Recommended Hardware for RealityCapture:
Like most software applications, there is a list of system requirements for RealityCapture that can be used to help ensure the hardware in your system will work with their software. However, most "system requirements" lists tend to cover only the very basics of what hardware is needed to run the software, not what hardware will actually give the best performance. In addition, sometimes these lists can be outdated, list old hardware revisions, or simply outright show sub-optimal hardware.
Here at Puget Systems, we have taken the time to put together benchmarks for RealityCapture and run them across a wide variety of hardware. Based on this testing, we have come up with our own list of recommended hardware - as well as a specific workstation configuration tailored with these recommendations in mind.
Each step in the RealityCapture workflow utilizes the CPU differently, but overall we found a clear value winner in our testing:
- AMD Ryzen 9 5950X 16 Core - This processor offers the best performance in RealityCapture, ahead of the more expensive Threadripper models as well as everything Intel currently offers.
- AMD Ryzen 9 5900X 12 Core - For those who want to save a little money, this processor is only a tiny bit slower than its bigger sibling while costing a few hundred dollars less.
- AMD Threadripper 3960X 24 Core - For those who want more than the 128GB of memory that the Ryzen processors max-out at, moving to AMD's Threadripper series will allow for up to 256GB with only a small drop in performance.
- RealityCapture 1.1: AMD Ryzen 5000 Series Performance
- Intel Core vs AMD Ryzen and Threadripper Performance Comparison
- RealityCapture Benchmark
RealityCapture requires a NVIDIA graphics card for full operation because it uses CUDA for some of the key processing. Without that, you can technically run the program and perform some basic steps like registering images - but you cannot create a mesh / 3D model.
We found that there isn't a huge difference between modern mid-range and high-end video cards, but there is a little bit of a performance gain from spending more. However, using two video cards instead of one also provides a sizable benefit. As such, our recommendation is to use two video cards if your budget allows - and these are some of the sweet spots:
- GeForce RTX 3070 8GB - Only a little slower than the 3080, this model offers great value for those wanting to save some money.
- GeForce RTX 3080 10GB - This is our top single-card recommendation, with excellent performance in RealityCapture and a solid price. The RTX 3090 is a little faster, it is true, but also costs more than twice as much - not a worthwhile upgrade for most users.
- Dual RTX 3070 or 3080 cards - For a further boost in processing speeds, consider doubling up on video cards (increasing the power supply capacity is important to handle the added load)
- NVIDIA GeForce RTX 3070, 3080 & 3090 Performance
- NVIDIA GeForce and Titan Performance Comparison
- RealityCapture Benchmark
Memory requirements in RealityCapture are actually fairly modest, especially compared to other photogrammetry applications. This program was designed to use "out-of-core" algorithms, which do not require all of the data being worked on to be stored in the main system memory at same time. As such, very large projects can be handled with modest amounts of RAM - the developers even claim that 16GB is sufficient for image sets with thousands of photos. Here at Puget, though, we generally recommend 32GB as a minimum since Windows and other programs running in the background also need some memory space.
We strongly recommend using solid-state drives on all computers these days. They have a huge impact on every aspect of computer usage, from faster boot times to more responsive operation. RealityCapture will also load images more quickly from a fast drive, and because of its out-of-core design it also reads and writes data from the drive throughout processing. Image sets can take up a lot of space, so having a secondary hard drive for archival of projects and other data may be helpful.