were building a new machine I would probably start with 1 GTX (FERMI based) video card with 2gb of on-board memory depending on the reviews of those cards. If the reviews didn’t look good for those cards I would start with a 1.5gb FERMI based gaming card. Try that with 3ds Max to see if most of your scenes fit into that 2gb of on-board memory. If most of your scenes do render with a 1.5gb or 2gb GTX type card then buy one or two more & enjoy some very quick rendering (but lot’s of heat). Just don’t be surprised if the gaming cards don’t hold up under the extreme stress of long periods of rendering, unless you water cool them.
If most of your scenes don’t fit onto the 1.5/2.0 gb GPU’s THEN you should explore the Quadro/Tesla series…OR wait & hope for someone to develop a GTX based video card with a lot of on-board memory. The other option is a Tesla 2070 that has 6GB of on-board memory. It would be awesome to have two or three of those! However that configuration would be really expensive but it might pay for itself if you can turn projects around quickly with that configuration.
NOTE: You may also check my iray/GPU F.A.Q. HERE.
Example iray animations: