CoolColJ's test pics thread

[CoolColJ]

cool stuff, just need a star background, or maybe model each individual star

I returned my faulty 6GB memory kit to the store and got a brand new one. Now each kit will run at 1500+mhtz with CPU at 4.09ghz stable without a BSOD in sight, and CPU voltage at 1.25volts. It runs pretty cool like this under full load, barely cracks 60 degrees with ambient of 18. So this bodes well for summer when it will be 10 degrees warmer. I still can't believe after having Indigo render for 4 hours straight that the heatsink is barely warm! This Noctua NH-D14 is very effective!
Might be able to reduce voltage further with more testing. Back when I was using my faulty memory kit and 1.3 volts, I was cracking 70, with lower ambient temperture....

With both kits in for 12 gigs, It won't run past the post screen at 1500+ mhz though... so I had to knock the memory mulitplier from 8x to 6x, and now the memory is at 1180+. But no BSODs in windows, and voltages are the same all round. Still higher than stock memory speed of 1066 though.

I see Corsair now sell a matched 12 GB kit of the same memory that is rated to run at 1600mhtz with all 6 sticks installed..... ugh.... well at least I have both for $150 cheaper, and that extra memory speed doesn't really make that much difference from what I see in the real world, apart from benchmarking. An extra 5k samples at most. I could probably run it higher with more voltage on the QPI/VTT and CPU, but that will bring more heat.

Even a 24 GB kit
http://www.corsair.com/products/corei7/default.aspx


edit - running at the slower memory speed allowed me to tighten the timings down to 7-7-7-20

[CoolColJ]

I mentioned before that I thought the default Indigo settings for MLT were only good for previews (Large Mutation probability 0.4, Max Change 0.01), and since they smooth out fairly quickly, but if you wanted accuracy you had to make the Max Change value smaller.

I found a Max Change of 0.003 and under takes way too long to smooth out, but 0.005 worked well.

So to put my theory to the test I did 2 Bidirectional MLT renders, one using default settings, and another with with my values of Large Mutation probability 0.5, Max Change 0.005.
I let them render for 5 hours each to see if they would converge to the same image, and it looks like they won't

The default settings don't give you accurate details. Look in the oven and in the glass doored cupboards, plus reflections/caustics in other areas. There is one highlight on the right glass cupboard that didn't show up.
I also had Aperture Diffraction on with Dag's obstacle map posted earlier

[CTZn]

Yes... where !? You successfully cut an hair in 4 along its lenght (also called nitpicking)

Seriously the differences are small, and whereas the default is showing perhaps less features (relatively to the perceptible differences) they are also better defined. With a smaller max_change you are invoking more details but they would need also more time to smooth up.

You should test n different kinds of scenes to generalize your assertion (int, ext, scale etc), because when it comes to tweaking MLT you know it is scene dependent (specially max_change). A large nature scene with trees or even larger structures might not benefit from the tweak.

At this point it's pretty subjective - you may be right though

[CoolColJ]

I dunno the difference are pretty obvious to me

Off course I'm talking about things that need MLT, if you have an outdoor scene, grass and trees and what not, then basic path tracing will do the trick just fine and faster.

I first noticed this in my Water tests scenes, plenty of MLT needed here. Smaller max change lit the under water objects way better. Caustics were also brighter. With the larger Max change values, you'll find small thin reflections and caustics will be somewhat fainter and not as well developed. Small specular highlights on glass things tend to take way longer to show up as well, or may not show up at all.

[CTZn]

I dunno the difference are pretty obvious to me

That's the point, they are not obvious to me, though I tried hard. Apart that and for what I know, you are correct in saying that a smaller m_c resolves small details better. Above all, your - significative - experience prevails, CoolColJ !

[CoolColJ]

Look in the oven, pretty big difference there

[CTZn]

Aaalright, that's something I missed... but this was unfair, only you have the 1:1 reference

[CoolColJ]

that's the size I rendered both of them, 800x600, just make sure your browser isn't shrinking the image down

[CTZn]

[Silmä]

Here's the difference as seen by Photoshop. Brightness is ramped all the way up.

I wouldn't say there's too much difference. The oven and that aperture diffraction bloom are the biggest differences, otherwise it's probably more due to jpeg-compression than indigo.

[CoolColJ]

Like I said, only in areas where you need MLT.
They still didn't converge to the same results, so settings do matter. If you do a render of a pool or lots of glass and caustics etc, where MLT is extensively needed, then you'll see a big difference. I have with all my tests

[CoolColJ]

Just another new path tracing code test
Good old "Arnold" style render for those who remeber
Used to take a whole night to do em, now just a few minutes in the 10 years since http://www.3dluvr.com/marcosss/

And rerendered one of my old Sketchup scenes. Took a long time to get clean!

[Godzilla]

Just another new path tracing code test
Good old "Arnold" style render for those who remeber
Used to take a whole night to do em, now just a few minutes in the 10 years since http://www.3dluvr.com/marcosss/

Rendering certainly has come a long, long way.

Took a long time to get clean!

You talk about this render like it went to a rehab facility.

[CoolColJ]

Testing out Indigo 2.4.9's better MLT. It feels faster, and Samples per pixel rate is higher now, but the under water displaced object is still painfully slow to light up!
The normal flat surfaces converged after 30 mins though

4+ hours of Bidrectional MLT

Straight MLT - 2 hours or so, worked pretty well without bidirection. Hardly splotches now and they dissappear after a while.

Check the differences in samples per second

[Stromberg]

I just love this topic, i hope you get the time to update it soon
Or maybe you had a rendering going from your last post let's hope so