Mega Machine - Slow rendering?

[thevisad]

Hey guys, just some questions as far as what I should expect to see with generations.

I have the following machine

Dual, Quad Core Xeon 2.0 ghz nehalem chips
12 gigs of ram
3 x 15000 RPM sas harddrives - single disk mode
Geforce GTX 295
Windows 7 - 64


I upgraded from a Dual core 3 ghz, 4 gigs of ram and a geforce gt 8800

After the upgrade, I can tell things are definitely a big faster. But considering the major upgrade I would expect this to dump maps a lot faster. I took a map dump from http://srtm.csi.cgiar.org/ for michigan (256m picture) and generated the map, 6016x6016. I then imported it to VTP, exported the BT and then began creating all of the maps for the system in L3DT. Now, I am sitting at 9 hours of generation, 30% of the Normals are done and it still has 4 hours left on the normals.

I also, forced 8 cores and synced the processes before running the generation.


Am I doing something wrong/missing soemthing or is this what I should expect. I mean, I realize the map is HUGE! I just assumed things would be a bit faster on this rig.


Thevisad

[thevisad]

Ok, I found that I had chosen and loaded aaron recommends on the Normal, generating a 8x texture file. I dropped it down to a 4x texture file and it generated the normal in 10 minutes, less 4 hours 50 minutes. That's more what I was expecting!

S

[Aaron]

Hi Thevisad,

The normal map for a 6k x 6k pixel heightfield at 8x resolution works out at a tad over 2 gigapixels. If I use an 4x attributes map, the estimated time for the normal map will be ~7 hours on my 3GHz quad-core. If I use an 8x attributes map with the same normal map settings, the calculation time for the normals map drops to ~1.8 hours. This is because there is a significant computational efficiency in generating normal, light, texture and alpha maps that the same size as the attributes map. This largely explains why, in your case, generating 4x normal map was so much faster than an 8x normal map; your attributes map was 4x resolution also.

I also, forced 8 cores and synced the processes before running the generation.

I think you will find the default settings are better. Synchronising tile threads makes the normal map calculation slower by ~13% on a quad core machine, and it would be worse on an 8 core machine. Forcing 8 cores is also unnecessary (though harmless), as L3DT should default to use all available cores*, which would be 8 in your case. If L3DT was not defaulting to 8 cores before you forced this value, please let me know, as that would constitute a bug.

Best regards,
Aaron.

* The core count is read from the process' thread affinity mask, which as far as I know is the accurate way to measure core counts on Windows systems. Forcing either more or fewer threads than the count given by the affinity mask will result in a degradation of performance. The default settings should therefore be the fastest.