TG2 Mac and OS X 10.6 Snow Leopard

[jo]

Hi,

Now OS X 10.6 is available I can give you some information about how TG2 Mac runs on it.



Please note: The edit field display problem described below is now fixed in Terragen 2.1, which is available now. See here for more information:

http://forums.planetside.co.uk/index.php?topic=8268.0

Scaling for threaded renders is now improved and using up to 16 threads can be worthwhile in some scenes now. OS X 10.6 is still markedly faster for rendering than OS X 10.5.

The Free edition with these improvements will not be available until January 2010, so I'm leaving this information about the edit field display workaround etc. for now.



First the bad news. The most important thing is that there is a problem with edit text field display on OS X 10.6. It looks as if the text has been shifted off to the left. Some edit text fields, those which are displaying numbers for example, might look as though the numbers have been clipped in half. In the worst cases, which are in things like the node name parameter in parameter views, the text may not be visible at all.

Fortunately there is a simple workaround for this, which is to simply click in the text field or tab to it. The text will then display correctly. You do need to watch out for this though because, for example, it can give the impression that there is no heightfield blending shader when in fact there is.

I have been able to work around this problem and it will be resolved in the next public release. I'm not sure when that will be as there are a number of big improvements which are being tested by alpha testers and we need to make sure everything is ok with those before we can make a release incorporating the fix for the edit text field issue on OS X 10.6. I apologise for the inconvenience in the meantime.

Now for the good news. OS X 10.6 is noticeably faster for rendering on multicore machines, and I would expect the improvements are greater the more cores you have. As an example the 3dspeedmachine.com TG2 benchmark is 21% faster on OS X 10.6 than on OS X 10.5, using the same v2.0.3.1 of TG2. That isn't due to Grand Central Dispatch or anything, more down to Apple fixing a performance problem in a low level OS function. The speed improvement will vary between scenes, but it's good to get a speed boost just from upgrading the OS. Upgrading to OS X 10.6 is certainly worth considering.

TG2 Mac still has scaling issues and isn't yet able to fully utilise multiple cores past 4 or so. However with OS X 10.6 the 3dspeedmachine.com benchmarks is at least a bit faster when rendered with 8 threads than with 4 threads. The good news is that I have a good understanding of where the major scaling problems are happening and it should be possible to resolve them relatively easily. Due to the tricky nature of multithreaded development we will have to do some thorough analysis and testing to make sure everything is working as it should, so there are unlikely to be further scaling improvements in the next public release.

Regards,

Jo

[Falcon]

For those of us who've held back on 10.6 partially because of TG2: Is there an update on this, or a (very rough would be ok) ETA on the next release?

[jo]

Hi,

A very rough ETA would be 3 weeks to a month unfortunately. I'm not sure it would be practical to put out a new release just with the text display fix, it would need a week or two for testing first anyway.

Regards,

Jo

[pclavett]

Hi Jo

Thanks for the information as I was about to open a thread to see if others had experienced this ! The text fields are mostly empty and the numbers are effectively all shifted to the left ! I thought that all my files were corrupted but they are fine when opened in the version on my 10.5 system disk. Most field do correct on clicking on them and the trick is to not change things by doing so.... such as very good typist like myself accidently adding a number or a letter which of course would then corrupt the file. For now will await your correction and deal mostly in TG2 on the 10.5 system... except possibly for final render.... if it is faster on 10.6...and my 8 core MacPro. I would then need to tell the software to use the 8 cores right ??? I have not been attributing all 8 cores to TG2 for now...after reading, I think it was Oshyan's, article on rendering stating that adding cores would sometime slow the rendering. What would be your advice for an 8 core MacPro... to maximize performance during final render ??? Appreciate your comments !

Paul

[old_blaggard]

I've noticed that I can usually get very good scaling on my Mac Pro (Nehalem - 8 cores, 16 threads) by always increasing the "Subdiv cache" setting to at least 150MB/thread. Clouds and skies can render using all 16 threads at maximum efficiency, but terrains (and objects) typically start to slow down at 6 or 8 threads. It's still not ideal, but it's definitely *much* better than it was under 10.5.

[jo]

Hi,

Regarding how many threads to use on an 8 core Mac pro - it depends a bit on the scene. It's hard to give a concrete number I'm afraid.

Generally speaking the GI pass will scale pretty well with more threads. That might suggest that if you're using higher GI settings it might be worthwhile to increase the number of threads.

I think overall my suggestion would be between 4 or 6 threads. 6 threads might be faster but possibly only by a little bit, not the sort of amount you might expect from adding 2 extra render threads. Any speedup you may get from using more is likely to be minimal compared to the risk it may end up taking longer.

These are some times I got on 10.6 for a simple scene of a heightfield and 3d cloud layer, looking up more at the sky:

16 threads : 18:30
8 threads : 16:12
8 threads, 1200 MB subdiv cache ( 150 MB per thread ) : 16:41
6 threads : 13:57
4 threads : 13:58

A more complex scene with more work to do might render a bit faster with more threads because it will take longer to hit the parts of the renderer which cause bottlenecks with more threads, but it's hard to say.

I have to say I didn't see that there was better processor utilisation while rendering sky, just during the GI pass.

Regards,

Jo

[echrei]

I'm planning to get a Gulftown Mac Pro which will have 24 logical cores. Were those numbers with the currently released build or one in development? I know the Windows version handles more threads better, but does it work well with 24 threads yet?

[jo]

Hi Echrei,

The numbers are from the development build but we haven't made any major changes to the threading for the next release so they should be similar.

I don't think anyone has a 24 core machine to test on, so we can't say.

Regards,

Jo

[Falcon]

May I ask again what the ETA is, because the month or so mentioned above has long since passed.

I don't want to rush you, just know whether it's worth waiting for the update before I upgrade to Snow Leopard.

[Oshyan]

We're in the latter stages of testing now, so it's safe to say it should be relatively soon. However this shouldn't hold up your upgrade to Snow Leopard. There are performance improvements that come with the upgrade itself, even without the changes to TG2.

- Oshyan

[jo]

Hi,

Terragen 2.1 with a fix for the edit field display problem is now available:

http://forums.planetside.co.uk/index.php?topic=8268.0

As we've developed the renderer there have been some good improvements to the threaded performance and scaling on OS X 10.5 and 10.6, but 10.6 is still definitely the one to have for best performance.

Regards,

Jo

[echrei]

I ran some tests using the benchmark from 3d Speed Machine to see how Terragen 2.1 with the object raytracing compared to 2.0. Using Terragen 2.1, it took longer to render in Windows 7 than in Mac OS which was a big surprise. Terragen 2.1 took longer without raytracing to render for both operating systems. These tests were done on a 2.6GHz MacBook Pro.

(updated benchmarks in following post)


Object Raytracing:
                     Mac OS 10.6       Windows 7
On                     23' 52"             25' 53"
Off                     27' 57"             31' 40"

[Oshyan]

Very surprising results. Did you do these tests one-after-another, or were you comparing to older results from previous testing?

- Oshyan

[echrei]

The 2.0 results were from a previous testing and I had my times posted on 3D Speed Machine. The Windows version was always consistently faster before, I'm not sure why it isn't now. I ran the render a couple times on Windows just to make sure.

[Oshyan]

Honestly the results are so odd I wouldn't trust them without re-testing on the current system with both versions. Has anything changed in your software environment - new antivirus, etc?

- Oshyan