As I had some serious issues with the stability of Houdini with my GTX 480 which persisted through all Drivers and all Houdini builds I decided to update the grafic card.

After going through all the posts which handle graphics and related issues i decided to go for a Quadro k4000, which arrived two days ago. The good news is it seems to run more stable than the GTX 480. The performance though is nothing but disapointment.

On geometry the GTX outperforms the Quadro almost by a factor of two an on volumes it is even worst, it outperforms the Quadro almost by the factor of three. On particles an fluids I did not test because here the GTX would constantly bluescreen my machine.

First I thought there is an issue with the Quadro but after running some benchmarks and testing with other apps I found that it is performing very well. In 3ds Max I get more then twice the speed with the Quadro.

I tested on Windows 7 and on Ubuntu 12.04, the results are the same.

I do not understand how a more than 3 year old gaming card which cost me less then 200.- Euro can can outperform one of the latest workstation graphic card for almost 850.- Euro.

I realy hope I am doing somthing wrong and stupid here and there and that there is a solution I didn´t think about.

Cheers

It could be fill-rate limited. The 480 has 48 ROPs (Render OutPuts, which are fixed-function units responsible for committing pixels to the framebuffer and doing depth testing). This gives it 33.6Gpixels/sec fill rate. The K4000 has half the number of ROPs, 24, giving 19.4GP/s. It doesn't tell the whole story, as this only accounts for roughly half the frame rate reduction. The 480 also has a slight floating-point rate edge (1.35GFLOPs vs. 1.25GLOPS). The different in shader architecture could also play a significant role.

If you simplify the shader by turning off lighting in the first case, does it help at all? (the "No Lighting icon on the right toolbar, near the top). I'm not suggesting this as a workaround, but as part of investigating the issue. While I do have a GEForce 480 available for testing, I don't have a K4000 (or close equivalent).

Hi Twod,

I'm glad someone of the developers takes care of this issue though wonder a bit that you don't have a Quadro k4000 at hand as I thought you are provided by nvidia with the lates hardware for testing.

I just tried to turn of the lighting but there is hardly any difference. The framerate goes from 4.6 to 4.8 on the geometry file. With pyro there is no difference at all.

Cheers

We do have a K4000 for testing, I just don't have one personally. I'd have to steal it from a co-worker's workstation But I'll see if they can give it a test on the scene.

If No Lighting doesn't change the draw time much, then the card isn't limited by shading computation, so it likely is fill rate or vertex-processing limited.

The Copy SOP is doing an X-ray pass to show the guide geometry through the Copy SOPs result, which might be affecting the draw time with multiple passes. If you append a Null SOP and display that you'll likely get a better draw time.

Just tried to add a null and put the display flag on it, but still the same, 4.6fps.

What you mention sounds to me like limitations affecting the geometry draw times, but would this also explain the reduced performance on pyro effects? Here the GTX is almost 3 times faster.

I think the volume case is strictly fill rate limited. The volume shader is dead-simple and so is the geometry, but it draws a lot of parallel planes with blending. The K4000 has better texture sampling than the 480, so I don't believe it is the limiting factor, unless the texture sampling and framebuffer writes are competing for memory bandwidth and being limited there.

You can try reducing the Volume Quality to Low in the Geometry tab of the display options, and also the Max 3D texture size in the Texture tab (say, to 128).

I already did all this while trying to get the GTX 480 to work properly some time ago :wink: Houdini seams to be very picky on graphic cards.

Just for curiosity I did some testing with the Quadro. I switched the Houdini viewport to wireframe and get 6.2 fps on the geometry file with 8 million polys. Now I set up the same scene in Maya and 3DS Max. In Maya I get 58 fps and in Max it is 91 fps.

Not that you get that wrong, I consider Houdini by far the best 3D package out there but would of course love to see the same numbers for its viewport.

Just for curiosity I did some testing with the Quadro. I switched the Houdini viewport to wireframe and get 6.2 fps on the geometry file with 8 million polys. Now I set up the same scene in Maya and 3DS Max. In Maya I get 58 fps and in Max it is 91 fps.

Not that you get that wrong, I consider Houdini by far the best 3D package out there but would of course love to see the same numbers for its viewport.

misterluqman
Just for curiosity I did some testing with the Quadro. I switched the Houdini viewport to wireframe and get 6.2 fps on the geometry file with 8 million polys. Now I set up the same scene in Maya and 3DS Max. In Maya I get 58 fps and in Max it is 91 fps.

Not that you get that wrong, I consider Houdini by far the best 3D package out there but would of course love to see the same numbers for its viewport.

I also agree. IMHO pro grade cards are for really very specific type of works, lik eif you need opencl but not with half floatoing point accuracy for scientific research for example.

I heard so many stories like yours i myself did not go with a quadro or firepro but went and purchased a 3 GB 7950 and shot my self again in the foot.

It turns out this card HD7950 sucks in Houdini big time,

I am not joking but a meshed sea surface with waves 3.5 mil poly only , on default viewport in maya 3.1-4 fps (its not acceptable but for a consumer card its ok) and then we jump into viewport 2.0 powered by dx11 and i get between 80 and 115 fps can you beleive it. VP 2.0 is limited but it gets its update slowly, and from what I saw at the technology preview at annaheim itr seems like it will be the new default of maya soon……

now the same scene in Houdini the card perfroms in GL3.2:1fps and H11 viewport gives me actually 7fps but no shading just black object.

Nvidia However works fine as far as I know.
So whats the deal here.

So in default opengl modes maya same card is 3 times faster than houdini, but Nvidia performs similar between the applications.

I can never replace Houdini with Maya, just when I finally converted from maya 9 months ago .. please fix the viewport performance, apparently even a last gen quadro suffers.

Wireframe and the wire-shaded modes use a geometry shader, which seems to substantially affect performance in the case of large models. It's a tradeoff made to improve update performance.

There is a performance improvement for AMD cards and large models in the next major version of Houdini.

tricecold
It turns out this card HD7950 sucks in Houdini big time,

I am not joking but a meshed sea surface with waves 3.5 mil poly only , on default viewport in maya 3.1-4 fps (its not acceptable but for a consumer card its ok) and then we jump into viewport 2.0 powered by dx11 and i get between 80 and 115 fps can you beleive it. VP 2.0 is limited but it gets its update slowly, and from what I saw at the technology preview at annaheim itr seems like it will be the new default of maya soon……

now the same scene in Houdini the card perfroms in GL3.2:1fps and H11 viewport gives me actually 7fps but no shading just black object.

Nvidia However works fine as far as I know.
So whats the deal here.

So in default opengl modes maya same card is 3 times faster than houdini, but Nvidia performs similar between the applications.

I can never replace Houdini with Maya, just when I finally converted from maya 9 months ago .. please fix the viewport performance, apparently even a last gen quadro suffers.

Can you share your test scene? I recently tested the exact same 25million polys model in Maya 2014 and Houdini and Houdini was better on AMD 7950. This is on Os X.


Edit - this is the test; check the last posts and screenshots I made in the link below. I'm sure I'm doing something wrong as everyone else seems to say that Maya's viewport 2.0 is far superior to Houdini. Let me know what I'm doing wrong.

http://www.sidefx.com/index.php?option=com_forum&Itemid=172&page=viewtopic&t=29138&highlight=7950 [sidefx.com]

Edit 2: A very quick read of Maya 2.0 viewport shows its generating proxies on the fly? Can it be written for Houdini/ any downsides to it? http://images.autodesk.com/adsk/files/viewport_2_0_api_gold.pdf [images.autodesk.com]

twod
Wireframe and the wire-shaded modes use a geometry shader, which seems to substantially affect performance in the case of large models. It's a tradeoff made to improve update performance.

There is a performance improvement for AMD cards and large models in the next major version of Houdini.

Could you give us a hint when that might be ? :wink:

I'm afraid I can't comment on release dates, sorry. The AMD fix was complicated enough that I was leery of backporting it. However, as it is tested in our development version, I may consider a backport once I'm confident it is stable.

Sounds like good news for AMD owners.

Could you reproduce the Quadro issue, or should I give up hope and maybe go for a 7970?

Yep, I can reproduce the Quadro slowness with large models in wireframe. There seems to be a sudden decrease in performance past a certain mesh size when using a geometry shader (which is how wireframe is drawn, by using the original shaded mesh). I've tried a few things to get it running faster, but it seems to be the mere presence of the geometry shader itself that is causing the slowdown. Straight wireframes (computed on the CPU from the polygon mesh) are faster, though cause slower update speeds due to the additional work to generate the connectivity.

The FirePro W8000 shows similar performance characteristics, so I wouldn't run out to buy a 7970, as I don't believe this would resolve it.

A small update on the wireframe issue - I've improved its performance by roughly 75% on Nvidia cards, and nearly double (100%) on AMD cards. For my K5000, I'm now getting a 10M quad grid running in wireframe @19fps.

This will be in H13 only, but it'll be out relatively soon [vimeo.com].

whoah! Os X/GL2.1 version too?

Edit: Always wondered if its normal for Scene level to have different fps/performance than Geometry level? Houdini seems to slow down to process when diving to the Geometry level on a 10mil test scene.

Pure Happynesssss!!!!

MartybNz
whoah! Os X/GL2.1 version too?

No, just GL3. GL2 doesn't use geometry shaders and so wasn't affected by the fix.

Edit: Always wondered if its normal for Scene level to have different fps/performance than Geometry level? Houdini seems to slow down to process when diving to the Geometry level on a 10mil test scene.

The SOP level often displays selection, which is additional display work. Especially primitive selection on GL2, which forces promotion of shared points to vertices because per-primitive vertex attributes are not supported by OpenGL. So for a 10M quad polygon mesh, the presence of the primitive selection cause the mesh to use 40M GL vertices instead of 10M GL vertices. The result is a lot more vertex shader work. GL3 uses geometry shaders to index the selection instead, keeping the GL mesh at 10M vertices.

The SOP level display also defaults to wire a over a shaded, which is a bit slower to display than straight shaded.