PDA

View Full Version : Hyperthreading... that's no joke!



jamesl
03-06-2004, 12:21 AM
Now, I never was one to buy into Intel's marketing hoopla, but... I have to say, I just found out that hyperthreading is for real. I recently put together a dual xeon 2.8 Ghz machine, and had my first heavily raytraced render to throw at it. My previous workstation was a dual PIII 500, and I seldom saw a significant performance increase going over 2 threads. In fact, 4 threads usually introduced so much processor overhead that it took a second or 2 longer. Well imagine my surprise when I got the following numbers under XP and my Xeons:

1 thread: 3:13
2 threads: 2:06
4 threads: 1:23
8 threads: 1:20

This is enhanced medium AA, ray recursion 7. Even though it seemed to plateu at 8 threads, there was still a slight performance increase! I'd like to see some benchmarks run with dual G5's, to see if there is an analagous performance curve, because I'm pretty impressed with this hyperthreading stuff.

Ah, who cares, we're all going to be rendering interactive GI next week with Fprime anyways... :rolleyes:

j

Psyhke
03-06-2004, 01:55 AM
Which brings up the question: Will FPrime utilize dual processors/multi-threading??

Capt Lightwave
03-06-2004, 04:54 AM
Have you asked them?

bloontz
03-06-2004, 07:55 AM
Do you have Intels hyperthreading turned on? I think that Newteks Multithreading is something different, in tests that I did Intels hyperthreading seemed to give mixed results, sometimes actually slowing a render slightly. I turned it off but really need to explore it further.

mamurphy
03-06-2004, 11:55 AM
I find those benchmarks very interesting. I decided to do my own comparisons between a dual 1Ghz p3 Dell workstation WinXP and a single 1Ghz Apple Powerbook OSX. Both with 1 Gig Memory.

I rendered the Radiosity_ReflectiveThings on default settings.
PC:
1 thread : 3:08s
2 threads: 2:15s
3 threads: 1:51s
4 threads: 1:46s

Mac:
1 thread : 1:44s
2 threads: 1:51s
3 threads: 1:57s
4 threads: 1:59s

Neither has blazing speed, but I sure thought the PC would destroy the Mac. For goodness sakes. It's a workstation versus a laptop. I'm not trying to start a platform war, but I sure hope I can convince my job to upgrade this workstation. I would prefer a dual G5, but a faster PC would be fine;)

Aegis
03-06-2004, 01:10 PM
The dual G5's are still a way behind dual 3.x Ghz PC workstations - a single G5 2Ghz CPU clocks in at about the same speed as a 2.53Ghz P4 - the promised 2.6Ghz G5's should level the playing field.

Incidentally, Hyper-Threading is great on a single PC but it causes problems when network rendering - in order to get full 100% usage out of the CPUs it's necessary to run 4 ScreamerNet Nodes on the machine, each of which need enough RAM to render the scene - in this kind of scenario a PC can chew up 2Gb of RAM scarily fast. That said, all the tests I've done indicate that even as part of a ScreamerNet, a PC will render faster with Hyper-Threading on - very cool.

Oh, and my dual Xeon does Radiosity Reflective things in around 26.7 seconds :)

mamurphy
03-06-2004, 01:26 PM
Well, you need to talk to my boss and tell him to buy me a new machine!!!:D

Time is money






...I still love my powerbook...Go little guy, Go.

Aegis
03-06-2004, 01:29 PM
Send me the moolah and I'll build you one :D my dual 3.06Ghz box cost just a tad under 2000 (including VAT) - and that includes 3 x 10,000rpm Western Digital "Raptor" drives for my VT[3] - a pure LightWave workstation could come in at a lot less than that. Dell sells a lesser specced workstation for around 3700 :eek:

mamurphy
03-06-2004, 01:58 PM
Sorry, I'm saving my $3000 for those shiny new dual 2.6 Ghz G5s:cool:

And I know my workplace won't spend any money anytime soon. I just got them to buy Lightwave!!!!:D

tjacobs
03-06-2004, 02:05 PM
Hyperthreading with 8 threads has always provided the best speed on my dual 2.4 zeons.

Phil
03-06-2004, 02:11 PM
Be aware that there are known issues with LW's multithreaded rendering. It's surfaced here a number of times and NT are aware of the problems.

Most of the time you won't have a problem, but it's worth checking jobs with significant usage of fog/volumetrics/lens flares, etc. and ray tracing. We've had jobs that can only be rendered single threaded - multi-threaded renders cause major problems within the same frame (different intensities between different threads) and additionally between those same segments across different frames. There has been no comment (beyond confirming the bugs) as to whether this will be fixed in 8.0.

This affects LWSN as well as Layout, so is a bug in their render code.

jin choung
03-06-2004, 07:38 PM
"hyperthreading" only exists on p4s and better....

so the p3 wouldn't count as an example of a hyperthread.... also, i have no idea why the hell you would be able to get EIGHT threads on 2 hyperthreaded processors (that should be 4...).

jin

toby
03-06-2004, 07:49 PM
You can set a single to 8 threads if you want - might even go faster -

I used to leave it on 8 with my dual G4 because it was fastest 90% of the time, but somewhere along the line, an upgrade of LW or the system, 2 is now the fastest 90% of the time. It also depends on the scene you're rendering, and probably how the rest of the computer is configured too, I wouldn't give hyperthreading all the credit.

bloontz
03-06-2004, 08:07 PM
I think some here are missunderstanding what hyperthreading is. It's a scheme that sets up a processor to act as two. It is a setting turned on in the bios of compatible computers and is supposed to allow for better responsiveness when using mulitple applications simultaneously. If I have hyperthreading active on my dual xeon machine the system information panel shows 4 processors.

Lamont
03-06-2004, 10:33 PM
James, when you look in the Computer Management tab, how many processors are shown?

I thought only Win2K Server saw 4 processors?

bloontz
03-06-2004, 11:15 PM
I'm running XP also.

GruvSyco
03-07-2004, 12:45 AM
Originally posted by Lamont
James, when you look in the Computer Management tab, how many processors are shown?

I thought only Win2K Server saw 4 processors?

Win2K server can support up to 4 out of the box. I believe you can get support for more directly from MS or vendors that sell systems wth more procs. Win2k Pro only supports 2. XP pro will support 2 while XP home only supports 1. Just an FYI :)

Psyhke
03-07-2004, 02:09 AM
Originally posted by Capt Lightwave
Have you asked them?

No. But I believe I shall.

bloontz
03-07-2004, 06:54 AM
Originally posted by GruvSyco
Win2K server can support up to 4 out of the box. I believe you can get support for more directly from MS or vendors that sell systems wth more procs. Win2k Pro only supports 2. XP pro will support 2 while XP home only supports 1. Just an FYI :)

W2K Advanced server supports 8 and W2K Datacenter server supports 32.

That doesn't have anything to do with the virtual processors that are diplayed when using hyperthreading though, as XP Pro will show 4 procs on a dual machine.

wavk
03-08-2004, 06:01 AM
That's weird! When I got a new machine at work, some time ago, I did some tests also like the ones you did. I have a 3 Ghz P4 with hyperthreading. My test results were completely different, my machine is faster without hyperthreading. I set it on and off in the bios and did the test with number of threads in lightwave also. And for single-threaded apps, the hyperthreading sucks bigtime, it gives you a 50% speed decrease.

Aegis
03-08-2004, 06:26 AM
Maybe H.T. just works better in a dual CPU environment? I don't have access to a single CPU machine to test the theory...

Bytehawk
03-08-2004, 05:44 PM
I don't have access to a single CPU machine to test the theory...

that's not fair, you 're spoilt rotten !!!!

don't tease us like that

:D

Aegis
03-08-2004, 11:25 PM
that's not fair, you 're spoilt rotten !!!!

Heh! What can I say? The studio here is full of dual Xeons - I've got 3 :D :D :D

samartin
03-09-2004, 06:32 AM
Originally posted by wavk
That's weird! When I got a new machine at work, some time ago, I did some tests also like the ones you did. I have a 3 Ghz P4 with hyperthreading. My test results were completely different, my machine is faster without hyperthreading. I set it on and off in the bios and did the test with number of threads in lightwave also. And for single-threaded apps, the hyperthreading sucks bigtime, it gives you a 50% speed decrease.

I find what you found above weird, here's my results (P4 3.06 laptop) :-

HT -

1 thread - 54.3 & 1m 5s while browsing CGTalk (very responsive feedback when using other programs).
2 thread - 51.9
4 thread - 49.6
8 thread - 48.6

No HT -

1 thread - 1m 29s
2 thread - 1m 26s
4 thread - 1m 28s
8 thread - 1m 8s

If I don't use HT I can forget about using any other program as it slows down dramatically...

*edit*

That was rendering reflective radiosity things...

dwburman
03-09-2004, 08:08 AM
I've noticed that when rendering something very simple (renders in a few seconds or less) HyperThreading or at least rendering with multiple threads does slow things down over all. But that's a fairly specific situation.

lede
03-09-2004, 09:33 AM
I was doing a small test on my 3.0 HT machine the other day and a simple scene would render in 10s but when I bumped up the threads to 2 it took 15s. Now when I throw a more complex scene at it that takes 45s on one thread it would take 39s on two threads.

I'll have to play around with the setting some more, but it looks like the dual processor is much faster than a single.

-Lede

wavk
03-09-2004, 10:28 AM
Samartin, I find it hard to believe that with 1 thread, hyperthreading is faster than without. On my machine, when I use hyperthreading, it fools win2k of having a two processor machine, you can see this in the task manager. When one thread is used, only one of the processors can get to work, so only half of the real processor is used. Are you sure you have no processor heavy stuff running in the background, like a virus scanner? That would make sense, so win2k would use the other (virtual) processor, which isn't used, for rendering. Hmm... weird test results. I guess not only the processor counts, but maybe also which win version or which mobo or the number of memory banks, or whatever. My tests were really thorough and with large scenes as well. Hm.... wait now I come to think of it, was I using LightWave already back then? I don't think so, so these tests must have been with YafRay! Now I'm not so sure anymore if I also tested with LightWave. Should not matter because it's still multi-threaded, but could make a difference.

Okay, I'm giving it another try tomorrow with LightWave! See you then!

Have fun,

Wybren van Keulen

samartin
03-09-2004, 10:58 AM
If it makes any difference, I am using XP Pro and maybe that is more suited to HT technology. But I have to say what I have noticed tho' cos' it is a laptop, when not using HT my fan makes more noise and it looks like the render passes slow down when they get to pass 3/5. I use HT all the time now as I am not hugely concerned with rendering speed, I like to browse when I am rendering.

I hope you find out why it doesn't work so well for you...

*edit*

I do have norton virus running in the background but I wouldn't have thought it would make a huge impact on performance, I'm not as techy these days as I once was...

toby
03-09-2004, 05:37 PM
Some processes will take longer with multiple threads because splitting the data into multiple threads actually takes longer than the benefit of rendering with 2 processors, i.e. 4 second renders.

It's just like driving your car across the street - it's faster to walk.

kyfsu00
03-10-2004, 11:48 AM
no amd people here?

i use a p4 3.06 ghz with 1gb memory at work and at home an amd athlon 2800+ with 1gb memory and home renders a whole lot faster. almost twice as fast with more background programs running.

mattclary
03-10-2004, 12:31 PM
Originally posted by kyfsu00
no amd people here?

i use a p4 3.06 ghz with 1gb memory at work and at home an amd athlon 2500+ with 1gb memory and home renders a whole lot faster. almost twice as fast with more background programs running.

Well, you are seeing the opposite of what most people see, myself included. I like AMD, but for pure render power, the P4 is currently top dog (for single cpu systems).

http://www.blanos.com

samartin
03-10-2004, 12:48 PM
I used to have a 2800+ OC'd to 3200+ and it's roughly the same speed as my 3Ghz machine althought when GI comes into play it is way way faster due to the SSE2 instruction set...

amorano
03-12-2004, 10:07 AM
should have a dual opteron (244's) setup with 2gig of ram going on moday when the drives get here. I'll try to remember to post some results here, and most definately will on blanos.

jeremyhardin
03-12-2004, 12:46 PM
Radiosity_ReflectiveThings.lws --default settings

Mac Dual 2 Ghz G5 1GB RAM

1 thread: 55.5 seconds
2 threads: 44.6 seconds
4 threads: 36.2 seconds
8 threads: 34.1 seconds

wavk
03-13-2004, 04:41 AM
Finally got some time to test it. I don't know the exact number of seconds the renders took, but hyperthreading indeed works with LightWave! Cool.

With two threads it was a couple of seconds slower, with 4 threads way faster and eight threads sped it up just a little bit, also. But I guess there's no threading in FPrime;)

mamurphy
03-13-2004, 09:47 AM
I sure like the look of those G5 numbers. Hopefully Apple is releasing their dual 3Ghz at the end of the summer.