Hardware comparison for Agisoft PhotoScan

This page shows a comparison of several CPU types for use with Agisoft PhotoScan.

Choosing a system configuration

The four processing stages:

  • Align Photos (CPU)
  • Build Dense Cloud (CPU, GPU)
  • Build Mesh (CPU)
  • Build Texture (CPU)

Only the "Build Dense Cloud" step is able to utilize the GPU. However, that one step takes longer than all the others combined.

Optimal CPU-corecount / GPU balance

For a good price-performance ratio use a mimimum of 6-8 CPU cores, add one core for every additional GPU. For Multi-GPU one core must be reserved per GPU. While multi-core CPU and multi-GPU efficiency decreases for every added CPU-core/GPU, multi-core CPU efficiency increases when addig more GPUs:

  • multi-core CPU efficiency 50%-92% (depending on step and data)
  • multi-core CPU efficiency increases by ~5% per GPU
  • with more GPUs, add more CPU cores
  • disable one CPU core per physical GPU in the system
  • second GPU increases performance by 30-40%
  • third GPU increases performance by another 20-25%
  • fourth GPU increases performance by another 12-15%
  • for GPUs, more is better (2x GTX 970 is better than 1x GTX 980)

Other performance factors

  • operating system
  • network processing (cluster)

CPU comparison

The last 3 columns show the buildtimes in seconds (total and stages 1 and 2) from anandtech 2017 CPU tests win10 - photoscan 1.3.3 large.

cpu socket mem max-mem cores frequency turbo cost total s1 s2
i9-7980XE 2066 ddr4-2666 128 18 2.6 4.2 2049 2081 344 1292
i9-7960X 2066 ddr4-2666 128 16 2.8 4.2 1699 2026 343 1269
i9-7940X 2066 ddr4-2666 128 14 3.1 4.3 1444 ? ? ?
i9-7920X 2066 ddr4-2666 128 12 2.9 4.3 1189 ? ? ?
i9-7900X 2066 ddr4-2666 128 10 3.3 4.3 999 ? ? ?
TR-1950X TR4 ddr4-2666 1024 16 3.4 4.0 999 2374 333 1507
i7-8700K 1151 ddr4-2666 64 6 3.7 4.7 399 2477 598 1537
i7-7820X 2066 ddr4-2666 128 8 3.6 4.3 609 2515 508 1608
TR-1920X TR4 ddr4-2666 1024 12 3.5 4.0 779 2560 377 1647
R7-1800X AM4 ddr4-2666 64 8 3.6 4.0 374 2769 510 1809
R7-1700X AM4 ddr4-2666 64 8 3.4 3.8 334 2769 537 1882
i7-6850K 2011v3 ddr4-2400 128 6 3.6 3.8 399 3261 808 2020
i7-4960X 2011 ddr3-1333 64 6 3.6 4.0 ? ? ? ?

Other system components

component type amount capacity total (GB) cost
board MSI X299 LGA 2066 8 ddr4-2666 128 352
board MSI X99A LGA 2011v3 8 ddr4-2666 128 391
board MSI X399 TR4 8 ddr4-2133 128 389
board MSI X370 AM4 4 ddr4-2133 64 175
board MSI Z270 LGA 1151 4 ddr4-2400 64 185
GPU GeForce GTX 1080 Ti 1 gddr5x 11 1200
GPU GeForce GTX 1070 Ti 1 gddr5x 8 750
GPU GeForce GTX 1060 1 gddr5x 6 450
memory 16G 8 ddr4-2666 128 1600
memory 16G 4 ddr4-2666 64 800
case 1 midi tower 150
psu corsair RM 1 550 W 105
psu corsair RM 1 650 W 115
psu corsair RM 1 750 W 130
psu corsair RM 1 850 W 180
psu corsair RM 1 1000 W 190
psu corsair HX 1 1200 W 230
ssd m2 2280 1 512 240
hd sata 1 2000 100

Load requirements [W]

Add ~50W for PSU selection.

cpu 1x1080 2x1080 3x1080 1x1070 2x1070 3x1070 1x1060 2x1060 3x1060
i9-7980XE 609 800 1045 488 663 839 427 545 663
i7-7820X 534 775 1019 463 638 814 401 520 637
i7-6850K 534 775 1019 463 638 814 401 520 637
i7-8700K 484 725 969 417 588 764 355 470 587
TR-1950X 575 815 1060 503 678 855 442 561 678
R7-1800X 488 729 974 417 592 768 355 474 592

Example system configurations

All costs are estimated using consumer prices (Mar 2018).

Base hardware:

  • case, ssd, hd: 500
  • psu: 110 - 230
  • board (2066, 2011, TR4): 350 - 390
  • board (AM4, 1151): 175 - 185

Configurations:

# vendor cpu gpu(s) mem max psu [W] [CHF]
1 intel i7-8700K 1x 1080 64 64 550 3190
2 amd R7-1800X 1x 1080 64 64 550 3180
3 intel i7-6850K 1x 1080 64 128 650 3410
4 amd R7-1800X 2x 1070 64 64 650 3470
5 amd R7-1800X 2x 1080 64 64 850 4430
6 intel i7-7820X 2x 1070 64 128 750 3900
7 intel i7-7820X 2x 1080 64 128 850 4850
8 intel i9-7960X 3x 1070 64 128 1000 5790
9 intel i9-7960X 3x 1080 64 128 1200 7180
10 amd TR-1950X 3x 1070 64 128 1000 5130
11 amd TR-1950X 3x 1080 64 128 1200 6520

Tests at ISG (Windows vs. Linux)

We made some tests to figure out what operating system performs better for agisoft photoscan. The used scene, test settings and system configurations are listed below.

  • Version: 1.4.1
  • Test scene: monuments
  • Batch jobs: Align Photos, Build Dense Cloud, Build Mesh, Build Texture

Test system 1

  • Board: ASUS X99 Deluxe
  • CPU: i7-5820K 6x3.3 GHz @ 4.7 GHz (over-clocked)
  • GPU: ASUS GeForce GTX 980 Ti STRIX
  • Memory: 4x4 GB DDR4 @ 2400 MHz
  • SSD: Samsung 960 Pro 2 TB
os quality (dense cloud) total s1 s2 s3 s4
Ubuntu 17.10 Highest (Ultra High) 2005 35 1670 111 189
Ubuntu 17.10 Default (High) 485 31 290 60 105
Ubuntu 17.10 Default (Medium) 240 30 72 46 91
Windows 8.1 Highest (Ultra High) 1612 28 1266 119 199
Windows 8.1 Default (High) 480 22 306 54 97
Windows 8.1 Default (Medium) 244 30 86 42 86

Test results are similar for medium to high settings. Using ultra high settings, windows takes the lead. It is unclear however if the highly over-clocked system is the reason for this difference, as it was initially optimized for running Windows.

Test system 2

The Institute of Geodesy and Photogrammetry uses photoscan and was in need of additional computing power for their research projects. We helped them out configuring a system for that purpose, which led to this blog post. The result was a dedicated workstation with the specs listed below. We were able to get our hands on this shiny piece of hardware for a few days and ran the same test again using Linux and Windows. The settings were left all at defaults, except dense cloud set to high.

  • System: DALCO X299 Workstation black (Tower)
  • CPU: Intel i9-7960X 16x2.8 GHz
  • GPU: 2x nVidia GTX 1080Ti 11G (GPU PCIe Speed: x16/x16)
  • Memory: 64GB DDR4 2666MHz
  • SSD: Samsung 960 Evo 500GB M.2 2280

Below are the test results comparing Windows and Linux. This test was finished 70% faster on Linux compared to Windows:

os quality (dense cloud) total s1 s2 s3 s4
Ubuntu 18.04 Default (High) 257 24 87 107 40
Windows 10 Default (High) 365 20 185 97 63

I made some addtional tests on Linux (Ubuntu 18.04) using the same settings as above to compare the performance between version 1.3 and 1.4 and using OpenCL:

os quality (dense cloud) version total s1 s2 s3 s4
Ubuntu 18.04 Default (High) 1.3 251 19 97 103 30
Ubuntu 18.04 Default (High) 1.3 OpenCL 233 19 86 98 30
Ubuntu 18.04 Default (High) 1.4 294 35 72 124 61

Conclusion

The performance comparison of Agisoft Photoscan running on Linux vs. running on Windows is inconclusive from these tests. The results may highly depend on the hardware specs, the used settings and the scene to render. But choosing the right operating system may very well result in noticeably shorter render times. It is probably a good idea to make some tests with the actual hardware and if possible using a real scene as test data to get some useful representative results.

An important side note regarding GPU (graphics processing unit): Using consumer grade Geforce Graphics Cards, as listed above, will not allow to use Remote Desktop to run Agisoft Photoscan. If you want to use Remote Desktop to run Agisoft Photoscan, you need enterprise grade Geforce Quadro cards. Of course this comes with its price.

References: