Video card again - GTX disaster

So, my 8 month old home workstation computer seemed to be undamaged from the fire, but when I tried to start it up again, it failed - lights and fans, but otherwise black screen… Since we weren’t sure if it was the vid card, the HD, the mainboard, or something else, and since it was “totaled” anyway by the insurance company, my computer builder built me another one. Pretty much the same config, just a slightly faster i7. However, since the original had a Quadro 2000 and there has been so much noise around Quadros not performing any better than GTX’s with Rhino, I told him - how about putting in a good GTX instead? He suggested a 780 with 3Gb, but said it wouldn’t fit in the original sized case - too long - so we would have to go with a huge Coolermaster case. I said, OK, what the heck… Foolish me…

New computer built, new screens (same as earlier, dual 1920 x 1200 displays), both disks from the original were OK, so they got imaged over, so the machine is almost identical to the original with the exception of the vid card.

The performance of the 780 positively sucks with Rhino. There isn’t a Holomark available currently to test, but just making 100 10mm spheres (jagged and faster) and running TestMaxSpeed (shaded mode, 4x antialiasing) takes 4.6 seconds. Compare that to my nearly 2 year old laptop with a Quadro 2000M/2 Gb which does the same thing in 1.75 seconds. With 16x AA to boot…

Edit: my Quadro 4000 workstation at school does testmaxspeed with the 100 spheres in 0.6 seconds…

I thought, this can’t be right, there must be some driver problem, so I completely uninstalled all video drivers all the way back to VGA and then installed the latest WHQL Nvidia driver using “clean install”.

Exactly the same result.

So now, on top of everything else, I am now the proud owner of a brand new handicapped computer… So, what to do now? Go back and pay for a Quadro out of my own pocket? :rage:

GTX as good as or better than a Quadro? Not anywhere near in my experience…

–Mitch

I’m testing with a Titan not a 780 and the results are only 1 second faster with your sphere array test. I also tested drivers from May 'til the last one from the end of July with no speed difference so I agree it’s not the driver. I don’t have a Quadro here to compare.

For the product sized models I work on I find these cards to be quick in most cases. Of course, faster is better and your Quadro in the laptop sounds great. So really the only thing that might help ease the pain is that you now have 2300 Cuda cores at your disposal for rendering. Give the Octane plugin a try to see what I mean.

@jeff is there any hope the GTX cards will work faster in v6 or is this a strictly hardware difference with the Quadro line?

I will have to look into this pretty deeply to see why the differences are so big/bad. I have Quadros here and a 680 GTX, so I should be able to duplicate Mitch’s example. I currently have the 680 installed and an 5x5x4 array of 10mm spheres (100) at maximized viewport yields about 3.5 seconds here (not sure why it’s not as bad as Mitch’s score atm)…but I’m seeing very odd differences between Rendered mode and Shaded, especially if I’m only working with the render meshes…Using Render meshes only, shadows off, I get about 0.5 seconds TestMaxSpeed…which is what I would have expected because the VBOs kick in…but shaded mode doesn’t seem to be affected…Hmmmmm, very odd. But this will take me some time to delve down into what exactly is happening. But I’d like to first try the same thing on a Quadro to confirm Mitch’s differences.

Mitch, you’re one of the most savvy Rhino users I know… but I still have to ask this:

Are your Shaded mode settings EXACTLY the same on both your laptop and new desktop? Any small change can make a huge difference in frame rates… Even a one-pixel thicker edge or isocurve can make a difference. So I want to make sure everything is the same…the safest test (IMO) is to first “Restore Defaults” for Shaded mode on both systems, and then compare results… But I’m guessing you’ve probably already done that…I’m just making sure.

Something definitely doesn’t “feel” right here, however, if one system is working well and another is not, it’s probably not going to be a setting in Rhino that can fix it (again, assuming all settings are exactly the same on both target machines)… I will investigate this today, once I finish my Monday morning regiment.

-Jeff

Thanks for looking into this one Jeff. I see what you mean with Rendered mode being faster (isos and edges on but shadows off). I get 2.5 secs here rotating a maxed viewport of 100 spheres. I can get the same speed in Shaded if I change the lighting scheme to ‘no lighting’ and I can make it super slow (6+ secs) if I enable advanced GPU lighting which Rendered also uses but without the slowdown.

Your last test is slower because Shaded mode uses 3 light sources, Rendered mode only uses 1 (unless there are some in your scene). 3 lights, means 3x the shading and lighting calculations per frame.

-J

I see what you mean and can match the time in shaded with Adv GPU mode if I use Scene lighting.

Hi Jeff,

The shaded mode was custom, but the same on both systems, as one was imported from the other. However, just to make sure, I reset Shaded to default on both computers and used the same file (from Dropbox) to test, single maximized viewport of the same pixel size (1389 x 808). I even disconnected the second monitor on the workstation to see if that would make a difference - it didn’t.

The result was the same, a consistent TestMaxSpeed of about 1.76 seconds for the laptop with the Quadro 2000M and about 4.64 seconds for the workstation GTX 780. Pretty disappointing. Visually it’s a LOT slower and a file with more elements is just unmanageable. I can’t work with it at all, so it’s probably going to end up on the shelf or in the trash. :confounded:

–Mitch

Ok…thanks… But I wouldn’t do anything yet…you might make a bad situation worse… Just because your laptop is performing better doesn’t necessarily mean that putting a Quadro in your desktop will make things “equal”.

Please give me some more time to see if this is indeed a reproducible case between Quadros and GeForces on my end… I have a Quadro K2000 here, that I will be testing out shortly. If things look like they are 3x to 4x better than the GeForce 680 GTX I have, then I will be glad to give the go-ahead… I might even be willing to trade you the K2000 for the 780…depending on the results.

-Jeff

Well, I’m thoroughly disgusted. I’m seeing major differences between the K2000 and the GTX 680. Granted, the K2000 is a brand new Quadro, but still. The horsepower that is supposed to exist in the 680 just doesn’t seem to be there…the K2000 is a single slot, very narrow, very “small” card with no additional power requirements. The 680 is a double-slot, HUGE, card requiring 2 additional power plugs (1x6pin and 1x8pin PCI-E adapter)… And the results of this simple test are ridiculous to say the least.

Using a simple model, 5x5x4 array of 10mm spheres (100 total), running TestMaxSpeed in a maximized viewport in Shaded mode with max AA enabled:

The K2000 yields 0.86 seconds!!!
The GTX 680 yields an abysmal 4.0+ seconds!!!

So WTF?

Because this makes absolutely no sense to me, my brain is telling me that this must have something to do with settings and profiles somehow. Rhino sets up NVidia Profiles (hidden) on installation and/or during the first run of Rhino… why the profile settings would make such a huge difference is beyond me…or even if that is what is happening. I will now start picking apart the profile settings one-by-one to see if they are what is causing such huge differences on such a simple test case.

In any event, Mitch your findings are exactly correct…and not only do Quadros seem to perform better, these new K series Quadros are looking like they’re blazing fast… But I’m not saying anything definitive until I do more investigating.

-Jeff

And the K5200 is yielding 0.5 seconds!!! I have smaller K series cards here too, I will be testing them as well… This just really doesn’t seem right…something must be going on between Rhino and GeForces… or perhaps even Rhino and NVidia (eh em)… I’m no conspiracy theorist, trust me… but something just don’t seem right at the moment.

-J

Ok last post on this today…but this has really become pretty ridiculous for GeForce… I just tried the same test using the K420 (no I did not leave off a ‘0’ zero)…that’s the K Four Hundred Twenty Quadro…

It gets 1.65 seconds vs. the GeForce GTX 680’s 4.0+ seconds. Just doesn’t make any sense… and I have yet been able to find any rhyme or reason behind it.

For visual comparisons, I’ve attached photos of the two cards… Who would ever guess that the 420 would out perform the 680?

-J


Don’t trash your GTX just yet Mitch, I tried this test on my Dell with K2000 (x2) and it took 4.2 secs, maybe something other than the card is at fault?

Jeff, 0.86s sounds too good to be true. I set up the sample a second time using default shaded displaymode and maximised viewport , see screenshot fwiw:

Jeff!

thanks for looking into this, it’s awesome to see you on the case.

Wow, great result. My Quadro 6000 needs 2.6 seconds. GPU usage during the test 0%!!! I hope there is a software solution to get speed.

Jeff, I happy to see you are looking for a solution. Good luck.

Hi guys, I agree with ATH, it’s great, wonderful actually, to see you on this case!

This is a bit OT, but I think it can be a good read for some:
(Valve is the developers behind HalfLife and the Steam platform)

This is what interested me: “Microsoft has relentlessly marketed DirectX, too — and who can forget the release of Windows Vista and Microsoft’s OpenGL smear campaign? Vista’s bundled version of OpenGL was completely crippled, forcing many devs to switch to DirectX.”

If that is true, that OpenGL was crippled when it was shipped with Vista, and now that OpenGL is bundled with the drivers, as I understand it, can this still be the case?

I know that GeForce 400 and above have floatingpoint limitations, but does OpenGL use floatingpoint calculations in Rhino? And if so, could that data be “down sampled” before the GPU get’s to work with it? (I mean, is it a theoretical possibility to run OpenGL less accurate and fast?)

Edit:
BTW: One of the odd things I found during the Holomark beta was that ONE GeForce Card got Quadro perfomance on the meshes.
Darn Outliers… :smile:

1 Like

Just to confirm the findings for @BrianM

Test: Quadro 4000:
5x5x4 spheres, r=5, space = 12, default mode, 8x AA
4 views: 0.78 sec
Maximized: 1.2

File units mm 0.01 tolerance

No AA and modified dispmode:
(No material, default light, no specular)
4 views: 0.5 sec
Maximized: 0.56 sec
(edit: note: this is with an old i7 950, on a 64bit windows 7, so the card can perform even better on a newer machine where the cpu isn’t the bottleneck)

Same test on my laptop with a 330M:
Default, 4x AA, 4 views: 3.4 sec

Yep, as I said in my first post, I actually had to opt for a huge case for it to fit into, whereas the Quadro 2000 fits easily into a much smaller “standard” mid-tower case.

–Mitch

0.64 seconds for my Quadro 4000 here at work. I used a 10 x 10 sphere layout, but that doesn’t matter, got the same result with a 5 x 5 x 4…

–Mitch

Having a GTX680 here as well, and I get the same slow results.

This is interesting:

-C-H-A-R-L-E-S-

I just did a test on an older system with a Radeon 7700 card, and it took 3.9 seconds and used 33% of the Core2Quad 9550 CPU. So it seems it’s not only a nVidia thing…

@Jeff is that an expected CPU load for that test?
My Quadro i7 system uses only 7%, Both chips are quad core, and only the i7 is HyperTreaded.