Forum

Author Topic: Revisiting Graphic Card OpenCL Benchmarks  (Read 6861 times)

MitchellT

  • Newbie
  • *
  • Posts: 16
    • View Profile
Revisiting Graphic Card OpenCL Benchmarks
« on: November 02, 2012, 03:32:44 PM »
While finally buying my custom Photoscan computer, the salesman recommended the Nvidia GTX 660ti over the 580. After parting with my credit card info, I looked into the previous forum posts about Graphics Cards and Memory http://www.agisoft.ru/forum/index.php?topic=379.0 and found my way here: http://clbenchmark.com/
They have a very handy comparison feature, but I don't understand which of these seperate benchmarks are PhotoScan relevant:
•   Physics: SPH Fluid Simulation
•   Graphics: Raytrace
•   Vision: Optical Flow
•   Image Filter: Sobel - Global memory usage
•   Image Filter: Sobel - Local memory usage
•   Image Filter: Sobel - Image usage
•   Image Filter: Seperated Gaussian Blur - Global memory usage
•   Image Filter: Seperated Gaussian Blur - Image usage
•   Image Filter: Median3x3 - Image usage
•   Programming Principles: Bucketing - Parallel Associative Local
•   Programming Principles: Reduction - Parallel Commutative case with Local memory usage
•   Programming Principles: Scanning - Parallel Associative case with Local memory usage
•   Programming Principles: Scanning - Sequential Associative case
•   Programming Principles: Sum - Global atomic add
•   Programming Principles: Sum - Local atomic add
•   Programming Principles: Bitonic Merge Sort
•   Programming Principles: Tree search
Can anyone decode this for me? While the two cards have similar total number of benchmark points, there is quite a bit of difference for each of the individual benchmarks. Since I am not a gamer, I am assuming the "Programming Principles" are the ones of most importance, but there are significant differences between each of these benchmarks. For me, a little knowledge, is a dangerous, and an expensive, thing. Thanks!

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #1 on: November 02, 2012, 09:55:00 PM »
Mitchel just run the benchmark from the forum here...... best for comparsion.... The 6xx series are just so fast like the 5xx series. Nvidia is cripling performence in drivers fro Copmute - Open CL they want that we buy the PRO card....... Afther a rought weeks will hopefyly have some time to prepare results and they are very promising for clearenace on this.....
----------------
www.mhb.sk

Matt

  • Full Member
  • ***
  • Posts: 104
    • View Profile
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #2 on: November 03, 2012, 07:16:24 AM »
I have had a play with the nvidia 470, 570, 580, 590, 680 and the 580 or 590 (essentially 2 580's) wins hands down in terms of price performance. Newer Nvidia drivers cards have a lower bandwidth and significantly decrease compute performance. My advice is buy a second hand 580 or 590 to start with to get yourself up and going and spend the money you save on extra ram. Its either that or have a play with the 7000 series Radeons and hope the drivers stay stable. I will have some new 7970 OC benchmarks for for Wishgranter next week.
« Last Edit: November 03, 2012, 02:33:13 PM by Matt »

MitchellT

  • Newbie
  • *
  • Posts: 16
    • View Profile
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #3 on: November 04, 2012, 05:14:45 AM »
Thanks guys! I switched my order back to the 580.
Just to satisfy my hacker wannabe curiosity, can anyone comment (are you listening Agisofters?) on which of listed cl  benchmarks in my original question are the most important for PhotoScan? ???

Matt

  • Full Member
  • ***
  • Posts: 104
    • View Profile
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #4 on: November 04, 2012, 08:08:01 AM »
Hope you got the ASUS DCII as you can overclock the snot out of it without a waterblock and increase the performance substantially. Have a play with the older drivers as they seem to offer better uncapped open cl performance. There are a few suggestions on the forum for benchmarks but you are better of running the test project and playing around with the CPU/GPU core slider. This enables you to tweak your system specifically to the type of processing you are doing. 4 of 16 cores work best for me but i am running dual cpu's.

Wishgranter

  • Hero Member
  • *****
  • Posts: 1202
    • View Profile
    • Museum of Historic Buildings
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #5 on: November 04, 2012, 01:18:41 PM »
its mostly uses scalar computations, but hard to say what precise, because of its specific feature.....
----------------
www.mhb.sk

MitchellT

  • Newbie
  • *
  • Posts: 16
    • View Profile
Re: Revisiting Graphic Card OpenCL Benchmarks
« Reply #6 on: November 10, 2012, 10:39:03 PM »
Thanks all,
Matt, I ordered a used Asus DCU for less than the 660. The rest of the build arrives soon.
I will post the results.