GeForce GTX Rendering Benchmarks and Comparisons

Banner

Since a lot of you are asking about GPU rendering (and the NVIDIA GeForce cards have become so popular for rendering) we decided to compare the current high-end GeForce cards using a more realistic benchmark than the exterior car scenes which have become so commonplace.

When it comes to GPU rendering, it doesn’t get more real than Kevin Margo’s short film CONSTRUCT which was rendered on a BOXX GPU-dense workstation using V-Ray RT. Thanks to our friends at Chaos Group, we were able to get our hands on one of Kevin’s 3ds Max CONSTRUCT scene files for some testing. In this article, we’ll discuss the pros and cons of using high-end GeForce GTX cards for final frame production rendering.


Rendering Performance
Let’s get right down to it and talk about render times. The chart below compares the average render time of nine frames in an animated sequence. There were ten total frames in the test sequence and for each, the first frame was discarded because it was significantly faster than the subsequent frames. The GPU was cooler on the first frame, yielding a faster than average result. But as the card reached temperature, there were consistent render times going forward.
renderGraph

The GTX Titan Black became the champion in our tests, besting the GTX 980 and 970. Given the new Maxwell GPU compute architecture available in the NVIDIA 900 series, we expected the GTX 980 to put up a stronger fight, however, the 980 and 970 still posted respectable times. Below is a wireframe of the scene used in the tests.

Note: It is still a bit early to pass full judgement on the performance the 900 series Maxwell based GTX cards. Future updates to V-Ray as well new NVIDIA drivers could significantly improve the results above. 

screengrab


What GPU specifications should you look for?
For GPU rendering, you want as many CUDA cores as your budget allows. With that understanding, the Titan Black continues to lead the GeForce lineup by a wide margin. If you compare the chart above to the one located below, it’s easy to appreciate the relationship between CUDA cores and rendering performance.

cudaCores

Despite the importance of CUDA cores, they aren’t the only spec you should consider. Because all of your scene’s assets (including the scene file, textures, etc.) need to fit within the GPU’s memory, you need to make sure that the GPU you choose can accommodate your dataset. For most users this isn’t an issue, especially if you opt for the Titan Black with 6GB of memory. But for some, the 4GB memory limit of the 980 and 970 may be too constricting.


Which is most cost effective?
Based on the chart above, we know that the amount of CUDA cores on board has the greatest influence on render times. Therefore, if you want the most compute power for your money, you should think about how much you are paying per CUDA core. The GTX 970 at just 21 cents/CUDA is the best overall value if you on a strict cost consideration.

That said, many artists want a higher compute capability than a single GTX 970 has to offer. If your goal is compute density (packing as many CUDA cores in a machine as possible) then cards like the 980 and Titan become more favorable.

price3


Pros and Cons

GTX Titan
Pro – Most CUDA cores and large memory footprint.
Cons – Most expensive GTX card. Higher upfront cost. Highest price/CUDA core.


GTX 980
Pro – Respectable amount of CUDA cores ( 2048). Lower upfront cost.
Cons – High price/CUDA core. Lower memory footprint.


GTX 970
Pro – Low upfront cost. Cheapest price/CUDA core.
Cons – Lower memory footprint. Low compute density.


Choosing a the right workstation

If you’re ready to dive into GPU rendering, you won’t find a better workstation solution than the BOXX APEXX 4 which supports up to four full-size GPUs. To learn more, watch the video below.

If you wish to speak with a BOXX performance specialist, please complete the form below.

10 thoughts on “GeForce GTX Rendering Benchmarks and Comparisons

  1. This is great info as far as absolute performance goes, but the $/CUDA core metric is not relative, at least not between two different architectures.
    What is relative, is actual performance that each $ buys you, not how many cores.
    And if you do a similar graph on how much $ per gained second you have to spend going either GTX980 or GTX Titan Black over a GTX970, I believe the results would be quite the opposite of what you project.
    The value of the GTX970 proposition is unmatched. If you want more speed than what a 970 offers, you can always add more than one card. Needless to say that a 970 duo will humiliate a GTX Titan, while being considerably cheaper, while even a 970 Trio is within reach for everyone that can afford a GTX Titan and has the PCIe slots to spare.
    Note that 2x GTX970 will actually consume marginally more power than a GTX Titan Black, so the compromises made are very few going this way.
    Now, if what you do requires more than 4GB of VRAM, or you need the maximum compute performance in a single box for VRay RT , a set of GTX Titan or 780Ti cards could be your best bet, but even mATX based BOXX workstations would be faster and cheaper with 2x 970s vs. a single GTX Titan.

    • This is exactly what I was looking for but regarding the footnote on your table – can we actually assume that? Has anyone been able to test this?

      • Previous generations of GTX cards (5xx / 6xx / 7xx etc) were scaling close to 100% with Vray RT GPU (e.g. if one card would do 3:00min for a frame, 3x cards would do 1:00min and 6x around 30 sec), so it is safe to assume the same for the 9xx series.

  2. I don’t get the result you have posted for the “price per Cuda”.
    a basic math and you get a whole different results:

    1.GTX Titan Black- $1000/ 2880 = $0.34 per Cuda
    2.GTX 980- $600/ 2048 = $0.29 per Cuda
    3.GTX 970- $350/ 1664 = $0.21 per Cuda

    GTX970 is the best value and has the lowest price per Cuda core

    • Hi Yondon,
      Yes you are right. I’m not sure how I mangled the numbers on the price/CUDA core chart. I think it had to do with trying to combine it all into one chart. I’ve made the corrections. The performance data is unchanged and still valid. Thanks for the comment.

  3. What version of driver are you using?
    I use the GTX 970 but vray always stops working
    I am using the latest nvidia drivers

  4. Id like to add that all your cores dont have to be on one card… or even the same generation.

    I have a 680 and a 970 in one machine and they both work in tandem perfectly fine together.
    Having said that though, each square area of render, is done faster on the 970
    WHen it starts you dont know which card is doing what, but soon youll see the faster box, then move on to the next
    and then the 680 will finish a few seconds before the 970 finishes its second one.

    So… a card you have laying around somwhere is an EASY way to add some extra cuda power to ANY machine
    that has the PCI slot for it.

    Ive also set up that 680 as a dedicated PhysX card, and i do NOT put them in SLI.

    If im not using blender, or playing a game that uses PhysX, the 680 is pretty much OFF.

Leave a Reply

Your email address will not be published. Required fields are marked *