www.kraytracing.com

I have done some tests with Kray AA and native Lw AA.
Kray is the fastest and lw is the slowest. What! ( i hear you say)
Yes i have also been concerned about how "slow" AA work's in kray and Grzegorz have also worked hard to improve it alot. So why is it so slow and why does i say it's fast? Mainly when we think of how AA "use" to behave we compare it to standard AA in Lw.
I have done tests there i have compared the time of pure AA between Kray and Lw. I have measured the time from when the frame is done and starts it's AA pass until it's done, both for Kray and Lw.
I have used Adaptive AA grid3 treshold 0.1 in Kray and in Lw adaptive AA classic Low (5pass) treshold 0.1. I used Montecarlo 4x12, noisereduction and one Bounce in Lw. In Kray, lightmapping, fg100/800 and 16bounces.
The testscene took me 3minutes 27 sec to finish in Kray but only 1 minute and 9 sec in Lw.
The differences between them is that lw's AA is optimized for one ray or bounce. Kray uses 2-2000 bounces.
If i raise the numbers of bounces in lw to 3, it takes the same time or longer than Kray to do the AA. If i raise the bounces to 8, it takes so long time that i couldnt wait for it to complete.
My conclusion is that Kray is relatively insensitive to the numbers of bounces used while lw is highly sensitive to the numbers of bounces. When i have compared AA in raytrace mode, Kray AA is faster than Lw AA.
Grzegorz is still not done with the AA algorithms so i guess we will see faster and better AA soon. I have hopes of that it's possible to "disconnect" the AA from the rendering completly and therefore gain big speedups but untill then i want to give some advices so you understand what slows down AA.

The more FG used the slower AA. One thing that could hog Kray totally is if there is very dense clouds of FG samples.

Try to spread the samples further apart rather than using even more FG to get rid of splotches/artifacts.

Adaptive AA can be very efficient together with "clean geometry" without textures,bumpmaps and reflections but this is rarely what most people render. For most scenes "Fullscreen grid3 AA + some undersampling like 16x16 provides a faster solution since it doesn't have to check anything, just render. FS grid3 AA works well for most things but not DOF. For Dof i recommend low value Stochastic AA instead.

The main reason for this is that when you're doing anti-aliasing in Kray, you don't have to fire more final gather rays. When you're rendering in monte carlo mode in Lightwave, every extra anti-aliasing sample will also increase the global illumination sampling. For this comparison to be more fair, you should turn off adaptive sampling in Lightwave, and reduce the number of rays per evaluation accordingly. No anti-aliasing and 6x18 rays per evaluation will produce about as much noise as enhanced medium anti-aliasing and 2x6 rays per evaluation (6x18 RPE * 1 AA pass = 108 samples per pixel; 2x6 RPE * 9 AA passes = 108 sampes per pixel).

Am I missing something here? Whenever I check FSAA (box/cone grid 3), AA-times go skyhigh...

Yes FS mean you render the same image many times (god for mblur and such) and that takes time. In the future we will implement smarter algorithms like Vektor blur.

That's hardly a fair comparison.
Using pure MC in Lw makes no sense. Nobody uses that and its the least powerful Lw radiosity mode....

except wrote:That's hardly a fair comparison.
Using pure MC in Lw makes no sense. Nobody uses that and its the least powerful Lw radiosity mode....

uhm guys, Silver's post was from Wed Mar 15 2006 (six).
Nothing to see, please move on

except wrote:That's hardly a fair comparison.
Using pure MC in Lw makes no sense. Nobody uses that and its the least powerful Lw radiosity mode....

IMHO Monte Carlo is the best GI. Once processors will be much more powerful than now, even games will use Monte Carlo GI. Interpolated/cached GI is just fake for current slow machines.

TrueArt Support wrote:[
IMHO Monte Carlo is the best GI. Once processors will be much more powerful than now, even games will use Monte Carlo GI. Interpolated/cached GI is just fake for current slow machines.

IMO Metropolis Light Transport is the best GI now. Seeing that I rendered my first raytraces over 20 years ago and yet no realtime raytracing in computer games now, my guess it will take at least 20 years before I see any monte carlo in realtime let alone realtime Maxwell
Maybe just before I die, but I think I don't care much about it by that time

I agree with Erwin, certainly now that Moore's law turns out to no longer be valid, I don't think we'll be increasing render speeds that much.
I'll be surprised if I have a 128-core machine on my desk in 4 years, and even if that's the case, a render in maxwell which now takes 24 hours then still takes over 10 minutes.
In order for something like that to render real time you'd need an increase in render speed of at least 2.6 million times. I don't see that happening in a long long time. Especially since HDTV resolutions are now becoming a standard.

According to Moore's law power doubles each 24 months.. So, after 4 years it has to be just 400% faster than now.. So, if currently dual quad-core are sold, after four years there should be dual 16-core. It's easily doable IMHO..

Aren't Intel going to be adding Hyperthreading to their current quad-core processors soon? Not quite an octo-core, I'll admit, but it will mean eight threads rather than four - and a certain speed increase.

I wonder how many cores they'll be able to fit on one chip though. They'll run into heat dissipation problems pretty rapidly.

Phil

TrueArt Support wrote:According to Moore's law power doubles each 24 months.. So, after 4 years it has to be just 400% faster than now.. So, if currently dual quad-core are sold, after four years there should be dual 16-core. It's easily doable IMHO..

Well, technically, Moore predicted that the number of components would double every two years, not necessarily that performance doubles.

There's also Writh's law that says : "Software gets slower, faster than hardware gets faster."

...

what?