Forum

Author Topic: CPU and GPU benchmarks  (Read 165045 times)

MeHoo

  • Full Member
  • ***
  • Posts: 136
    • View Profile
Re: CPU and GPU benchmarks
« Reply #45 on: January 17, 2015, 04:50:23 AM »
CUDA cards don't show any improvement over basic openCL cards like the GTX's.  I am using a quadro k4000 and two 580 GTX's and the 580's are exactly double the performance of the quadro.  Since, as Alexey mentioned to me in an email recently:
Quote
As for the graphic cards, professional class cards from Quadro series are optimized mostly for CAD applications and double-precision calculations that are not actually used by PhotoScan. So high-end gamer cards show much better performance due to higher number of GPU cores and frequency. Tesla cards will show similar performance to the recommended gamer-class GPUs, but will cost much more.

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #46 on: February 09, 2015, 02:46:41 PM »
Any test using dual Xeons yet?  Considering a build with 2x Intel Xeon E5-2660 v3.  Anyone know if Agisoft will benefit greatly from the 40 hyper threaded cores?  Or should i just save a ton of money and go for a  i7-5960X? 

Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: CPU and GPU benchmarks
« Reply #47 on: February 09, 2015, 02:53:31 PM »
Hello Igor,

For dual-Xeon systems I can suggest to use 6-8 core Xeons with highest possible frequency. Using 40 cores with 2.2 GHz could be even slower than desktop six-core i7.

We haven't yet made direct comparison between 5960X and 5930K, but assume that second option may be faster due to higher CPU frequency, even the number of cores is six versus eight.
Best regards,
Alexey Pasumansky,
Agisoft LLC

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #48 on: February 09, 2015, 06:39:19 PM »
So in other words its a complete waste of money going with dual Xeons for Agisoft?  A single CPU with maximum available clock speed is the best option?  In that case the clear winner should be the i7 5930 with 6 cores and 3.5GHZ.  Intel Xeon E5-1650 v3 is 6 core at 3.5GHZ and 15mb Cash should also be a contender?  Throwing in 2  E5-1650 v3 will not increase performance significantly? 


Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: CPU and GPU benchmarks
« Reply #49 on: February 09, 2015, 07:09:28 PM »
Hello Igor,

Dual-Xeon with high CPU frequency will be faster than single CPU, but as I've already said, it's better to use 6-8 core CPUs. Also Xeon-based configuration will allow you to install more than 64 GB RAM, if it is required by your tasks.

For example, dual Intel Xeon E5-2667v2 configuration was about 20-40% faster than single E5-2667v2 on Align Photos stage on our test aerial datasets.
Best regards,
Alexey Pasumansky,
Agisoft LLC

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #50 on: February 09, 2015, 08:30:30 PM »
Ok that clarifies it.  Thank´s  for the explanation. 

holgar

  • Jr. Member
  • **
  • Posts: 62
    • View Profile
Re: CPU and GPU benchmarks
« Reply #51 on: February 10, 2015, 03:42:28 PM »
I have just tested my new MacPro with the Building-Benchmark Scene:

New MacPro, 6-Core, 64GB Ram, Dual D700


Dense cloud Medium-setting 0/12 agressive:

Device 1 performance: 667.114 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Device 2 performance: 693.546 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Total performance: 1360.66 million samples/sec

Finished processing in 82.2601 sec (exit code 1)


Dense Cloud High-setting:

Device 1 performance: 819.165 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Device 2 performance: 834.194 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Total performance: 1653.36 million samples/sec

Finished processing in 434.088 sec (exit code 1)


Dense cloud Ultra-setting 0/12 mild:


Device 1 performance: 804.472 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Device 2 performance: 832.658 million samples/sec (ATI Radeon HD - FirePro D700 Compute Engine)
Total performance: 1637.13 million samples/sec

Finished processing in 3347.54 sec (exit code 1)

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #52 on: February 10, 2015, 05:30:30 PM »
Alexey,  you said dual CPU will improve Alignment stage with 20-40%.  How About build dense cloud and build mesh stages.  How much roughly if any improvement should i expect? 


Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: CPU and GPU benchmarks
« Reply #53 on: February 10, 2015, 05:37:07 PM »
Hello Igor,

During our tests (note that we haven't tested a lot of dual-Xeon configurations on various projects) for dense cloud generation (except depth maps estimation) we have about 30% speed up for 2x Xeons and mesh generation stage speed up was only about 10-15%.

But I assume that values may vary depending on the CPU, scene type and even OS.


You can check the following topic, though it is quite old: http://www.agisoft.com/forum/index.php?topic=1330.0
Best regards,
Alexey Pasumansky,
Agisoft LLC

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Re: CPU and GPU benchmarks
« Reply #54 on: February 11, 2015, 02:32:36 PM »
Anandtech is using Photoscan to benchmark CPUs so you can pretty much find all your answers there.
Check out the recent 14core Haswell Xeons for instance: http://anandtech.com/show/8730/intel-haswellep-xeon-14-core-review-e52695-v3-and-e52697-v3/2

For the record, all of their Agisoft benchmarks were most probably recorded using pre 1.1 version, but I don't think there should be any game-changing differences between 1.0.x  and 1.1.x benchamrks.

-Andrew

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #55 on: February 11, 2015, 03:15:17 PM »
Thanks for the link. .Not very impressive results for the Xeons given the high price  Xeon ES2697 is  slightly slower when running it as dual CPU incredibly enough according to the test.

 Looks like the CPU i already have the i7 4930 is actually a very smart choice.  The top 12 core Xeon ES2690  is around 14% faster but is 10x more expensive.  Glad i asked on the forum  before bying as it seems investing in expensive Xeons could have been a huge loss of money for nothing. I use a lot of other software though so have to consider this to. 

 Disappointing though that there seems to be no way of getting a descent speed increase  in Agisoft no matter what CPU you use. Aligning 1500 36 mpix images on my I7 4930 high setting  right now and the processing time is 55 hours.... Investing 3000 USD in the top Xeon cpu and the align might be finished in 47 hours instead of 55.   That is not nearly enough to justify the investment. 

I will have to go Xeon for the RAM though.  So trying to figure out which one.  ES1650 6core runs at 3.5GHZ so should be fast, no tested though?  Probably should just buy a single CPU and not double?   


Alexey Pasumansky

  • Agisoft Technical Support
  • Hero Member
  • *****
  • Posts: 14813
    • View Profile
Re: CPU and GPU benchmarks
« Reply #56 on: February 11, 2015, 03:20:01 PM »
Hello Igor,

55 hours for 1500 is quite a lot. Are you using any preselection? Also using lower values for key-point and tie-point limits will results in faster processing.
Best regards,
Alexey Pasumansky,
Agisoft LLC

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #57 on: February 11, 2015, 03:28:24 PM »
Thanks for asking!  Im using these settings


High setting - Diabled (not generic)
1000 tie point limit. 
40 000 point limit
No image pair pre selection.

55 hours is a lot but if that is what it takes i have to wait :-)  That is the projected time, usually it seems to get finished faster though but still 45 hours plus i would think.  If you have suggestions on how to speed this up with out losing accuracy I'm very interested though. 



dtmcnamara

  • Jr. Member
  • **
  • Posts: 73
    • View Profile
Re: CPU and GPU benchmarks
« Reply #58 on: February 11, 2015, 03:37:11 PM »
Glad i asked on the forum  before bying as it seems investing in expensive Xeons could have been a huge loss of money for nothing.

Not all Xeon CPUs are that expensive,  take the E5-2620v3 that I have. Performance wise, it is about 10% less powerful as your 4930K, and it only cost $339 new. Now while the i7 is faster on single core processes, the Xeon can a max of 768GB and almost every motherboard out there that supports Xeon CPUs will do 4 x16 GPUs.

Also one other thing that a lot of people might not know. Reducing memory hard faults will speed up your processing times. What is a hard fault you might ask? Well its the reason Agisoft recommends upwards of 700GB of RAM for some scans. Take a look at the link below for an explination.

http://www.brighthub.com/computing/windows-platform/articles/52249.aspx

igor73

  • Full Member
  • ***
  • Posts: 228
    • View Profile
Re: CPU and GPU benchmarks
« Reply #59 on: February 11, 2015, 04:02:04 PM »
Interesting with the hard faults.  I am running a huge image set now of 1500 images as i mentioned above.  Its been running for 12 hours now and 0 hard faults per second is reported by Windows.  CPU clock speed is 118% of normal and all cores at 99-100% so seems my system is running fine and the cooling is enough.  I have 64gB DDR3 and around 24 GB is used right now. 

I have 55 hours remaining to align so what i said before is wrong.  Total time for me to align is 67 hours not 55 for the 1500 images.   This is just one of 4 chunks of similar size.  The last chunk though:--)  All the others are already aligned, merged and dense cloud generated.  It has been a pretty  monumental processing task though .  That  is why I'm interested in more speed.