Octane 2.3 final will not have MLT. Tweet

[matej]

Hm, I don't know... this news scares me a little, because it looks that GPU programming is a lot harder and (currently) limited than CPU programming. This will undoubtedly bring other limitations in the future.

Anyway, I hope for the best. This new algorithm could be a huge marketing asset, if it will deliver such power.

And thanks for the straight explanation.

[abstrax]

Hm, I don't know... this news scares me a little, because it looks that GPU programming is a lot harder and (currently) limited than CPU programming. This will undoubtedly bring other limitations in the future.

Anyway, I hope for the best. This new algorithm could be a huge marketing asset, if it will deliver such power.

And thanks for the straight explanation.

Yes, that is a correct observation. The reasones for the "hardness" are manifold:

- Limited flexibility of GPUs
- Limited memory bandwith between CPU and GPU
- Complex cache architectures on GPUs
- Immature APIs and standards
- Immature drivers
- Non-existent or immature algorithms for GPUs
...

BUT, fortunately all these things are not physical barriers and will be overcome over time. One little baby step after another...

Another thing that makes me looking forward to the future: GPUs raw power and the challange to make use of it, reactivated the interest of a lot of developers, who became bored with the same old, same old of CPU rendering. Not that CPU rendering is dead already, but due to its maturity, big jumps in quality and/or performance become more and more unlikely. But the GPU stuff is an exciting new world to conquer

Cheers,
Marcus

[radiance]

Hm, I don't know... this news scares me a little, because it looks that GPU programming is a lot harder and (currently) limited than CPU programming. This will undoubtedly bring other limitations in the future.

Anyway, I hope for the best. This new algorithm could be a huge marketing asset, if it will deliver such power.

And thanks for the straight explanation.

Hey,

It's not in fact that bad. The issue is very simple: parallelization.

GPUs require algorithms to be parallelized, and some algorithms are easy to parallelize, others aren't.
We tried to parallelize MLT, which has failed, so we developed a parallel algorithm instead.

That's the only reason. It's not about 'limits' or being 'harder'. GPUs speak russian, CPUs speak english.

A masterwork novel in russian might not be translatable to english and have the same spirit, emotion and philosophy remaining in it. Languages are not just forms of communication, but also have perceptual differences between cultures that speak them.

We tried to translate a masterwork russian novel to english, but we were not happy with the outcome, so we decided to write a new novel in english instead.

That's my non-technical explanation
Radiance

BTW: today we implemented 3 new things in octane, which would have taken 3 days to implement on a CPU, as GPUs have lots of neat little toys in them that are designed for graphics. (some of them are mentioned in the release announcement of the 3rd pre 2.3 coming tomorrow in the RC forum.

[pixelrush]

I must admit I am a little disappointed about MLT but the RLT sounds ok.
Can we see a pic that shows the result of your new algorithm?

[matej]

That's my non-technical explanation

I like it. My scepticism comes mainly from lack of knowledge on the subject, and what is unknown is usually scary...
OTOH, I have hope in the experienced team that is working on Octane.

Looking forward to test the new pre2.3...

[radiance]

I must admit I am a little disappointed about MLT but the RLT sounds ok.
Can we see a pic that shows the result of your new algorithm?

No, because it is currently disabled, the controls are disabled, etc, we can't deliver the next pre 2.3 release tomorrow if we have to reenable everything, and we're not happy shipping a build with the algorithm inside (although invisible/unuseable)
also while working on area lights we need to disable some stuff here and there, so we can work faster.

let's focus on getting tomorrow's release out and we'll show you something soon, promised.
There is not reason at all to worry.

Radiance

[pixie]

(information above is not octane specific, but applies to all CPU based unbiased engines out there too)

I've mistakenly saved this three pictures with completely meaningless names and now I wanted to know which mode did I pick, the only thing I know is that I used PathTracing, BiDir Path Tracing, BiDir MLT, I do know it took 8 minutes to render each, and that pretty much the noise has scaled at the same rate and that the scene was made of two direct lights, could please someone help me?

[jakchit]

Hey radiance,
I think most people here can understand your development setbacks and frustrations.
This technology is still in its early stages and you are in the trenches of its conception.

We all appreciate that Refractive is trying to create a great GPU renderer with known and tested technologies.
But, I think what people will also have to realize is that you guys are also trying to break the mold and develop
even more advanced technologies that will make Octane more than just any old GPU renderer.

I for one appreciate that you can see past these roadblocks and are working to bring us something new and better,
not just run with the pack.

Thank you
Jack

[rgigante]

Reading the thread (and speculating a bit) it sounds to me like something relying on SPPM. I could probably get wrong but this could be a plausible solution considering the fact that it's relatively easy to implement it on GPU.

R.

[andrian]

Hi all,


Since the new an algorithm we've developed is not based upon any solid scientific paper/concept, it might take a few pre releases or RCs to get it working perfectly, but the few weeks more have advantages, as the new algorithm is somewhat similar to bidirectional path tracing without fireflies, eg you can place an emitter in a glass lightbulb, covered by a diffuse transmission cloth lampcrown, and have beautifull caustics on your walls.


Yours,
Radiance

Ok, this rise the question, since the new algorithm is similar to bidirectional path tracing , will we be able to navigate our scene as we do now with path tracing?