Maya 6 - 7900GT or 6800 Ultra -> QuadroFX 4000

EDIT: This post is outdated -- the results were skewed by a weird Maya bug, plus the fact I forgot to disable Vsync. See my post 3 msg down for the newer results.

----------------
Okay, here's some tests I did with my 2 cards. Not completely conclusive but still useful results.


6800Ultra softmodded to Q3000+
7900GT overclocked to 630MHz core (default 7900GT is 450MHz I think)

I didn't underclock the GT because I'm not completely sure how to do it. I bought this card already voltmodded
and bios clocked to this speed. I do have another 7900GT clocked to 550MHz which I could test however.

More suggested benchmarks methods you want me to try?

Samiam -- my conclusion at this point is that it's not really going to make much difference for you to upgrade to a 7900 unless you play games on your machine as well. Even though texturing shows a speed improvement in these tests, for my work on game level design, I see absolutely no difference when the geometry count gets very very high. (CPU is the limit there).


------------------

My custom benchmark comments:

Lighttest01:
Most of the tests came out identically. The big differences are in HQ rendering mode.
Quadro seems to have a large advantage in HQ rendering mode until we enable texturing.
The 7900UC is definitely faster at texturing.

Wiretest01:
I was surprised to see the 7900UC actually beat the Quadro in wireframe mode. Solidshade
goes to the Quadro, but textured again goes to the 7900. In this test though, it's always a
farily small margin.


Artisan test:
I took a high poly plane and used sculpt surfaces tool to draw on it. I made the plane high
enough polys so it would run around 30 fps while I was drawing. I signed my name in the plane
with the sculpt tool watching for lag and watching lowest-highest framerates. Both cards
appeared identical, both with a very slight lag behind my mouse.

I also have synthetic benchmarks at the end of the data. They don't seem to be representative of real-world results.

As to menus, etc. I really don't notice any difference between the quadro and the Geforce. I've been using maya for 2 years on a 5900GT card with no issues with menus or overlays or whatnot. I didn't see any difference using the SQ3000+ or the 7900GT either.



----------------------------

Code:

All tests were performed using Nvidia 91.47 driver.


My custom benchmarks: Timed with stopwatch so expect a margin of error

Custom Lighttest01 -
	64,760 polys onscreen
	scene consists of 81 blobby spheres, a floorplane, 8 colored lights (max that would show up in real-time view)
	The 8 lights rotate 3 revolutions over 450 frames
	By default, 2-sided lighting is on

		6800U->Q3000+	7900UC
Lit		7:71 sec		7:72 sec
Lit 1-side		7:75 sec		7:69 sec
Lit HQ		45:44 sec		1:06:07 sec
Lit		7:69 sec		7:69 sec
(wire on shd)
Lit		45:44 sec		1:07:53 sec
(wire on shd HQ)
Wireframe		7:77 sec		7:69 sec
Wireframe Smooth	7:75 sec		7:69 sec
Solidshade Xray	7:72 sec		7:69 sec
Textured		7:69 sec		7:69 sec
Textured HQ	40:07 sec		30:91 sec
Textured HQ	45:32 sec		33:95 sec
(smooth wire on
shade)
Textured		15:21 sec		15:08 sec
(transparency)
Textured		1:21:91 sec	54:69 sec
(transp. HQ)
Textured 1-side	15:13 sec		15:13 sec
(transparency)


----------------------------

Custom Wiretest01 -
	~1 million polys in scene, 600-800K on screen
	A game level
	default lighting
	Camera moves around scene over 450 frames

		6800U->Q3000+	7900UC
Wireframe		44:01 sec		38:38 sec
Smooth Wireframe	46:46 sec		44:97 sec
Solidshade		1:54:44 sec	2:05:46 sec
Textured (100 	54:44 sec		52:82 sec
   frames only)





SYNTHETIC BENCHMARKS:
--------------------
SpecViewPerf 8.1
--------------------

7900GT Uberclocked (630 core)
maya-01 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   58.5     N/A   7      True    8    8    8    0     24      0   
  2   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  3   11.00   20.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   20.0     N/A   7      True    8    8    8    0     24      0   
  5   12.00   19.9     N/A   7      True    8    8    8    0     24      0   
  6   11.00   60.0     N/A   7      True    8    8    8    0     24      0   
  7   11.00   60.1     N/A   7      True    8    8    8    0     24      0   
  8   11.00   60.0     N/A   7      True    8    8    8    0     24      0   
  9   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-01 Weighted Geometric Mean =   35.34

6800U->Q3000+
maya-01 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  2   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  3   11.00   59.4     N/A   7      True    8    8    8    0     24      0   
  4   11.00   59.5     N/A   7      True    8    8    8    0     24      0   
  5   12.00   46.9     N/A   7      True    8    8    8    0     24      0   
  6   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  7   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  8   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  9   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-01 Weighted Geometric Mean =   57.99


--------------------
SpecViewPerf 9
--------------------

7900GT Uberclocked (630 core)
maya-02 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   60.1     N/A   7      True    8    8    8    0     24      0   
  2   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  3   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  5   12.00   20.0     N/A   7      True    8    8    8    0     24      0   
  6   11.00   19.9     N/A   7      True    8    8    8    0     24      0   
  7   11.00   19.9     N/A   7      True    8    8    8    0     24      0   
  8   11.00   19.8     N/A   7      True    8    8    8    0     24      0   
  9   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-02 Weighted Geometric Mean =   26.92


6800U->Q3000+
maya-02 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  2   11.00   59.6     N/A   7      True    8    8    8    0     24      0   
  3   11.00   34.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   59.5     N/A   7      True    8    8    8    0     24      0   
  5   12.00   59.6     N/A   7      True    8    8    8    0     24      0   
  6   11.00   59.2     N/A   7      True    8    8    8    0     24      0   
  7   11.00   56.3     N/A   7      True    8    8    8    0     24      0   
  8   11.00   51.2     N/A   7      True    8    8    8    0     24      0   
  9   11.00   59.3     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-02 Weighted Geometric Mean =   54.69



----------------------
SpecCapc for Maya 6.0
----------------------
7900UC	6800U->Q3000+
4.99		5.03	wireframe
27.31		0.14	smoothShaded
14.19		0	Texture Load Time
8.07		0.03	textured
7.57		0.03	texturedSelected
9.98		0.03	texturedHilited
scene	werewolf.ma
1.74		1.72	Project File Load
5		5.04	wireframe
5		5.04	smoothShaded
13.21		10.84	Texture Load Time
8.19		7.2	textured
7.72		7.79	texturedSelected
9.99		10.04	texturedHilited
scene	hand1.ma
15.01		15	Project File Load
1.17		1.18	wireframe
2.39		2.93	smoothShaded
0.03		0.05	Texture Load Time
2.39		2.95	textured
3.08		3.56	texturedSelected
3.45		3.64	texturedHilited
scene	Insect.ma
1.33		1.36	Project File Load
32.36		32.61	wireframe
43.16		43.36	smoothShaded
5.08		4.88	Texture Load Time
97.2		87.03	textured
99		94.53	texturedSelected
107.89	108.7	texturedHilited
scene	Squid.ma
1.22		1.23	Project File Load
7.73		7.76	wireframe
7.74		7.77	smoothShaded
18.96		18.63	Texture Load Time
9.64		10.47	textured
9.56		9.56	texturedSelected
15.45		15.54	texturedHilited
scene	handDef.ma
15.64		15.6	Project File Load
628.69	642.83	smoothShaded

 

@augie: it seems to me that your vertical sync is set to the default "Application-controlled" setting,
since none of the SPECviewperf results are above 60 frames/second.
It is capped to the update refresh of your monitor, which is 60 Hz.

When testing, set it to "Off" instead, then it will not be capped, and this will give you much higher speeds...!

 

Whoopsie, you are right. I don't know why I didn't check that. I got nearly 3x the speed on the basic lit benchmark.... I'll have to redo the benchmarks.

update: There appears to be quite a bit of randomness in certain benchmarks depending on how long maya has been running. Some benches, specifically the HQ on the 7900 were like 4x as fast as previously in this run. Then even on this run I might get a several second difference between two benches depending on how long maya has been open... And what's even more weird is that sometimes maya seems to get faster after being open for a while! Other times it seems to get slower. I'm trying to run the questionable benchmarks many times under different conditions and see what's the best.

Augie

 

NEW RESULTS:

The above results (three posts above this) were not qute accurate. I redid all the benches, plus some new ones because I did not have Vsync off. Additionally, I discovered some inconsistencies with Maya. Depending on what files were loaded before each custom benchmark, I might get slower or faster numbers. (I actually get slower numbers on some benches if I load Maya fresh, even on reboot, than if I load another file first, switch to HQ mode, then load the original file again...) I think I found the order of loading that would yield the best numbers for each case and tried to be consistent between both cards. (I have no idea why Maya behaves this way. It was most obvious with the HQ rendering mode, which is leading me to think it's doing some weird stuff with texture baking/caching or something) This weird behavior is the reason for the huge differences in HQ rendering of this test run vs the previous test run above.

6800Ultra softmodded to Q3000+
7900GT overclocked to 630MHz core (default 7900GT is 450MHz I think)

New conclusion:
It looks like the 7900GT, at least at 630 core, matches or beats the 6800U->Q3000 in pretty much every case. Not by much in most tests, most notable in HQ rendering mode where the 7900 comes out on top by a good margin. It's really a minimal speed difference in MOST of the benchmarks. (Within margin of error) I never use HQ rendering mode, so the 6800U->Q3000+ is nearly identical to the 7900GT UC for my purposes. If you use HQ rendering mode, the 7900GT UC wins hands down.

When the 6800U is not in Quadro mode, it keeps up in a lot of the benchmarks, but definitely is slower in HQ rendering. (I didn't list these benchmarks below, but I tested a couple to ensure there was indeed a difference between Quadro and non-quadro mode. Ex: 50:91 sec vs 39:25 sec HQ Textured) The couple of non-HQ benchmarks I tried either came out a bit slower or identical to the Quadro mode.

To sum it up for Maya 6.01 - If you're not using HQ mode, your limitation is probably going to be your CPU and not your video card.


------------------

My custom benchmark comments:

Lighttest01:
Most of the tests came out identically. The big differences are in HQ rendering mode.
Now the 7900 seems to have a large advantage in HQ rendering.

Wiretest01:
Essentially a tie. 7900 slightly faster in wireframe, but this could just be a fluke. I got better scores on the older benchmark for the Q3000, but I couldn't reproduce this on this benchrun.

Artisan test:
I took a high poly plane and used sculpt surfaces tool to draw on it. I made the plane high
enough polys so it would run around 30 fps while I was drawing. I signed my name in the plane
with the sculpt tool watching for lag and watching lowest-highest framerates. Both cards
appeared identical, both with a very slight lag behind my mouse.

I also have synthetic benchmarks at the end of the data. They don't seem to be representative of real-world results.

As to menus, etc. I really don't notice any difference between the quadro and the Geforce. I've been using maya for 2 years on a 5900GT card with no issues with menus or overlays or whatnot. I didn't see any difference using the SQ3000+ or the 7900GT either.


----------------------------

Code:

All tests were performed using Nvidia 91.47 driver.
V-Sync disabled.
Maya 6.01
E6600 @ 3.0GHz, 2GB memory @ 667MHz 4-4-4-12
Windows XP Pro

My custom benchmarks: Timed with stopwatch so expect a margin of error

Custom Lighttest01 -
	64,760 polys onscreen
	scene consists of 81 blobby spheres, a floorplane, 8 colored lights (max that would show up in real-time view)
	The 8 lights rotate 3 revolutions over 450 frames
	By default, 2-sided lighting is on

		6800U->Q3000+	7900UC
Lit		2:71 sec		2:60 sec
Lit 1-side		2:63 sec		2:63 sec
Lit HQ		31:32 sec		25:50 sec
Lit		3:69 sec		3:32 sec
(wire on shd)
Lit		32:13 sec		26:26 sec
(wire on shd HQ)
Wireframe		2:03 sec		1:94 sec
Wireframe Smooth	2:12 sec		2:12 sec
Solidshade Xray	3:89 sec		3:22 sec
Textured		6:37 sec		6:57 sec
Textured HQ	39:26 sec		30:66 sec
Textured HQ	39:71 sec		31:77 sec
(smooth wire on
shade)
Textured		10:15 sec		10:57 sec
(transparency)
Textured		1:05:72 sec	54:15 sec
(transp. HQ)
Textured 1-side	9:88 sec		10:44 sec
(transparency)


----------------------------

Custom Wiretest01 -
	~1 million polys in scene, 600-800K on screen
	A game level
	default lighting
	Camera moves around scene over 450 frames

		6800U->Q3000+	7900UC
Wireframe		42:21 sec		38:31 sec
Smooth Wireframe	41:69 sec		41:38 sec
Solidshade		2:05:44 sec	2:04:44 sec
Textured (100 	54:44 sec		54:58 sec
   frames only)


----------------------------

Custom Spintest01 -
	Same scene as lighttest01
	64,760 polys onscreen
	scene consists of 81 blobby spheres, a floorplane, no lights
	Camera rotates 3 revolutions around scene aiming at center
	By default, 2-sided lighting is on, default maya lighting on
	Tested in hotkey "5" smooth shading mode unless otherwise noted

		6800U->Q3000+	7900UC
Solidshade		8:50 sec		8:19 sec
Solidshade		8:70 sec		8:39 sec
  (no lighting mode)
Solidshade		11:50 sec		10:88 sec
  (wire on shade)
Solidshade		11:32 sec		11:30 sec
  (wire on shade,
   selected)
Solidshade HQ	31:64 sec		30:16 sec
Solidshade HQ	34:76 sec		32:84 sec
  (wire on shade)
Solidshade HQ	34:82 sec		32:70 sec
  (wire on shade,
   selected)
Flatshaded		16:64 sec		14:13 sec


-----------------------------


SYNTHETIC BENCHMARKS: (same as before)
--------------------
SpecViewPerf 8.1
--------------------

7900GT Uberclocked (630 core)
maya-01 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   58.5     N/A   7      True    8    8    8    0     24      0   
  2   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  3   11.00   20.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   20.0     N/A   7      True    8    8    8    0     24      0   
  5   12.00   19.9     N/A   7      True    8    8    8    0     24      0   
  6   11.00   60.0     N/A   7      True    8    8    8    0     24      0   
  7   11.00   60.1     N/A   7      True    8    8    8    0     24      0   
  8   11.00   60.0     N/A   7      True    8    8    8    0     24      0   
  9   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-01 Weighted Geometric Mean =   35.34

6800U->Q3000+
maya-01 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  2   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  3   11.00   59.4     N/A   7      True    8    8    8    0     24      0   
  4   11.00   59.5     N/A   7      True    8    8    8    0     24      0   
  5   12.00   46.9     N/A   7      True    8    8    8    0     24      0   
  6   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  7   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  8   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
  9   11.00   59.8     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-01 Weighted Geometric Mean =   57.99


--------------------
SpecViewPerf 9
--------------------

7900GT Uberclocked (630 core)
maya-02 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   60.1     N/A   7      True    8    8    8    0     24      0   
  2   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  3   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
  5   12.00   20.0     N/A   7      True    8    8    8    0     24      0   
  6   11.00   19.9     N/A   7      True    8    8    8    0     24      0   
  7   11.00   19.9     N/A   7      True    8    8    8    0     24      0   
  8   11.00   19.8     N/A   7      True    8    8    8    0     24      0   
  9   11.00   30.0     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-02 Weighted Geometric Mean =   26.92


6800U->Q3000+
maya-02 Viewset
-------------------------------------------------------------------------------
Test Weight  Frames    DLB  Visual Double     Frame Buffer     Depth  Stencil  
  #    Per Sec   Sec   ID    Buffer Red Green Blue Alpha Buffer  Buffer  
-------------------------------------------------------------------------------
  1   11.00   59.7     N/A   7      True    8    8    8    0     24      0   
  2   11.00   59.6     N/A   7      True    8    8    8    0     24      0   
  3   11.00   34.0     N/A   7      True    8    8    8    0     24      0   
  4   11.00   59.5     N/A   7      True    8    8    8    0     24      0   
  5   12.00   59.6     N/A   7      True    8    8    8    0     24      0   
  6   11.00   59.2     N/A   7      True    8    8    8    0     24      0   
  7   11.00   56.3     N/A   7      True    8    8    8    0     24      0   
  8   11.00   51.2     N/A   7      True    8    8    8    0     24      0   
  9   11.00   59.3     N/A   7      True    8    8    8    0     24      0   
-------------------------------------------------------------------------------
maya-02 Weighted Geometric Mean =   54.69



----------------------
SpecCapc for Maya 6.0
----------------------
7900UC	6800U->Q3000+
4.99		5.03	wireframe
27.31		0.14	smoothShaded
14.19		0	Texture Load Time
8.07		0.03	textured
7.57		0.03	texturedSelected
9.98		0.03	texturedHilited
scene	werewolf.ma
1.74		1.72	Project File Load
5		5.04	wireframe
5		5.04	smoothShaded
13.21		10.84	Texture Load Time
8.19		7.2	textured
7.72		7.79	texturedSelected
9.99		10.04	texturedHilited
scene	hand1.ma
15.01		15	Project File Load
1.17		1.18	wireframe
2.39		2.93	smoothShaded
0.03		0.05	Texture Load Time
2.39		2.95	textured
3.08		3.56	texturedSelected
3.45		3.64	texturedHilited
scene	Insect.ma
1.33		1.36	Project File Load
32.36		32.61	wireframe
43.16		43.36	smoothShaded
5.08		4.88	Texture Load Time
97.2		87.03	textured
99		94.53	texturedSelected
107.89	108.7	texturedHilited
scene	Squid.ma
1.22		1.23	Project File Load
7.73		7.76	wireframe
7.74		7.77	smoothShaded
18.96		18.63	Texture Load Time
9.64		10.47	textured
9.56		9.56	texturedSelected
15.45		15.54	texturedHilited
scene	handDef.ma
15.64		15.6	Project File Load
628.69	642.83	smoothShaded

 

You managed to force an NV40 based 6800 Ultra into NV35 based Quadro FX 3000?

NV40 based GeForces should be modded into a NV40 based Quadro FX 4000.

My recent experience with Maya in wireframe and PaintFX showed me that a 7900GT is no match for a 6800 GT PCI-E softmoded into Quadro FX 4000.
The interactivity on the softmoded card is lots faster than the choppy handling of wireframe models on the 7900GT.
For the speed tests, I was using scenes that contain about 3 million faces.

 

Sorry, it's my mistake. I thought the 6800 converted to FX3000. Ok, it's 4000. It doesn't change the benchmarks though.

I should think I would have seen some difference if it's so great considering I tested a wireframe scene with about 1.1 million polygons. I'll try tripling that.

I'd like to pursue this. What driver version were you using? And what version of Maya? (I'm doing tests on 6.01, but I do have access to Maya 7 as well.)

PaintFX is something I should test too.

Augie

 

Maya 8 with certified 91.36 drivers.

You can spray a lot of polygons using PaintFX, making a 3 mil scene is a snap...

 

Thanks Augie

augie, thanks for the benchmarks. Interesting results.

Makes sense the 7900 would do better on HQ rendering as it has more raw power and HQ mode is a lot like high end game rendering. Did you have light shadows enabled or normal maps in your HQ render? I'd imagine the 7900 could potentially do better on that as well.

Also, did you test HW rendering (actual rendering to a file) with the 7900? I read somebody claimed the 7900 couldn't do it...

Let us know how the 3mil poly test goes, though I wouldn't use particles. Real geometry with textures is more realistic.

Alec ~
To what do you attribute the Quadro advantage in high poly scenes? Driver? I'm unaware of hardware features on the Quadro which manages high poly better, but not sure about that. Could be better memery management?

 

Yep, it's the driver. Smart usage of hardware overlays, display lists, culling lists, triangle strips and full usage of the tranformation engine of the GPU (the T in T&L).
My bet is that when drivers detect a gaming card, they disable hardware acceleration on lot of functions that are never used in game engines.

 

Ah! Now THIS is an interesting thread!
I've been dealing with this stuff for a lot of time: I've been working on Maya since 3.0 and in these years It was the only 3D software I used. I have three systems: a old Athlon XP 2000 with a GeForce MX 440, an Athlon 64@2.3Ghz with a GeForce 5900XT soft and hardmodded to Quadro FX3000 and a notebook with Sempron 3000 and a X700 softmodded to FireGl V5000. I've been testing a lot because I like to get the max from a card and I also wanted to be sure to have high productivity in Maya while retaining a great price/performance ratio.. so I will tell you what, for me, in my very practical experience, is the truth (BEFORE PEOPLE START FLAMING ME: THE FOLLOWING IS VALID FOR MAYA ALONE ):

Short answer: QUADRO/FIREGL are NOT worth the price to be used in Maya, hence go for a strong geforce. An exception to this is that, if you don't have upgrade issues, the soft 6800U->FX4000 is, right for now, the best choice (more on this later).

Now for the long answer:
I'm sorry for my long message but I hope you find it interesting (as I do):
First of all we must make a distinction between compatibility and performance:
quadro and firegl cards WILL give you COMPATIBILITY -> they will provide, more or less, a nearly perfect implementation of the opengl hardware features that maya uses/supports. What are these features? Two sided lighting? No. Accelerated antialiased lines? Partially. Improved viewport performance (that is fps while spinnig a model)? No. Hardware overlay planes? Yes. Improved artisan/paintfx (consequence of hw overlay planes)? Yes. What does all this mean? It means that with these cards you will have fewer if none graphic glitches: some minor display bugs in the menus will fade away and you will have substantially faster hotbox/marking menus artisan/paintfx so....
the PERFORMANCE: to have performance in Maya means to have a fast (possibly real time) interaction AND a fast workflow. You don't use HQ rendering or antialiased lines while working: these are good just to take screenshots. Most of the time you spin wires and plain shaded models. You use hq texturing only when texturing but very often an IPR or a test render will provide a much better feedback for this even if you're using a quadro (if you make some non-basic shading network Maya will not show very well the texture even if you have a real quadro). And now the most important feature: hardware overlay planes. You do use this all the time in Maya but for different pourposes: hw op allow a portion af the screen to be placed in top of another and than to be removed without the need to redraw what was underneath it. Think of this like transparent slides: you put a slide with a background and than another with a menu on top of this. Than you remove the slide with the menu: the slide with the background is still there. This is how quadro/firegl work. Now imagine instead to print the menu on top of the background with a stamp: if you want to remove the menu you have to erase it somehow AND to redraw the portion of the background that got covered by the stamp: this is how geforce/radeon work. This is used every time for hotbox and marking menus! The more complex the scene is (slower redraw time) and the bigger the superimposed menu is (larger area to redraw), the bigger is the difference between workflow speed you get with a quadro or geforce. In years I have learned that hotbox (which gets the most advantage from hw op since it IS a big menu) slows you down and can be replaced more effectively by marking menus/shelves. Marking menus benefit form hw op too, but since they are so small the difference is barely noticeable. So to have a quadro/fgl for this is superflous. The other pourpose of hw op is another matter: artisan/paint fx: if you massively use these you WILL be much more productive with a quadro (the mightier the scene the more you will benefit from having hw op). But if you don't use these features very often (or almost never) you can afford an occasional slowdown for not having paid the price of a quadro.

Given also these considerations I experienced PERSONALLY (no friend, no rumors, no bull****) the following:

1- In general Nvidia cards perform much better than Ati cards in Maya: take a good geforce (not even a quadro) and it will beat the crap out of ANY firegl (this at least till now and till V7100). Real life example: took I x850xt platinum.. modded it successfully to firegl (I know it has not an exact equivalent but it works trust me): scores improved a LOT in specviewperf and apparently it crushed my soft Fx3000 but in the true Maya things are completely different. Same situation with my v5000: much better overall than the x700 but not even near (despite the higher spec scores) to my 5900xt in the real thing.

2- Never trust SpecOPC tests for true application performance for, at least, two reasons: first they implement a fake graphic module for each application (which mimics the real one) and second.... ati and nvidia cheeeeeaaaaat! Like clokkevi stated they cheat! To be polite we could say that they introduced some "optimizations" for the "very reliable" spec tests. SpecOPC is good just to see if you have successfully modded your card. Real life example: yesterday I installed the latest firegl for my softmodded v5000: good 10% improvement in maya score in specviewperf 8.1.... real maya viewport performance improvement: 0, not even half a fps.

3- SpecAPC (the test executed with the real program) provides you a more reliable benchmark result but it is incomplete: it does not take hw op into consideration nor paintfx and that explains why quadros have no apparent advantage on geforces. If you see a benchmark stating otherwise there are just two cases: either they're lying/telling bull**** or they messed with drivers/configuration in at least a "strange" way.

4- Maya is, indeed, very CPU/system bound and doesn't make a massive or correct use of some hw opengl features that have a big impact on other applications performance. Real life example: I hardmodded my geforce 5900 to get accelerated antialiased lines and two sided lighting: this alone provided a good improvement in specviewperf and some cad applications but provided almost no improvement at all in Maya.

5- You have no difference between 45.28 powered softquadro and recent drivers equipped gf: that's, sadly, perfectly normal: the 45.28 drivers are simply too old and the newer ones have really improved. Still you retain advantage in hw op though.

Some things that I personally think are true but I'm not completely sure of (bull**** ):

1- Regarding SpecOPC tests I believe that some workstation producers cheat on their own: they maybe ask ati and nvidia for "specially optimized" drivers to make a better score compared to the one of another producer equipment. That may explain the difference found sometimes between identical or similar hardware equipped workstations, the only true difference being a "slight" different driver version.

2- Nvidia have worsened the opengl support for newer geforce cards. This has many benefits for them: first it widens the gap between their professional cards and their gamer counterparts: this helps to justify the expense for a quadro and, at the same time, helps the overall sellings a lot: they have a lot of older gpu equipped cards still to sell: think of FX1XXX or FX3XXX series. This also makes me believe to AlecRyben when he states that the 6800U->Fx4000 performs better than a newer 79XX: superior raw power alone is no longer enough to make the gf beat a quadro (even if inferior from hw point of view).

And this makes me consider the 6800U->FX4000 the best buy (upgrade issues apart) nowadays. The 6800 is powerful and its quadro support is perfect.
If upgrade is an issue go for a 79XX. In the nearby future it would be possible for higher end FireGl to reach or surpass the performance of, at least, expensive geforces thus making them a good choice especially if you do use a lot of paintfx (plus no glitches): I'm not very moved to this though.

Some recommendations regarding driver settings for tests and some maya considerations:

When benchmarking for specviewperf or maya you will want to:
Select a generic profile for the driver (specviewperf)
Select maya profile (real maya)
Disable any background application
Disable visual themes (just to be sure)
Be sure your overclock, if any, is rock stable
Test your memory with memtest86 to be sure everything is ok
Disable vsync
Defragment the drive
Use certified drivers
Check your softmod is ok with rt AND something like OpenGl Extensions Viewer

Maya usage:
Check overlay planes are on: look at the time slider in Maya if the bar thick under current frame number is solid dark gray you have no overlays. if you have a dotted/grid dark gray behind the current number thay are ok (put the range slider to some like 10 to see better).
Depending on your taste you could want to enable vsync for comfort to the eyes (worst performance but smoother)

This all that came to my mind in this hour. I hope not to have wasted my time and to have been useful at least to somebody.

Greets.
Anax

 

I found your post very interesting and informative. Hopefully some of this info will help me benchmark and test the differences more intelligently.

On just a quick test though, I do not find any difference between marking menu/hotbox speed or PaintFX real time preview. I loaded some huge scenes and tested the lag of marking menus -- I found them the same. Laggy at times, but generally quick.

Another thing, you mention the animation time bar method of seeing the difference between HW overlays and non HW overlays. Well, on both the 7900 and the Q4000 I see the pixelated grid, so apparantly the overlays are enabled on both cards?

As for PaintFX, I made a benchmark consisting of a large mass of painted trees, then rotated the camera around. Both cards came out at 6 fps on that test. The actual process of PAINTING in paintFX never seems to slow down on either card.

I just haven't seen anything yet that confirms this for me. I guess we could be comparing apples to oranges though considering AlecRyben was using Maya 8.

I tested a scene with 8,240,000 polygons in wireframe mode. Preliminary benchmarks put the cards pretty much equal. I need to test again later in more modes when I have a chance.

I do need to try the OpenGL Extensions Viewer as you mentioned. I'll have to find this.

Augie

 

In the firegl control panel and in the quadro one you have an explicit option to disable hardware overlay planes: when this is disabled you ALWAYS get a dark gray back for the frame indicator (I attached an image showing this: I disabled overlay planes so you can see dark gray under the 3). This is true for every video card I ever tried: never got a grid background with geforce or radeon (not even in linux or macosx). If you get a grid with the 7900, then it seems that hw op are enabled and this is really new to me (and interesting).

True. I doubt that hw op could improve the overall speed of display of paintfx while you are basically spinning a wireframe/mesh (because this is what a paintfx stroke becomes after being drawn). Personally I can spot a small difference (on my sempron 3000 with firegl v5000) only when painting a canvas at depth (not the scene with meshes). I assume that our systems are not powerful enough to provide this kind of difference... and this brings us to ->

No I don't think so. In the years I've never found some performance difference between geforce and quadro to be caused by a new/different maya version, so I doubt this is the case with Maya 8.

Have you seen Alec specs? He's running a powerful dual opteron system!
You are basically comparing (from a hw point of view) the performance of a 6800 ultra to that of a 7900GT overclocked... and you're getting the same results!!!! This has only two possible explanations in my opinion:
1- The quadro driver is making the difference (leveling an inferior hardware to a 7900GT)
2- You're system is not powerful enough to push both cards in maya thus clamping them to the same level.

You CAN get out of all your doubts simply doing this: use the 6800 as a normal geforce and compare its performance to that of the 7900GT if you still get pretty much the same results in terms of performance THEN solution 2 applies.

Maya is VERY cpu/system bound and my guess is that unless you have a really powerful system, the differences in performance brought by using hw op are simply made inconsistent by having an insufficient hardware. That's why I recommended the 6800U to FX4000: in any case you get a quadro (so you're sure you get all the features) AND a cheaper videocard (compared to a 7900) that performs very well.

Ciao
Anax

 

Actually, for operations which cannot make use of his 4 cores, my system is faster than AlecRyben's. I'm running a Core2Duo E6600 @ 3 GHz, which by all benchmarks I'm aware of is faster than a 2.6GHz opteron. His system would cream mine at rendering, however.

That is the purpose of this thread, yes. To see which one is faster/better for Maya and where those difference lie, but I'm not as concerned with which is the better value moneywise. A $50 price difference is negligible, IMO. (Can't get a 6800U for less than $150, and 7900OC's are going $200 used ... I actually got my 630MHz 7900GT w/Zalman cooler for $150).

I did tests with the 6800 in non-Quadro mode, just not very many. And yes, it was slower in non-Quadro mode. Significantly so in HQ rendering mode, but not in wireframes, lighting, solidshade, textured, etc. viewports. I will do a round of these tests in non-Quadro mode for confirmation as well. In general, I have to assume that Maya relies MUCH more on the CPU for these ops than the video card.

I will give PaintFX a very thorough testing, then try some hardware particles.
I will also try real-time shadow mode if I can figure out how to make it work. (Not too much luck so far.) I will also try to downclock the 7900GT to regular speeds if I can figure out how without flashing the bios.

I really just want to know WHERE the benefits lie in using the FX4000 over the 7900GT if there are any. It's a bit tough to track down and identify these supposed advantages at this point.

So far, on as fast a CPU as you can buy (clocked higher than Extreme Edition C2D), these cards are equal in 90% of Maya's interface.

That's precisely what I want to avoid. I don't want to assume Quadro is going to be better for Maya because it's Quadro. I want hard facts and figures that can either back that up or burst that common-knowledge bubble.

I'm sure we're making some headway. Hopefully I'll get some good benchmarks over the weekend.

Augie

 

Hmmmm.... I work in game and pretty technical so I'm aware of what game engines do, and they do most of that with the exception of overlays.

1 by 1:

Overlays: Quadro is capable of something like 8 cutouts if I remember correctly, while typical game cards only do 2 or something. Regardless, that's for UI interaction, things like cursors and menus. shouldn't impact rotation of high poly scenes unless you have UI elements all over or something.

All 3D cards do geometry culling. Most advanced culling is software based, such as BSP but that's irrelevant to HW tests. the remainder, such as poly occlusion, is certainly done on game cards as well.

All 3d HW do tri strips including game cards. In fact on consoles like the PS2 which are finicky, tristripping is essential to prformance.

All modern 3d cards do T&L including game cards. Games make extensive use of T&L. Over a decade ago my $3000 FireGL 3000 (comped from maker, k-ching!) was cutting edge bcus it did T&L, but for the last several years everything has.

So that's why I'm saying, aside from overlays there really isn't a big difference between game and pro functions anymore. It's true that game drivers need to be optimised for multipass and texture memory management, but shouldn't be a big deal for Nvidia and in no way justifies the price.

 

Hey Anax,

Agree with almost all of what you wrote. i do think 2-sided viewing and HQ rendering are pretty useful, but that's just my workflow as opposed to yours. For example HQ mode shows normal mapping well which is very important to some users. (and should be totally doable on game cards in Maya btw)


Totally agree specOPC is a joke for real-word performance. Typically SpecOPC will show 100% improvement or more, while real-world Maya shows 0% improvement.

SpecAPC is somewhat better, but like you said also incomplete. It doesn't demo workflow for example. A better SpecAPC would actually fake all the graphic calls to interactive work of building a scene (like the way Painter can playback strokes) including texture painting and such.

FYI: I think you should be able to run more recent drivers than 42.51 on your modded Quadro. I'm running 91.36 on my 6800-Q4000, which are great.

I'm not sure that Nvidia really wants to ruin OpenGL driver for games. Some important games still use OpenGL and they're benchmarked in places like Toms Hardware, who sell a lot of video cards. Having said that, of course Open GL game drivers are optimized differently, but, aside from a few features like overlays, most of what pro apps do is also done by games these days. What pro apps do that's unique (such as a lot of AA wireframe) shouldn't be a problem for newer game cards to handle with brute force. For example, I may have a scene with 3 mil polys, but am I really going to look at it all in wireframe in full detail? Of course not, I only have 2mil pixels in my screen. lol

 

btw, some info on overlays~

Overlays are as others have pointed out a bit like slides being overlayed on the screen for UI elements. These can be UI elements like menus, or brush cursors, etc. It works by saving/cutting-out the background render of geometry, overlaying your UI element, and then replacing the buffered background when the UI is removed. That prevents the scene from having to re-render the geometry each time a UI is overlaid. Basically any UI element that's quickly drawn and then moved or removed can be done well with overlays.

As i recall (not certain, but could check) game cards support something like 2 "cutouts" for such a purpose as menus and such, while Quadros support 8 or so.

Here's the thing though, for most ppl it shouldn't make a difference most of the time. For example, Artisan brushes benefit from overlays, but most ppl are probably only painting scenes they get a high FPS on anyways, so a redraw doesn't matter. why? Because most ppl want to rotate models frequently while painting so they're already optimizing display in most circumstances. Also, painting is either being faked on a 2D Z-plane (like Zbrush which is fast) or being done on a true 3D geometry, in which case the paint operation of something like raycasting done in SW, if done on heavy geomentry, will bog down your CPU probably before the GPU does.

So, I don't really see where Quadro should matter much except is a small number of rare circumstances. And if Nvidia wanted to those could probably be optimized for game cards too. Overall, I think Quadros have become a ripoff to perpetuate the "professional" business unit of NVidia, tying to justify "professional" features now that were actually neccessary several years ago but aren't rally anymore. The sooner we stop buying them and stop the farce, the sooner Nvidia will give up the sham and just sell one card with different drivers for all users.

 

That's very correct. Maya isn't really multi threaded for most GFX benchmarking purpose. Rendering is another matter of course. The core 2 duo starting at the E6600 CPU is the fastest on a per/core basis.

I'd be curious to see those too. I can do some, but your CPU is faster which should matter more. speaking of that, I need to upgrade to a 6600 soon as this issue is decided!

great

Use Rivatuner to downclock. For shadows, just enable shadows on a light, and then in the lighting menu in hotbox.

I'm leaning to think they don't exist except in very limited applications making heavy use of overlays, which is rare, and which can be worked around, in drivers and in workflow choices.

Agree. Certainly Quadro used to be better, but these days I think it's a sham perpetuated by the "professional" division and Nvidia's business model. Quadros are massivly profitable afterall. For software makers, requiring a $2000 card actually hurts thier business model.

Thanks, this info is really helpful to the community!

 

Gosh! I love this thread! :rock:

@augie

I'm sorry probably I made a mistake. Did you update your specs or did I read someone else? (I tought you had a P4 3.0 ghz or something alike). I wouldn't have even imagined solution "2" otherwise. The E6600 is great, good choice!

Well this is a bargain. In Italy I can find a 6800U on ebay for 190 euros new, and that's NOT a good price. While a 7900 GT goes for 330 something...

No I'm not bubbling about anything. I clearly stated that a strong geforce is better than a real quadro for maya. The exception is the 6800 -> Fx4000 because:
1- At least in Italy its a lot cheaper
2- Is an overall quite powerful card
3- You would want to use also other software (at least compositing and possibly mechanical etc) in which a quadro could prove a benefit (we're ALWAYS talking of a softquadro )
4(possible)- If some graphic glitches/problems arise in new versions of maya it could be very likey that they would be fixed for quadro drivers and not geforce. Because nvidia should care (mission critical?). Never happened till now AFAIK though

A test on hardware render/particles would be welcome! There's a major rumor about quadros providing a "BIG" benefit with these features. I, personally, have never spotted a big difference but I never made careful tests with this.

And, by the way, do you get the dark gray background on the timeline as in my attachment or not (with 7900)?

@Samiam

I, of course, agree. Everybody has his own workflow and, unfortunately, I don't happen to use normal maps very often for my work (in this real time fashion).
I'm not sure what you mean with 2-sided viewing though.

YES! Precisely what I would like to see too. And I will tell you more! Try to open a scene from specapc for maya (ex werewolf) and just press play reading the fps count while it goes looping: with both my cards I get a REALLY inferior framerate compared to that of the same scene run through MEL script.

Well mine is just a sensation, I normally would have expected the 7900GT to crush the 6800 quadro or whatnot (IN MAYA). Also the pro opengl stuff is different from what you expect by a gaming opengl driver (and thus an opengl driver could be good for games and crap for pro apps)... but as you stated... it's even true that on modern engines these features are somehow converging. I would be curious to see if a geforce can match an equivalent quadro in, for instance, ProEngineer since this software seems to break the balls a lot with opengl pro hw features ultra reliable and blahblahblah...(I hate engineers )

As i recall (not certain, but could check) game cards support something like 2 "cutouts" for such a purpose as menus and such, while Quadros support 8 or so.

Interesting I didn't know this. The firegl control panel calls the planes: double buffered 8 bit overlay planes. Don't know how these store data...
I will disable the planes and try to spot some difference when I have the viewport cluttered with many menus.

So far I've learned that Maya is even more cpu bound than I tought!
In the past, at a studio, I had experienced a strong improvement in paintfx performance (I was working on the vegetation of a valley for a video) when they replaced a 5900 Ultra with a Quadro FX3000 on a custom assembled dual cpu machine: I tought it was due to the quadro features (that's why I stated that if you use paintfx alot you will be more productive) but from I heard from you it seems that it depended on something else...
one reason more to go geforce!

Greetings,
Anax

 

I agree the 6800->Q4000 is a good choice. That's what Ihave afterall.

But, I think beginning with the 7800 and up, game cards are probably the way to go. That's especially true if availability of 6800 is an issue, if someone doesn't want to deal with modding, or if one also plays games.

Also, I think it's a good idea to pressure software makers to better support game cards. There's not really an excuse for "professional" cards anymore that are really just drivers but cost $10x.

 

Right, that's all the UI overhead and perhaps memory fragmentation.

There is the OpenGL standard and it has some calls mainly for professional use, like overlays, but most are totally universal. 95% or more of the functions a pro uses are the same as a gamer.

Keep in mind the Quadro is the same hardware, just different drivers. Driver development can't justify the Quadro price, or even close. Even if there is a little extra silicon for few "pro" features, still no justification for the price. I'd rather pay the Maya engineers for better software features than a NVidia shareholder for marketing fluff.

I agree if someone is a heavy PaintFX user they'll get more from overlays. If someone really needs overlays then Nvidia has them over a barrel and can charge extortion prices, as is thier current business model. But AFAICT that's a really small percentage of maya users. I hope Autodesk recognizes $2000 GPU isn't compatible with thier price point and business model anymore, and moves towards more general card support.