RealityCapture 1.1 – NVIDIA GeForce RTX 30 Series Multi-GPU Performance

Introduction

RealityCapture is a popular photogrammetry application which utilizes CUDA-compatible video cards for some processing steps within its workflow. In past testing, we found that RealityCapture benefited somewhat from adding a second video card. Is that still the case with the latest NVIDIA GeForce RTX 30 Series cards, and the recent update to RealityCapture 1.1 Blaze? Read on to find out!

Test Setup

Listed below are the specifications of the system we used for this round of RealityCapture testing:

We used AMD's Ryzen 9 5950X because we have found it to be one of the fastest processors for RealityCapture (alongside the other Ryzen 5000 Series chips). However, that does mean that we are practically limited to two GPUs. Some boards for this processor might be able to cram more cards in, but given the PCI-Express layout supported by this family of CPUs it wouldn't really be ideal to go beyond two… and stepping up to a bigger platform like Threadripper in order to support more video cards would add a lot of cost without a benefit in terms of CPU performance (though they do support more system memory).

To test these video cards, we used our in-house RealityCapture benchmark – both the standard and extended versions. Each was run twice per GPU configuration, and the best results were included in the charts below. Here is an overview of those projects:

• Rock Model – 45 photos at 20 megapixels each
• School Map – 51 photos at 18 megapixels each
• School Model – 278 photos at 18 megapixels each
• Park Map – 758 photos at 18 megapixels each

We make these benchmarks publicly available under the Creative Commons BY-ND license, so if you use RealityCapture you are welcome to download them from our website and compare your system's performance with the results below.

Benchmark Results

Here are charts showing the performance of single GeForce RTX 30 Series cards (in blue) versus pairs of the same cards (green):

For a detailed look at how adding a second GPU impacted performance in different processing steps, here is a table of the full results:

Analysis & Conclusion

Our testing of RealityCapture demonstrates a clear benefit to running two video cards, seen across all four of our image sets! Overall, the performance with two video cards ends up being 5 to 15% faster than a single card – with larger projects seeing the most benefit.

The full results tables shows that this increase in performance comes primarily from the Calculate Model step, though there are minor boosts to Colorize and Texture processing speeds as well. Align Images and the Ortho Projection step on maps appear to be unaffected by the video card at all.

What is even more interesting, though, is the comparison between two of the more affordable cards and a single costlier model. The most clear example of this is a pair of GeForce RTX 3070s (MSRP $499 each = $998 for two) versus a single RTX 3090 (MSRP $1499). The pair of less expensive cards was faster in all four of our image sets, and costs ~$500 less! Up to 5% faster results can be achieved by moving up to dual RTX 3080 cards, but the jump from 3080s to 3090s is always less than 2% – so a pair of those is probably not worth the added price for most users.

Are Dual GeForce RTX 30 Series Cards Good for RealityCapture?

Yes, the GeForce RTX 3070, 3080, and 3090 all showed a 5-15% boost from running as a pair versus a single GPU in RealityCapture! In fact, dual RTX 3070s outperformed a single RTX 3090 that costs substantially more. A pair of RTX 3080 cards is even faster, enough to be viable for some users, but the benefit from moving to dual 3090s is so small that it wouldn't be worth the added price in most cases.

As always, please keep in mind that these results are strictly for photogrammetry in RealityCapture. If you have performance concerns for other applications in your workflow, we highly recommend checking out our Hardware Articles (you can filter by software or other categories) for the latest information on how a wide range of programs perform with various GPUs, CPUs, and other hardware.