3D object generation software "Agisoft PhotoScanThe number of inquiries regarding "is increasing day by day. This time, using PhotoScan (Ver1.3.2), we will perform a comparative test using the new CPU and GPU of Intel and AMD, and consider the points of hardware selection.
Measuring method, tools, etc.
・ Agisoft PhotoScan version: Ver1.3.2 (ver. At the time of evaluation) / Ver1.2.4 for comparison
-The standard settings are CUDA enabled and whether or not the GPU is enabled on the software.
・ Sample project: Doll (Agisoft sample data download page)
After starting PhotoScan in the above environment,
Align Photos (High) → Build Dense Cloud (Medium) → Build mesh (Medium)
Execute in the order of, check the execution time in the output log, and compare.
Verification environment 1: Intel Core i9
※reference: TEGSTAR Premium
inspection result
Conclusion
It was confirmed that the update from the old version 1.2.4 to version 1.3.2 significantly improved the processing speed. It is thought that this is because the CPU has been used successfully on the software side and parallel processing has come to be performed.
In addition, the benefits of GPU processing have been fully obtained. In the Align Photo processing, the processing speed was improved by about 1080 times when the GTX3 was installed, compared to the case without the GPU. Although the difference between the specifications of 1080 and 1080Ti was not so large, the efficiency was improved by about 2 to 3% when two pieces were inserted. Considering the total processing time of a series of operations, adding a GPU is also an effective option.
Based on the above results, it may be possible to see a further parallel processing effect by using an AMD CPU that is more specialized in the number of cores. Verification environment 2). The above Intel Core i9 has 10 cores and 20 threads, while AMD Threadripper has 16 cores and 32 threads.
Verification environment 2: AMD Threadripper
inspection result
Conclusion
As a result of the actual measurement, the expected effect was not obtained, and conversely, turning off SMT (CPU multi-thread function) resulted in better results in this case.
The "reference" below is the CPU usage at the time of measurement, but the Intel Core i9 is evenly loaded on the CPU, whereas the AMD Threadripper is in a situation where the load is not full (use efficiency is poor) I will. Since the improvement can be seen by turning off SMT, it is optimized for the Intel CPU on the software side, and it can be said that the version of PhotoScan verified this time can not use AMD's CPU multithread function. That's right.
reference
CPU usage during processing in environment 1 (Intel Core i9)
CPU usage during processing in environment 2 (AMD Threadripper SMT standard)
* The condition where the load is not applied continues.
CPU usage during processing in environment 2 (AMD Threadripper SMT OFF)
Overall conclusion of this verification
In PhotoScan (1.3.2) used for the verification this time, we were able to confirm that the parallel processing and processing using the GPU, which were issues in the previous version, are working properly. In later versions including this version, it seems that work efficiency can be further improved by choosing the current Intel CPU or multiple GPUs.
It was thought that PhotoScan (1.3.2) was optimized for Intel CPUs, but it is possible that AMD CPUs will be considered in future updates. I will. It may be okay to select the AMD CPU even if you look at future version updates.
We plan to release information about differences in PhotoScan versions, patterns with increased physical CPUs such as Xeon, effects due to GPU specifications, so please look forward to it.
At Tegsys, we have introduced a number of proposals for PCs that assume the use of Agisoft PhotoScan.
Click here for the list of machine case studies for Agisoft PhotoScan |