V-Ray RT GPU benchmark test

Like everyone else in the 3D world, we were cock-a-hoop when we found out that the V-Ray RT GPU beta was available for testing. Once we calmed down, we got down to some serious testing to find out how fast it can render and which graphics card is the best choice.

As a benchmark, we rendered the scene in the image below using the standard production renderer. For this test, we used an Intel i7 920 CPU, which took 25 minutes 36 seconds to complete.

We also repeated the render using slightly lower settings that are more like the type of settings you would use for a draft render. The result: the render completed in 9 minutes 36 seconds.

The scene is made up of 1 million polygons and uses a single V-Ray dome light with an HDR image to provide varied reflections. Every material is highly reflective, which traditionally means longer render times. We used fairly high render settings to get as close to the final look as possible but without taking too long.

The test

So that the tests were fair, we let each GPU render reach 32 samples per pixel, which I hope you'll agree provides a decent enough image to base a creative decision on.

Of course there is still noise in the image but, for the type of feedback we are looking for, it is perfect. From this image you can see if the materials, reflections and lights you have applied are correct. So, instead of waiting 25 minutes (or less if you reduce the render settings), you can see within a few seconds if you need to change anything.

Results

At this point, we'd like to remind you that this is the beta of the GPU version. So, while these results are accurate, we can't say they are definitive. Not yet anyway.

We've only tested a few graphics cards so far, but in the near future we hope to be able to give you a much more comprehensive list of benchmarked GPUs, so you can choose which is the most appropriate card for your needs and budget. The results can be seen in the chart below.

We found that even an entry level NVIDIA GeForce graphics card will give you adequate feedback after 38 seconds. That's 15 times faster than the draft render using the traditional method. The Tesla C2050 is even more impressive, clocking in at about 55 times faster!

The difference between them? CUDA cores. V-Ray RT uses the OpenCL API instead of CUDA, but you can accurately predict the speed of a NVIDIA GPU by looking at the number of CUDA cores. In the above instance, the GTS 250 has 128 cores whereas the Tesla has 448.

So which GPU is best?

Luckily for you, we have created a useful table that shows the range of GPUs we can offer and the all important information about each one.

You may have noticed that we have only included NVIDIA GPUs and not ATI cards. So far we haven't been able to gauge how the number of processor cores on the ATIs correlate to render time. However, it seems to be widely accepted that NVIDIA are currently able to process much more efficiently using OpenCL than ATI are.

A few things to bear in mind

If you have a big scene that requires 3 or 4 GB of RAM, then in order to load this scene onto the GPU, you will need to go for a GPU that has this amount of RAM on board - which makes sense. So, for huge scenes, the Tesla range looks to be the best choice.

If your scenes are more moderate in memory needs, your primary consideration needs to be the number of CUDA cores, as this will give you a faster render.

We found that when we used a single graphics card to process the render and also drive the video output, artefacts occur on screen as soon as we clicked render. For best results, we recommend a standard GPU for the displays and a secondary GPU to be used solely for the processing

The Tesla C2050 requires dual 6-pin power cables and takes up the space of two PCIe slots. It also requires 238 Watts so check, not only your current power consumption and your PSU, but also that you have an additional 6-pin power cable as we only got one when we opened the box!

If you have any questions at all about this or would like to arrange a demo with your scenes, please feel free to email us at 3D@Jigsaw24.com or call on 03332 409 309. Just remember that this is a beta version, so there may be a degree of onsite troubleshooting going on!