.. I just figured out that I can have TG2 use 8 threads, although I only have 4 physical cores. This can be achieved through the Intel Core i7 supporting Hyperthreading. I can tell you HT is insanely effective with TG2, at least on this cpu. It really looks like twice the speed compared to 4 threads. INSANE!
I'm rendering a modified "cumulonimbus distant shot" scene, with 3 cloud layer and 0.9 quality, Gi 2/2, 700 cloud samples, 32 atmo samples at 1600x800 resolution. after 30 minutes, the GI prepass was done including 20% of the scene already rendered!
Guys, if you can make that possible, drop your whatever most desperate wish for 2009 and put a core i7 on your list instead ;-)
Cheers,
Frank
When we were discussing your machine on MSN I was a bit surprised that you had to use 2 instances to get HT to work. My idea was: if 1 instance can't use it, the second won't too.
But never paid too much attention to it :)
Great to see it works!
And of course: show us the result ;)
Martin
yeah, I really HAVE to set min threads to 8. If I have min threads = 4, it won't use HT.
Cheers,
Frank
damn ! T-U told me your computer was brutal but i didn't think that much brutal !!! :)
yeah, it's ridiculously brutal at TG2 ;-)
.. the image I just posted finished shortly after 2 hrs. The thick cloud part took a little longer ;-)
damn only 2 hours for those big fat good lookin' clouds ?! '-_-
incredible ! i need one of those processor ^^
... and it's the slowest of the available i7's. mine runs on 2.6 GHz, the fastest one at ~3 GHz (untuned).
Wow, that's fantastic news! I'm just getting more and more excited about the possibilities here :).
Slobber.
;D
Oh this sounds very exciting. I'm going to have to save up for one of these chips.
BTW Frank did you read this thread? Didn't sounds like it would make that much difference but it sound like it does.
http://forums.planetside.co.uk/index.php?topic=5032.0
I am not doing benchmarks, but I can tell you it feels like almost twice the speed with HT enabled. At the very least, the difference is clearly notable.
If and when my TG2 creativiy bursts needs a rest, I'll do a benchmark.
Cheers,
Frank
Ech I wish that GPU-CPU would scale 1:1 in terms of TG2 rendering. It would be beautiful to do some renders on my 96-shader GeForce ::)
Also I seem to remember reading on here that TG multicore support isn't particularly optimised yet, especially with more than 4 cores. So post TG2 we may get even more speed as the multi-threaded capability improves. It would be nice to the get the preview window multi-threaded as well.
multi threaded preview... that would be another productivity boost.
Quote from: FrankB on January 06, 2009, 10:47:49 AM
multi threaded preview... that would be another productivity boost.
I second that.
Hi Frank,
TG2's processor detection code currently ignores hyperthread virtual processors/cores and only accepts physical cores. This is because on older HT processors it was slower. Perhaps we'll need to change it for Nehalem family processors. In any case, you can set your preferred number of cores in the Startup preferences panel if you want to stick with 8.
Your news is interesting, as it makes me think I should definitely wait for a Nehalem based Mac Pro before buying a new machine. A dual quad core machine with HT ( which was actually fast ) would allow 16 threads. TG Mac in particular doesn't scale very well past 4 yet though.
Regards,
Jo
Hi Jo and all,
OK, I made a little benchmark for with and without HT
First of all, I took Karsten's benchmark scene, and was highly disappointed, because with HT, the benchmark was only like 3 seconds faster :-(
I though I would have to go back to this forum and apologize for the hysteric excitement and over-euphorical misinformation, but I thought "Frank, you know it's faster, you've seen it!"
I gave it another try with a slightly more difficult scene. With GI 2/2 and quality 1, and a heightfield covered with countless fakestones of all sizes. (still a simple scene, after all).
Results:
4 cores = 4:22 min
8 cores = 3:35 min
so... it's not twice as fast like I said I believe it was, but still alot faster - in this scene! I wouldn't be surprised if on other scenes, it would run close to twice as fast, maybe if I render sky, water, objects... I don't know.
Cheers,
Frank
Has anyone no mercy?
::) <-- fingers in ears and humming...
Quote from: jo on January 06, 2009, 11:12:48 PM
A dual quad core machine with HT ( which was actually fast ) would allow 16 threads.
Regards,
Jo
Hi Frank ... can you upload that file?
I just ran Karsten's test on an 'in-use'* 2x Intel Xeon Quad.
Using
2 Cores: 2:26
4 Cores: 1:13
7 Cores: 1:03
8 Cores: 0:56
15 Threads: 1:04
*'in-use': Well, the machine will be free for me from 7pm to 7am weekdays and the whole weekend ,-) - during the testrender it served about 20 Clients ,-)
good that you asked me for it! While digging this up, I figured that had set min thread to only 7, instead of 8 (maybe a typo or so). I've rerendered the scene with 8 threads and the new time is now 2:52 min.
I've attached the file for you. It's a stupid scene that decides to render background mountains first in the upper right quadrant, before covering them with the foregound later.... But for comparison, I'll leave it as is.
You'll have to generate the terrain first and use the quick render node for rendering.
Cheers;
Frank
By the way ... a well chosen title for this topic ,-)
The full 8 hardware-cores of my machine had to render ... 3'46"
Setting to 15 threads (did not dare to go for more ;D ) made 4'50" (while the render finished visually 1 minute earlier, and just waited for one cpu to complete its work)
Congrats to that hot machine! 8) 8) 8)
Volker
29:37 on my single core Xeon 2.8Ghz. (irwindale core (dell precision 470))
[edit] checked it again 29:35
:( :( :( :( :(
I need a dual i7 quadcore. (assuming one exists) really really badly :o
[edit]
no dual i7 is possible. but xeon versions of the chip (for servers) will be capable. Xeon 5570
Check out this amazing info:
http://www.anandtech.com/weblog/showpost.aspx?i=532 (http://www.anandtech.com/weblog/showpost.aspx?i=532)
Quote from: FrankB on January 07, 2009, 10:25:54 AM
good that you asked me for it! While digging this up, I figured that had set min thread to only 7, instead of 8 (maybe a typo or so). I've rerendered the scene with 8 threads and the new time is now 2:52 min.
Cheers for the file Frank. You say 2:52 on 8 cores? I'm guessing that's the 4 HT'd cores. I just ran it on my core 2 duo. 2:51 ;D
interesting.
and hard to believe, considering a single core xeon @2.8GHz took half an hour. Did you use the quick render node?
Frank
Nope. Left everything as it is except for the threads settings. Just set those to 2.
Edit: Oh I see you were using quick render. Yeah it is slower on that one. I couldn't even be bothered to finish it :D It was only half way through the render when it passed your 2:52. It eventually took 6 minutes exactly. The weird thing with this scene is that it renders the first two chunks as normal, then when the second of the two finishes the first chunk suddenly goes painfully slow and the render only works one chunk at a time until the last two chunks where it does them both at the same time again.
Hi,
Just for laughs I thought I'd try it on my dual 2.7 GHz G5. 12:23 - that new machine is looking better and better!
Regards,
Jo
Also for laughs :)
4:28
richard
I punched up 100mhz more to my clockspeeds so ~3,5ghz(q9400) now and results were: 3:01
Btw if anyone happens to have one those new Phenom II processors in their hands, I'd certainly like to see how the scale here
Quote from: FrankB on January 07, 2009, 10:25:54 AM
good that you asked me for it! While digging this up, I figured that had set min thread to only 7, instead of 8 (maybe a typo or so). I've rerendered the scene with 8 threads and the new time is now 2:52 min.
I've attached the file for you. It's a stupid scene that decides to render background mountains first in the upper right quadrant, before covering them with the foregound later.... But for comparison, I'll leave it as is.
You'll have to generate the terrain first and use the quick render node for rendering.
Cheers;
Frank
Sniff, 6:56s on my last year's Intel E8200 with 2gb ram, @ 2 threads... :(