The collected works of Brombear - Page 2

Since quite a lot things happened between the last post and now I thought it could be a good idea to give an update what we are doing now. This week I have been at fujitsu-siemens' inhouse-exhibition visit showing what we can do in realtime now. The workstation is the new v840 (opteron II based) equipped with two fx4500, the powerwall is driven in passive stereo with 2 fx5500 (for photos we switched to mono of course).

The model you see has approx 5.5 million polygons and is completely covered with shading materials. Lighting is completely done with IBL in addition with dirt mapping. The dirt mapping is precalculated and takes about 15-30 minutes on a multiprocessor workstation (depending on your experience) It renders on a single g70 gfx card with approx 10fps (depending on resolution) and on the new g80 cards with more than 20fps (way to go nvidia).









Regards

Matthias

Well, thanks but I guess that one goes to NVIDIA, the g80 series has endless shading power. Looks like we have to think about some new ways to burn it

Matthias

g70 = consumer 7800/7900 series (or quadro 4500/5500)
g80 = consumer 8800gts/8800gtx (at the moment). Since the GTX card is barely available, we did test it on the gts and imagine my face when I saw the model exactly followed every mouse gesture. I can't wait until I can test it on a gtx resp. the next generation of quadro cards with 1.5gb video memory that may be available in the first quarter of 2007



Another term for dirt mapping is ambient occlusion ( see here http://en.wikipedia.org/wiki/Ambient_occlusion ).

Matthias

Have made some additional shots within the interior. There are almost no texture coordinates used (except typical things like MMI interface and logos), everything is done with shaders. Reason is to be able to prepare a review session in less than a day once the geometry has been properly converted from the original CAD data. Most shaders (leather, rough plastic) have bump structures too. The dirt mapping I mentioned earlier has much more effect within the interior too, without it everything would look really flat and synthetic









Matthias

This is a special feature of the chinese edition ...



Not yet, but we are working on it Interestingly we heard that comment quite a lot during the exhibition. Seems even if the visualization isn't physicaly correct it is more than plausible for joe average ...

Next stop ... realtime raytracing

Matthias

no chance to get this running on sgi iron ... even ati cards fail for various reasons (no hdr filtering yet, gl state trashing, weird optical results and last but not least undefined crashes).

Matthias

Squeen,

your welcome, positive comments from people working in the same domain are always a special pleasure.

So, decided to put up a new image. If anyone here is at the hannover fair in germany next week, you could see our software (and me) live at the PNY boost (Hall 17 F60). We are showing the car within the picture on a 4 segment mono powerwall driven by a cluster of 4 pc with quadro FX4600 (FX5600 not yet available). Special features are image quality (high antialiasing quality in combination with our material shaders) and speed.



Matthias

Hey, that looks really nice. If you can limit your scenario so it fits, it will be a good solution.

For all others only realtime raytracing will be the solution. But the requirements to code it yourself are completely different:

- SIMD coding (adieu readability)
- acceleration structures (balance between updatebility, memory usage and speed)
- fast materials (phong is easy, texture filtering sucks)
- reading and understanding all newest papers on the net (raytracing is still a very hot research topic)
- performant multithreading
- dealing with jerky compilers (all of them suck, each one has his special areas where it sucks most)
- clustering (8 cores is simply not enough for a typical 1280x1024 resolution)

Matthias

Hi there,

just wanted to give a small update that we are still alive and founded a new company together with some 3D guys from berlin. Our new name is PI-VR GmbH and you'll find our website at http://www.pi-vr.de We are now also targeting automobil photographers and the first results can be seen within our gallery.

In the meantime we have extended our raytracer a lot, added some nice features to the OpenGL renderer (geometry lights, ultra high anti-aliasing), added a CAD interface and some things for marketing production rendering ( rendering layers, automated generation of images, etc.) and various other things that I might have forgotten. Even the ATI part of our OpenGL rendering does work now and the new FireGL 7700/8700 rocks like nothing else.

We also have a new product called VservR which is an online server based realtime 3D configurator.

So even if this is good news, there are some bad to share also. Nothing of this will work on your O2 nor your Octane/Fuel and not even on any kind of Onyx

Matthias

Some time has passed by and I would like to invite those that attend the hannover fair (20-24.04.2009 in hannover/germany) to join us at the shared booth of pny/nvidia/3dconnexion/cad.de (hall 17 booth F60)

We will showcase a very detailed model of a car consisting of 60 million polygons with a development version showing our global illumination implementation. Some images may be seen at our website already.

Regards

Matthias

A very good question ... coming from the visualization neadnertal times on onyxes with 120.000 polygons I often ask this myself. But our clients always see potential for improvement. It's just a question how near you want to go.

Speaking of limits ... we just have the limits of the main memory ... current nehalem workstation hardware supports up to 192GB. Our biggest model was 550 Million Polygons using 54GB main memory (all geometry, textures and hdr environments), so if I simply interpolate we should be able to visualize somewhere between 1.5 and 2 billion polygons ... should be enough for airplanes ...

but not enough for industrial complexes like a steel work ... I guess you would need another 10-100-1000x increase in memory

Speaking of photos ... most of the time the car is not there as physical prototype ... too early in engineering process and of course the rising variety of possible configurations of a car or whatever.

Hello Thomas,

thanks for the compliments, we really work hard for it.

Regards

Matthias

Since a few months have passed since my last post I would like to give an update.

We are about 2 weeks away from our 4.0 release with the major feature of global illumination which adds quite a lot of realism to interior sceneries. Other than that lots of preparing functionality (selection modi etc etc) so model preparation only takes few days. We have tested our software with big models and found out that 550 million polygons do not slow down raytracing. They consume somewhere around 55GB of main memory so there is quite a lot of room on todays workstations. Nehalem/icore 7 gave us approx 50% more performance.

@Squeen you have been somewhat silent about your project ... hope that you are still improving ... I would like to hear more facts about scaling on your ubercluster. Biggest we tried in the meantime was somewhere around 130 cores ... still way too slow to drive the Sony 4K projector @15hz

Matthias

PS: Images can been seen on our website at http://www.pi-vr.de

Hi Squeen,

very cool ... good that they forgot to sum up how much energy has been used to calculate those images

Are you continuing the hunt for faster rays and pixels ? You could try out that caustic thingy or play around with optix ...

Regards

Matthias

I just wanted to post a new picture that was produced today, I hope you like it. Click on Image for full HD resolution

Hi Squeen,

quite a lot of things to do, but hey that's what the fun is all about


Sorry, I didn't get that ?



1) will of interest for us too, especially for dielectrics
2) too
3) we did that alrey
4) BRDF is not enough for us ... we need real materials like metalic paints, leather, brushed aluminium etc. I think we will disclose some in the very near future.



24 core will be way too slow even only for a single side of the cave. We tested 200 cores for the sony 4K projector (8 MegaPixel) and it's still too slow.



Yes realtime graphics (with good quality) really change the way of working in a lot of fields. That is the reason most people can't think of anyone else once they have been infected with the realtime virus

Nice image, dielectrics are always visually attractive.

Regards

Matthias

Thanx for the kind words.

no photon map and no irradiance caching yet ... everything brute force with some gehirnschmalz as we germans say (my way of saying that even I do not understand everything myself) .... acording to my collegue we would really need some time to fire up vtune and friends and do some serious tuning session ... there should be some room for improvements.

I really love 3d motion blur ... this changes the look soo much ... and you never want to go back to 2d hacks again.

I dunno how often everything got refined during the last 2 years of development of the tracer, but also a whole lot of time was invested into the material system, the best tracer is worth nothing if you make it too powerful and complicated for your users to use it (that's the reason I would never achieve anything out of a regular dcc tool)

Regarding leadership we really don't care a fuss about, there's so many new selfdeclared leaders these days. I don't post on ompf anymore (even forgot my login) and after those useless bashing wars during the last weeks I am not unhappy about it.

In the end its all about having fun about what you do and having people that appreciate your work ... and best if you can make a living out of it

Have to add a small correction. The rendering uses 512 adaptive samples per pixel, but the computation time is correct. When using 256 samples you'll get some noise on the headlights and rims while rendering time goes down to 15 minutes.

We try to make some more benchmarks in the near future on various platforms (1way up to 8way systems) to see how it scales. Last test on 32core system in 2008 showed ~85% efficiency per core.

Matthias

Again an update what the current version can do on maxed out workstation hardware. Click on Image for full HD resolution



Forecast to version 4.2 (end Q1/2010)

- much better quality with motion blur
- polygons ... who needs polygons anymore ....

Sorry to bother you, but its 2x QuadCore = 8 cores + Hyperthreading. Hyperthreading adds another 22% speed increase which is very good according to intel. The mentioned time is of course with activated hyperthreading

About the noise I don't know myself (sampling is beyond my programming knowledge) but as long as its working I don't need to

I really would like to run our software on your cluster though ...

Following recommendations:

Master running on any kind of 64bit Windows OS.
Slaves running on 64bit Windows or 64bit Linux
fast and reliable tcp/ip based network between master and slaves. To be able to use multicast for initial spreading of the data the network topology should be very flat (one hop, maybe two ?)

Possible ?