Forum

Author Topic: Benchmarking a GPUs  (Read 117140 times)

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #60 on: February 05, 2013, 02:58:50 PM »
Hello Matt, some have send, but noone send me more of them..

its clear that as for now the 7970 are the fastest cards here, but my recomendation go to have at least one Nvidia 5xx or 6xx series in PC, because problems with VBO in ATI drivers, so cannot work with bigger objects ( over 20mil is on 7970 end, on 560Ti can go up cca 50 mil trias with realtime -- 50 mil trias take aprox 950 MB or GPU ram)

Try test the new beta version....... 50% faster that 0.9.0 version
----------------
www.mhb.sk

Matt

  • Full Member
  • ***
  • Posts: 107
    • View Profile
Re: Benchmarking a GPUs
« Reply #61 on: February 05, 2013, 03:28:39 PM »
Thanks, i thought as much.

I am in full production mode right now so will try the beta after this lot of work is out the door.
I have some photos to contribute for a Orthophoto benchmark that I can get to you in a week or so.

Alterco

  • Newbie
  • *
  • Posts: 26
    • View Profile
Re: Benchmarking a GPUs
« Reply #62 on: February 13, 2013, 04:17:44 PM »
Hi guys,

I have a question for you, GPU experts.

I am looking to buy a computer.
According to the Agisoft wiki, it is recommended to use NVidia GeForce GTX 580 or GeForce GTX 680 GPU.
So, I have chosen the GTX 680 GPU as reference.

After some documentation research, it appears that a SLI using 2 GTX 660 is more powerful than a single GTX680 for less money.

Then, I wonder why not use a 2 GTX 660 instead a GTX 680 with Photoscan.

Different thoughts :
- 2 GTX 660 are less expensive than one GTX 680
- the SLI appears more powerful
- 2 GTX 660 can provide 2 GPU instead of one with the GTX 680

Considering that the future SFM algorithm appear to use more and more the GPU computation, it seems interesting to multiply the GPU core number.

Finally, I can not find any arguments to not use a 2 GTX660 SLI.
Maybe just one, does/will Photoscan support this configuration ?

So, what do you think ?

Thank you in advance for you expertise.

Adrien

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #63 on: February 13, 2013, 04:28:15 PM »
SLI not work with OpenCL  ( it should but its still not iplemented ! and propably will be not for a while...  ). When thinking to use GPUs, thing on that every GPU need a 1 CPU core, so its better to have one card with better spec. Even better at least for Pscan is to use ATI solution, its faster, almost double the performance against NVIDIA cards...... In short, its  Pscan speciality of using raw performance of the ATI cards...... they are dubble so fast in single precision aritmetics as Nvidia...

----------------
www.mhb.sk

Alterco

  • Newbie
  • *
  • Posts: 26
    • View Profile
Re: Benchmarking a GPUs
« Reply #64 on: February 13, 2013, 06:59:15 PM »
Hi,

Thank you for your quick answer.
It is too bad that this promising configuration could not work properly with Photoscan...
It could have been interesting for other work too...

But I will keep the GTX680 instead the faster ATI solution because it seems that there is some issue with big scenes. Am I wrong ?

Regards,

Adrien

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #65 on: February 13, 2013, 09:36:33 PM »
Yes.... Stick with 680 and its OK.......
----------------
www.mhb.sk

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #66 on: February 13, 2013, 10:08:12 PM »
Hmm as from today test with new 13.1 drivers, it seems that ATI drivers are repaired and the VBO is working again.... So a large models can be opened and edited fast like on Nvidia cards....
----------------
www.mhb.sk

Matt

  • Full Member
  • ***
  • Posts: 107
    • View Profile
Re: Benchmarking a GPUs
« Reply #67 on: February 19, 2013, 01:57:13 PM »
We have a new toy to play with http://www.asus.com/ROG/ARES26GD5/

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #68 on: February 19, 2013, 02:25:07 PM »
hmmmmm how high will overclock it ? im think over 1200-1250 is possible....
----------------
www.mhb.sk

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #69 on: February 21, 2013, 08:04:49 PM »
Hmm early tests of Geforce Titan http://www.tomshardware.com/reviews/geforce-gtx-titan-performance-review,3442-10.html in OpenCl mode.......

It seems that will be disappointing speed in Pscan too.....
----------------
www.mhb.sk

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #70 on: February 21, 2013, 08:16:10 PM »
We did bring these issues up with Nvidia, and were told that they all stem from its driver. Fortunately, that means we should see fixes soon." I suspect their fix will be "Use CUDA".

Nvidia has really dropped the ball on OpenCL. They don't support OpenCL 1.2, they make it difficult to find all their OpenCL examples. Their link for OpenCL is not easy to find. However their OpenCL 1.1 driver is quite good for Fermi and for the 680 and 690 despite what people say. But if the Titan has troubles it looks like they will be giving up on the driver now as well or purposely crippling it (I can't imagine they did not think to test some OpenCL benchmarks which every review site uses). Nvidia does not care about OpenCL Nvidia users like myself anymore. I wish there more people influential like Linus Torvalds that told Nvidia where to go.
----------------
www.mhb.sk

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #71 on: February 22, 2013, 06:18:05 PM »
hmm interesant bench together

http://www.computerbase.de/artikel/grafikkarten/2013/test-nvidia-geforce-gtx-titan/13/

Yes that are common task, Pscan use very specific code that shine on ATI cards, OpenCl is better supported on AMDs cards. im think that do Nvida because they want support CUDA solution. CUDA is more "powerful" easier to code and do a lot of scientific code.......  But OpenCL is independed on hw that we use.......
----------------
www.mhb.sk

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
----------------
www.mhb.sk

Andrew

  • Jr. Member
  • **
  • Posts: 77
    • View Profile
Re: Benchmarking a GPUs
« Reply #73 on: April 21, 2013, 12:40:57 AM »
Wishgranter, you mentioned that latest ATI drivers seem to have fixed the VBO issues. Do you know if it is the case for all Windows versions (Win7, Win8 as well as Windows 2012 Server)?

It was also mentioned that Photoscan performs better (OpenCL-wise) on Windows 2012 Server than on Win7. Is it still the case using these latest drivers? What about regular Windows 8, is it as fast as 2012 Server or as slow as Win7?

Cheers,
Andrew

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Benchmarking a GPUs
« Reply #74 on: April 21, 2013, 10:21:29 AM »
Yes, VBO drivers are fixed.....
Yup, W2012server performnace it was best to see with Nvidia cards, 30 % difference in perfrocmannceon my 560Ti, but as im have read its posible that other cards are faster under W2012Server. Im have not so many time to test all configs under all win versions.. but take this as good hint, its for longer time why is this possible with W2012Server...
----------------
www.mhb.sk