Indigo Renderer GPU acceleration

[Pibuz]

actually we should buy one for you, considering what amazing work you do with Indigo!

I'd consider that :D (there is no mr green smiley available, but he'd fit perfectly here)

[Pibuz]

...btw I have a 8600GTS here at work and a 8800GTX at home. Maybe this one is enough too... I hope!
I am considering buying a GTX480 btw. I read wonderful comments and I know it is very efficient rendering-wise (Octane and Arion forums)

[CoolColJ]

The top priority for now is to improve OpenCL performance in general to the level we're seeing with CUDA. Even on NVIDIA cards we're not getting the same performance using OpenCL as with CUDA, and with ATI we have architectural differences to contend with too.

How do the ATI cards compare to Nvidia ones under OpenCL?

[lycium]

Right now, not very well. Better performance should be possible, but the newness of OpenCL as a platform is readily evident on NVIDIA cards as well: CUDA is much faster than OpenCL doing almost exactly the same work on exactly the same card and drivers. This has been noticed by other developers as well, most of whom just leave OpenCL support behind. We want to push on with support for this platform, and I much respect Chaos Group for their decision to OpenCL as well - this is a good thing for all of us since it provides incentive for better driver support!

The initial phase of the beta will focus on testing our CUDA codepath on Windows since that's the most mature right now; we're almost ready to distribute a build in this configuration, though we still have some important things to finish off. I'm just busy trying to get it to run on my GTX 470 at home (at 1:30am hehe), and it requires curiously much TLC to get working! Today we had another 10% performance increase as well.

We'll be sending out emails with more info to the beta testers as soon as these matters are dealt with; thanks for your patience!

[Russdigo]

The top priority for now is to improve OpenCL performance in general to the level we're seeing with CUDA. Even on NVIDIA cards we're not getting the same performance using OpenCL as with CUDA, and with ATI we have architectural differences to contend with too.

How do the ATI cards compare to Nvidia ones under OpenCL?

I've been holding off a graphics card update for years now. Nothing I use really would get much out of the boost (I'm no gamer and my photoshop performance is acceptable). My motherboards ATI X300 is fine... That is until now. I'd like to go ATI if possible, one look at the power requirements of the Nivida cards has me convinced I'd rather build a rendering machine than a 3 bar electric fire!

GeForce GTX 470 - £292.69 (via dabs.com) - 215 Watts TDP

GeForce GTX 480 - £409.99 (via dabs.com) - 250 Watts TDP

Mobility Radeon HD 5850 - £234.87 (via dabs.com) - 39 Watts TDP

Mobility Radeon HD 5870 - £321.12 (via dabs.com) - 50 Watts TDP

Will be interesting to see how they stack up best bang for the buck and best most efficient for the money.

[Russdigo]

In a years time we will see how the market has sorted itself out.

The math and feature sets balance will get interesting.

Buyers will start to sum up the whole package. For example its the price of the software plus the cost of the card to make the whole rendering package balanced against performance and weighed against the feature sets.

Indigo cost + ATI or Nvidia card

Octane + Nvidia card

Etc..

[lycium]

Fermi isn't a graphics card, it's a supercomputer which happens to have OpenGL and Direct3D drivers That's why it's no wonder it's getting spanked as far as gaming is concerned! Of course, properly used, R800 is an absolute monster with way higher peak floating point performance... each arch has its strengths.

BTW, I'm pretty close to getting the 470 working properly with Indigo now; it's looking crazy fast already, and I still have some options to tweak for much higher performance.


Russdigo: the other part of that performance picture is how long it takes to get a clean image

[Russdigo]

Russdigo: the other part of that performance picture is how long it takes to get a clean image :)

Sure. I would like to weigh up how long its takes to get to a clean image against the cost you pay for the tech which will now include the graphics card.

Personally I do many jobs that don't allow me to optimise and faff around with my scene. In fact this is one of the reasons I like Unbias renderers. If I have to juggle textures and polygons to get it to fit in my graphics memory like Octane I'm not going to be happy. I like how Indigo will use ATI and Nvidia. I hope the team will concentrate on work that is too big for its rivals to tackle without resorting to procedural textures and polygon culling.

[lycium]

Indigo should be very memory friendly in terms of how much data you can load, since we don't store the final image and a bunch of other things on the GPU. Plus we have support for instancing, allowing the virtual triangle count to easily scale into the trillions

I've gotten the 470 to work properly with Indigo; here's a 13 minute render of ENSLAVER's teapot vortex scene, rendered at 1920x1200 on my i7 920 + GTX 470 box at home: http://dl.dropbox.com/u/3038174/teapot_ ... 13mins.jpg

[neo0.]

Did anyone else get their email?

[dakiru]

Did anyone else get their email?

Patience, buddy! I guess Glare has a lot work to do now and that's good. :)

[CTZn]

Did anyone else get their email?

Patience, buddy! I guess Glare has a lot work to do now and that's good. :)

Affirmative, my concern is wether they're having enough sleep !

Please bear in mind that Indigo has more to offer than a brute and short path tracing (e: generally speaking). There's not one button to press to have a quality product out in this field, but... many keyboards already I guess haha !

The time involved in this strive for quality is meaningfull, neo0.... that's quite the paradigm of unbiasedness.

[Zom-B]

Better wait a extra week for a "usable" beta, then having a bugged one that does problems today!

[StompinTom]

Indigo should be very memory friendly in terms of how much data you can load, since we don't store the final image and a bunch of other things on the GPU. Plus we have support for instancing, allowing the virtual triangle count to easily scale into the trillions :)

I've gotten the 470 to work properly with Indigo; here's a 13 minute render of ENSLAVER's teapot vortex scene, rendered at 1920x1200 on my i7 920 + GTX 470 box at home: http://dl.dropbox.com/u/3038174/teapot_ ... 13mins.jpg

Gawdamn, that's some sweet 'n' smooth motionblur!

[Whaat]

I would like to see some test results from a complex sun-lit interior render. (I don't render blurry teapots very often...) ;)