Quadro vs Geforce performance of Cloudcompare ?

[Dimitri]

Dear all,

i'm about to change my laptop and workstation, and I'm wondering whether there's any benefit in using nvidia quadro graphics cards (with theoretically optimized opengl drivers) compared to geforce consumer cards ?
So far I've only used Geforce cards without any issue, but I'd like now to work with really big datasets (close to a billion points ?) and I'd like to optimize my workstation (beyong super big RAM, multicore XEON and SSD).

Has someone been able to do a side by side comparison ? it could be interesting to have a embedded test with a reference dataset to test fps in various configurations.

Is the graphic card memory important (beyond a minimum level) such that SLI configurations of 4Go medium graphics cards would be significantly better (beyond the strict increased GPU power) than a single high end geforce 4Go card ?

Thanks for your feedback

Dimitri

[daniel]

Well, looking at the way we currently use the GPU, I don't think a Quadro would help.

With the recent use of VBOs you now mainly need:
- a huge GPU memory
- and/or a very fast CPU/GPU transfer rate (if the cloud doesn't fit on the GPU memory, or if you play a lot with the dynamic scalar-field based coloring ;)

We don't use the GPU computing power at all :(

But apart for the improved display accuracy of Quadro cards (that could be interesting if you zoom in a lot on your clouds), the other differences are quite mysterious to me. So I'd be very interested in such a comparison (or more info on the other advantages of Quadro cards).

[Aarie]

Dear Dimitri,
i've faced the same question whether a Quadro offers substantial benefits over the Geforce cards. Short answer: It depends alot on your application and your budget.
Some benefits of the Quadro cards, which were particullary interesting for me, is the support of 10bit displays (for color grading) and support for Quadbuffer stereo. Other advantages, for example, are the use of ECC-memory and multi display sync.

The difference in performance usually sparks quite some heated debates: Usually the Quadros have a lower clock frequency since they are optimized for stability. So the raw perfomance is quite often worse, than their Geforce equivalents. On other hand I know that on Geforce Cards Nvidia artificially limit the viewport perfomance when using double sided shading since they consider this a "professional" feature. So basically much of the hardware is the same (except more VRAM on Quadro cards) but the main difference are the drivers. There used to be perfomance drivers for certain applications, which were certified etc. These could boost the viewport perfomance quite substantially.

I personally have a (lower end) Quadro card for color grading work and a Geforce for viewing large amounts of polygons or doing certain calculations.
And, last but not least, the price for higher end Quadro cards is quite painful.

Cheers.
Arie

[Dimitri]

HI all,

thanks for the feedback. My question was really specifically for Cloudcompare and the potential benefit of optimized quadro drivers for VBO.

Arie: I think there's a way for a quadro to emulate a geforce by choosing in the driver settings a "game developper" mode. Do you mind trying and testing the fps you obtain in CC for the same large point cloud ? This might give us some info.

Because I had good experience with Geforce and don't need the quadro for other applications, I'll continue with these and will experiments a SLI of 2 geforce 970 with 4Gb RAM each to see how it handles 1 billion of points.

Daniel: maybe we could post a reference "large" point cloud (200 millions points) in .bin format so that users could post their FPS and associated hardware configuration ?

[Aarie]

Dimitri,
I'll try to take a look at that tonight. Though the perfomance probably is more limited through the lower clock frequency.
Regarding SLI: If I'm informed correctly this will only increase perfomance if the application has a profil for SLI. Not every software automatically uses both gpus for viewport display.

Daniel, would 4GB of VRAM be enough for storing 1 billion points?

[daniel]

On the CPU, you can put about 90 M. blank points per Gb. So hardly 360 M. on 4 Gb (and for this price you don't have colors ;). If the GPU has the same internal representation (that must be true for Geforces, but maybe not for Quadros) we can expect roughly the same behavior.

We really need an dedicated octree structure for displaying huge cloud...

[Dimitri]

Ok, I think I've got most of the answers I needed after some experimenting:

In short: Quadro cards do not have an advantage over geforce (for the same GPU specs), and can actually be a bit worse. To test this, you just have to go in the drivers and select the profile "game development" and the card will behave as a "geforce". We did the test with a quadro K1000m that we had here, for blank, scalars, color and with or without EDL filter. In short the fps were similar or 10% worse in the "quadro" config compared to the "geforce" config (we restarted each time CC and used exaclty the same cloud and view). Arguably, even if this does not shows in the number, the 'quadro' mode is less reactive in general. So I would advise all people with quadro cards to use the "game development" mode.

SLI : this currently does not have any interest in terms of additional VRAM as the system stills see only the amount of ram on one card. THis could only improves the fps (which could be important for very large clouds). ATI is working on a solution to have the total VRAM accessible.

Now, using my old Geforce 540m, I explored the scaling between point cloud size and VRAM (as shown in the console). Here is the graph:

Vram_versus_PointCloudSize (Large).jpg (99.27 KiB) Viewed 17836 times

This shows how many points you could load on the best graphics card around (typical gaming cards can have up to 8Gb vram now).

And the following graph shows the scaling of fps with point cloud size.

Fps_vsNbpoints.jpg (41.41 KiB) Viewed 17836 times

I've not tested on my workstation with a more powerful card, but somehow I expect the scaling in Npoints^(-0.75) to hold for any card...

I'll wait a bit and will probably buy a Geforce GTX980 with 6 Gb Ram.

Hope this helps

[daniel]

Impressive!

Here is my (modest) contribution to this study. I used a blank cloud. My card is a GTX780 with "only" 3Gb VRAM. The 3D view display resolution was 1200x800:

cc_fps_blank_cloud.jpg (30.13 KiB) Viewed 17835 times

With a log scale, I get a -0.86 exponent (but of course that's not the same setup):

cc_fps_blank_cloud_log.jpg (20.57 KiB) Viewed 17835 times

P.S.: ok to share a big cloud as BIN file (do you have one? ;). My biggest "public" cloud is Bremen's Schlachte (173 M. - the one I used for this test). source: http://kos.informatik.uni-osnabrueck.de/3Dscans/ (but not as a BIN file). We could also get big LIDAR datasets.

[Dimitri]

Hi Daniel,

interesting to also find a powerlaw, with an exponent which is not that far !

As for the cloud. No need for the users to download a very large cloud: a 25 million point cloud, duplicated in CC twice or four times (and visible) will give the same fps than 1 single cloud of 50 or 100 millions points (I've checked). Only thing is to keep the same view. I think the most interesting and relevant benchmark would be for a point cloud with scalar and EDL activated (this is my typical use).

I can provide the 25 million points cloud of Lague et al., ISPRS, 2013, but it's already 300 Mo in .bin format.

[daniel]

No problem, I have plenty a public 25 M. clouds ;)