|
|
Hyperthreading
About Truespace Archives
These pages are a copy of the official truespace forums prior to their removal somewhere around 2011.
They are retained here for archive purposes only.
Hyperthreading // Hardware
Post by PWL // Mar 1, 2006, 4:20pm
|
PWL
Total Posts: 17
|
Find it hard to believe that this hasn't been brought up yet (maybe I overlooked it) but with the new duel core cpu's out, they do not support hyperthreading. I wonder what the trade off is in performance; Duel Core w/no HT VS. Single Core w/HT. |
Post by stan // Mar 1, 2006, 5:27pm
|
stan
Total Posts: 1240
|
it's called multithreading in truespace..look in the rendering option icons or if you have vray, right click the render icon and it shows up in stack view ..tools ..in the vray preferences:banana: .. |
Post by ProfessorKhaos // Mar 1, 2006, 9:24pm
|
ProfessorKhaos
Total Posts: 622
|
I think hyperthreading is basically a poor man's dual core processor. Think of it like a "siamese twin" when compared to the dual core's "identical twin" status.
Hyperthreading technology allows the processor to perform two operations in parallel if those operations happen in different parts of the chip. If both operations require the same circuitry, then one of the operations must wait until the other operation is done. This is better than the old way where only one operation was performed at a time regardless of how free the other parts of the chip were.
Hyperthreading is cheaper than a full blown dual core processor because the logic required to share circuitry is less than a full second set of circuitry. Component count is lower and the chip die size can be smaller as a result.
The cost of producing microchips is very strongly related to the final die size because the bigger the chip must be, the more likely an error will be found on the chip's surface.
For example, let's say a 10 x 10 cm silicon wafer is likely to have 1 flaw on it's surface such that circuitry drawn on that area won't work. If that wafer is used to create 4 chips (each 5 x 5 cm) then one out of 4 will be bad. If that same wafer is used to create 16 chips (each 2.5 x 2.5 cm) then 1 out of 16 will be bad. If the wafer is used to create 64 chips (each 1.25 x 1.25 cm) then 1 out of 64 will be bad. Remember, there's only 1 flaw on the surface.
Statistically speaking, a relatively small change in die size can make a pretty big difference in cost, especially when talking about microprocessor chips that are large to begin with.
From this perspective, the hyperthreading CPU is more cost effective than a dual core processor for the amount of performance you're buying, though it's actual performance will be somewhat less given an equivalent clock speed. As is most often true, you pay exponentially for the corresponding increase in performance. You also have to take into account that microprocessors require the supporting motherboard, ram, etc. to run and that has a certain fixed cost associated with it that might not increase all that much with a dual core processor. |
Post by SiRender // Mar 1, 2006, 11:35pm
|
SiRender
Total Posts: 38
|
I'd second what ProfKhaos wrote (cool analogy). And add that hyperthreading (HT) lets one thread do useful work when the other is stalled. For example, waiting on a memory access to be completed, recovering from branch mispredict, etc. This is useful even if you have more cores. So HT can boost the performance even in a dual core system.
More Intel processors will have both Multi-Core and HT, but there is a strange split right now (e.g. dual core without HT) but that is due to the development paths, time to market, and pricing choices for a processor not technical benefits of one vs. the other.
Hope that helps. I have an Intel HT processor now and it is nice for renders. I want to get a Dual Core with HT eventually.:) |
Post by hemulin // Mar 2, 2006, 11:40am
|
hemulin
Total Posts: 1058
|
...the hyperthreading CPU is more cost effective than a dual core processor for the amount of performance you're buying...
Did you know that the "virtual" cpu on a cpu with hyperthreading consumes nearly as much energy as the real cpu. I think this may alter your cost effectiveness statement.:) |
Post by ProfessorKhaos // Mar 2, 2006, 4:40pm
|
ProfessorKhaos
Total Posts: 622
|
True, but 90% of the time most processors are underutilized anyway (somewhere less than 50%) which means all that extra power is going for naught and the difference becomes a wash except when the machine is heavily loaded such as during an actual rendering session. In my case I'm spending far more time preparing models in the GUI than actually rendering stuff though I'd certainly agree with you in the case of a render farm (but also for other reasons). :)
Cheers!
P.K. |
Post by hemulin // Mar 3, 2006, 9:01am
|
hemulin
Total Posts: 1058
|
..True, but 90% of the time most processors are underutilized anyway (somewhere less than 50%) which means all that extra power is going for naught and the difference becomes a wash...
True, I failed to remember that the average use doesn't run folding@home, which blasts the cpu's up to 100% constantly.:rolleyes:
Oh and please don't somebody post back and say that running your cpu at 100% constantly is bad for its health because firstly its not (though there is a bit more explaining to that), and secondly I really don't care. |
Post by Freaky42 // Mar 6, 2006, 10:23pm
|
Freaky42
Total Posts: 28
|
Find it hard to believe that this hasn't been brought up yet (maybe I overlooked it) but with the new duel core cpu's out, they do not support hyperthreading. I wonder what the trade off is in performance; Duel Core w/no HT VS. Single Core w/HT.
Actually I've been looking around for Dual core CPUs for a little while and to my knowledge the AMD Athlon X2's and the newer FX-60 Dual cores are HT capable. It lists that in the Tech info in the areas I've read on them at least. It showed me that each core had a HT control section for it along with each individual memory cache each. Could be wrong I suppose but I do believe that AMD may be the only one at the moment. However Intel just realeased ,... um the 940 I think it was and I think it was HT capable too. Hope that helps. I'm up for correction if you find out anything else:p |
Post by Freaky42 // Mar 6, 2006, 11:48pm
|
Freaky42
Total Posts: 28
|
There we go that's what I was looking at for the AMD Athlon FX-60 one that's out now. Was wrong that it had a HT contoller for each core, aparentely it doesn't work that way since it's for the controll of bus management. whoops.:p |
Post by hemulin // Mar 7, 2006, 6:56am
|
hemulin
Total Posts: 1058
|
Erm, Intel Processors are the only processors that support/use Hyperthreading which is basically making a virtual CPU and as far as I know, that is what this thread is about. AMD processors are the only processors that support/use Hypertransport which is something completley different and bares no relevance whatsoever to hyperthreading, even though both (stupidly) have the same acronym.
Can I suggest that in this thread we use the acronym HT for hyperthreading and just use the whole word for hypertransport |
Post by hemulin // Mar 7, 2006, 6:59am
|
hemulin
Total Posts: 1058
|
Also I think (this bit I may be wrong on) that the most recent dual core AMD processors are the only ones that use Hypertransport because they do not have a northbridge on the motherboards. |
Post by Alien // Mar 7, 2006, 8:18am
|
Alien
Total Posts: 1231
|
I know all socket 940 AMD CPUs have hypertransport [Opteron 1xx, 2xx, & 8xx series], but I'm not 100% sure about socket 939 chips. I'm pretty sure they all do, just not 100% certain. My line of reasoning is this: Single core AthlonFX series chips are basically an Opteron 1xx series chip with 1 less pin & a different name [although I think there are now some 939 Opteron 1xxs as well, so... <shrug>] Anyway, if the FXs have it & they can be used on the same boards as non-FX chips, then it seems logical that the non-FX chips must have hypertransport as well, seeing as [as Hemulin correctly stated] it replaces the memory handling that would normally exist on a NB chip. |
Post by hemulin // Mar 7, 2006, 10:27am
|
hemulin
Total Posts: 1058
|
Alien, all the stuff you'v said is correct, you certainly know what your talking about. However where did Hypertransport come into it? I thought this thread was about hyperthreading. I'm not against threads changing half-way through, however i'm not sure why it has changed. Are some people getting confused in this thread and interchanging the two names. :confused: |
Post by Alien // Mar 7, 2006, 11:09am
|
Alien
Total Posts: 1231
|
Alien, all the stuff you'v said is correct, you certainly know what your talking about. However where did Hypertransport come into it? I thought this thread was about hyperthreading. I'm not against threads changing half-way through, however i'm not sure why it has changed. Are some people getting confused in this thread and interchanging the two names. :confused:
It would seem that way - I think I must have misunderstood 1 of the earlier posts - I thought someone was implying that only the more recent AMD CPUs [specifically the dual-core 1s] had hypertransport. |
Post by bigL // Mar 10, 2006, 12:39pm
|
bigL
Total Posts: 16
|
okay, here are some results of my test for analysing the performance of hyperthreading:
(always the same scene, rendered with truespace 6.5)
1 CPU without Hyperthreading, rendertime 1:47
1 CPU with Hyperthreading, rendertime 1:20
2 CPU without Hyperthreading, rendertime 0:57
2 CPU with Hyperthreading, rendertime 0:40
btw.: with Truespace 7 it takes 0:38 -))
this are very amazing results/timings....
regards,
bigL |
Post by jamesmc // Mar 18, 2006, 7:21am
|
jamesmc
Total Posts: 2566
|
My old Dell, P4 3.0 with HT, renders faster than my newer Athlon 64 bit processor! :( |
Post by hemulin // Mar 18, 2006, 7:32am
|
hemulin
Total Posts: 1058
|
what processor model is your Athlon 64? |
Post by jamesmc // Mar 18, 2006, 1:38pm
|
jamesmc
Total Posts: 2566
|
Here's the system information on the Athlon:
OS Name Microsoft Windows XP Professional
Version 5.1.2600 Service Pack 2 Build 2600
OS Manufacturer Microsoft Corporation
System Manufacturer MICRO-STAR INTERNATIONAL CO., LTD
System Model MS-7125
System Type X86-based PC
Processor x86 Family 15 Model 47 Stepping 2 AuthenticAMD ~1809 Mhz
BIOS Version/Date Phoenix Technologies, LTD 6.00 PG, 11/28/2005
SMBIOS Version 2.3
In plain English it's an AMD 3000+ for the 939 motherboard socket which can be overclocked, but running at 1.809ghz and it has Hypertransport. I have overclocked it up to 2.7ghz but didn't want to keep it there.
Funny thing is that on the speed tests I conducted with performance software, the AMD beat my old Pentium 4, 3.0ghz. However, in reality the P4 Pentium appears to do a better job at rendering.
In my case, the Pentium 4 Hyperthreading is better than the AMD 64 Athlon Hypertransport. :confused:
The difference may be in the Video Cards. On the AMD I have a PCI-E 6200 Turbo by NVIDIA. It states it comes with 128 ram, but borrows system ram to boost it up to 256.
The Pentium has a Nvidia 6000 series (blah forgot the number) and in an older type of video card slot (AGP.) However, it came with 256mb of ram and doesn't borrow from system memory.
I'm not sure how rendering works, but this may be the difference in my case. |
Post by Alien // Mar 18, 2006, 3:47pm
|
Alien
Total Posts: 1231
|
Here's the system information on the Athlon:
~1809 Mhz
Wow, didn't know they made A64s that slow!
In plain English it's an AMD 3000+ for the 939 motherboard socket which can be overclocked, but running at 1.809ghz and it has Hypertransport. I have overclocked it up to 2.7ghz but didn't want to keep it there.
Sounds like at stock speed even my socket 462 AXP [3200+] would beat it.
Funny thing is that on the speed tests I conducted with performance software, the AMD beat my old Pentium 4, 3.0ghz. However, in reality the P4 Pentium appears to do a better job at rendering.
Probably the speed tests relied a bit more on memory bandwidth/performance, which [thanks to hypertransport] 64bit AMD chips usually have the advantage in.
In my case, the Pentium 4 Hyperthreading is better than the AMD 64 Athlon Hypertransport. :confused:
Comparing Hyperthreading and Hypertransport is somewhat like comparing whales to elephants [or apples to oranges :)].
The difference may be in the Video Cards. On the AMD I have a PCI-E 6200 Turbo by NVIDIA. It states it comes with 128 ram, but borrows system ram to boost it up to 256.
The Pentium has a Nvidia 6000 series (blah forgot the number) and in an older type of video card slot (AGP.) However, it came with 256mb of ram and doesn't borrow from system memory.
It's not the graphics card - graphics cards have zero impact on rendering speed.
I'm not sure how rendering works, but this may be the difference in my case.
Nope, it's just that [AFAIK] you got the slowest 64bit Athlon AMD make. |
Post by hemulin // Mar 19, 2006, 9:35am
|
hemulin
Total Posts: 1058
|
HyperThreading Technology and Hypertransport are NOT equivalent technologies. Nor do they perform the same function. Hyperthreading creates a virtual 2nd instruction unit on Intel Processors that allows a second thread to run in parallel if the first is flushing its pipeline or not using a particular part of the CPU. For example if you had two separate applications, one that relied entirely on floating point calculations (those requiring REAL numbers, i.e. decimals), and another that relied entirely in integer calculations (whole numbers), ignore any cache/memory related issues, both threads would be able to run with little interference from the other, since one thread would use the FPU (floating point unit) and the other would use the CPU proper, and as such would see no significant performance hit (i.e. you would see almost 2x increase in speed over a non Hyperthreaded CPU). Hyperthreading was designed to overcome Intel CPUs long pipeline limitation, in that with branchy code, the branch prediction becomes very inefficient with a long pipeline. P4s can lose up to 30% of their CPU power because of branchy code which requires the pipeline to be flushed after ~10 ops (the pipeline is ~30 ops). While the pipeline is being flushed on the first thread, the second thread can work as normal, since the first thread is using no (or at least very little) CPU power.
Hypertransport is an AMD64 specific technology that does away with the front-side bus (FSB) and lets the CPU connect and communicate directly with the memory (via a dedicated memory interface) and the chipset (via the HT bus) without having to negotiate through a northbridge. AMD64 processors have a built in memory controller (whereas with other technologies this is on the northbridge),but this has nothing to do with the HT bus, although one cannot exist without the other. The memory controller is inherently faster since the path to memory is shorter due to the lack of a northbridge. The HT Bus's sole responsibility is to communicate data between the CPU and other peripherals at as close to 1GHz as possible.
In a multithread capable application such as TS, any sort of dual core processor is advantageous, including hyperthreaded processors for precisely the reason mentioned above. When one thread is flushing its pipeline, the other takes over. For the lower clocked AMD CPUs the hyperthreaded P4 will beat them, simply because it can do more work thanks to the virtual CPU, but as the AMD clocks get higher the P4s advantage is diminished and the faster clocked (compared to the low spec versions) but shorter pipelined AMD machines start taking over in the rendering time stakes. True dual core CPUs perform much better, since they have 100% each of a physical core to work with, and advantages would be seen with dual hyperthreaded Xeons for example, since you could run threads to give somewhere close to 300% performance over a single non-hyperthreaded CPU. The actual performance increase is HIGHLY dependant on the type of calculations being performed.
This quote is not from the tS forums (I just decided to use the forums quoting system to acknowledge the authors name(screen name)). |
Post by Emmanuel // May 31, 2006, 8:11am
|
Emmanuel
Total Posts: 439
|
Hi all,
I am reading this topic again because I have to purchase a new PC.
My current tS system is a P4 C 3.2 -often overclocked with good stability at 3.52 Mhz-, 2 Gb of CORSAIR DDR2 3200 RAM, MB ASUS P4P800Deluxe, 2 HD SERIAL ATA RAID 0.
Unlike BigL, I noticed that multithreading in tS6.6 was slower sometimes when rendering big scenes :
At 3.2, with no other app running in background, a scene takes 40s to render with multithreading... and 36s without it.
I think it may depends on some scene settings or maybe other system parameters, but, well, I am not convinced by the tS's multithreading feature.
Now we have some real dual core processors and a chance to speed up our calculations :
on my left the Athlon 64 X2 4800+ socket 939 for 620 €,
and on my right an Intel D960 Dual Core! socket 775 for 560 €.
I always purchased Intel processor but this article (http://www.hardwarezone.com/articles/view.php?id=1896&cid=2&pg=1)tends to say that today's best choice for a 3D computer seems to be in the AMD.
Do you have any comments ? |
Post by daybe // May 31, 2006, 8:22am
|
daybe
Total Posts: 562
|
I can't say too much about hyperthreading or dual core as I haven't investigated it very much and really have nothing to compare it to. I can say that I have had used intel for along time but switched to AMD a few years back and find performance is far better for me at least. Generally here AMD is alot cheaper as well for an Intel of equal specs, I currently am running with a Athlon 64 3200+ socket 939 and have been more than happy. They do tend to run a little hotter then intel but a good fan will take of that and you will need a have descent power supply, mine is 450 W and hasn't had any problems.
HTH |
Post by hemulin // May 31, 2006, 8:31am
|
hemulin
Total Posts: 1058
|
Personally I would go for the AMD, yes it is a little bit more expensive but the performance should be much greater than the Pentium D. |
Post by Emmanuel // May 31, 2006, 9:15am
|
Emmanuel
Total Posts: 439
|
Some articles say that heat and consumption of the new Athlon 64 X2 are lower than Intel's.
I am always suspicious toward these articles but the price difference seems to confirm the superiority of today's AMD processors. |
Post by Alien // May 31, 2006, 11:46am
|
Alien
Total Posts: 1231
|
on my left the Athlon 64 X2 4800+ socket 939 for 620 €,
and on my right an Intel D960 Dual Core! socket 775 for 560 €.
I've only tried 2 dual-core pentiums personally - my stepdad's Medion, which has an 800 series chip [either 3 or 3.2Ghz, 1GB RAM], & a Packard Bell in a PC World store, which had a 900 series chip [don't recall speed or exact model]. Whilst the Medion's ok/not-too-bad, the performance of the "Pack-o'-s***" was appalling - a dual core machine with 1GB RAM running more like a P2 with 1/2 or even 1/4 of that amount of RAM! :eek: |
Post by TomG // Jun 1, 2006, 3:02am
|
TomG
Total Posts: 3397
|
One thing to remember about multithreading (and even dual cores) is that it doesn't magically create extra power :)
Think of a car engine, with 600HP say, and a car with two 300HP engines. Ignoring any details of physics etc, the end result is a car with 600HP, in other words that can go just as fast.
There is no inherent benefit in splitting into two.
BUT imagine the car with two engines can actually split into two cars. Now there is a benefit, because when car A has to go to the grocery store, car B can go to the library. Now each car goes there at half speed, but the thing is before one task had to wait before the other could even begin.
This allows for a more efficient use of the 600HP, by driving to two different locations at once.
Of course, for rendering, this might not be a huge timesaver at all. Consider that rendering two lines at once isn't inherently better than rendering one line at twice the speed in terms of the overall render speed.
Of course there are micro details in the tasks (it's not just "one line at once or two lines at once") but all the same, the biggest difference that multithreading brings is the ability to do something else (browse the internet) while rendering without your machine being slowed to a crawl. Of course this is because your internet browsing or word processing gets given one 300HP engine, while your rendering gets given the other. With the one engine scenario, all 600HP have to go into one or the other at any given instant.
Given that your machine may or may not be doing other processes, it is possible that multithreading could slow things down, as your render may only be given a 300HP engine some of the time, rather than hogging the 600HP engine.
So, the long and short of it is that there is no "magic bullet" in multithreading and dual core that will make things automatically run faster - it depends on how many small tasks there are, how much delay was introduced into the process when one task had to wait for another, etc. Some processes, by their nature, will be as fast (or potentially slower) with multithreading than without it, that is true.
Now someone with more technical knowledge of rendering and multithreading / dual core can step in and correct my mistakes :)
HTH!
Tom |
Post by noko // Jun 3, 2006, 2:26pm
|
noko
Total Posts: 684
|
For rendering, Lightworks and Vray I see almost a two fold increase in render speed. I do this by enabling and and disabling affinity parameter in windows XP. Confirmed by task manager cpu usage of both cores. When TS7 is set to use both cores, during render both cores will be at 100%, when set to one only one shows %100.
Post processing and initial part of render, like creation of shadow maps seems to only use one core, but actual rendering is much quicker in the end with a dual core setup.
I have an Opteron 165, normal or stock speed of 1.8ghz, overclocked to 2.52 ghz as stable as can be. I expect to push that up once I do a volt mode to my motherboard for cpu but in no hurray. |
Post by Emmanuel // Jun 3, 2006, 9:11pm
|
Emmanuel
Total Posts: 439
|
Thanks for the comments Tom and Noko,
I am about to order that configuration :
Motherboard ASUS A8N-E (I don't think I will use SLI as tS's Lightworks and Vray are still using the CPU for rendering)
Graphic card ASUS Extreme N7600 GS (less electricity consumption and fanless)
4 Gb of CORSAIR sdram-ddr (2x TWINX 2Gb)
Processor AMD Athlon 64 X2 4800+
two HD Seagate SATA II 160 Gb in RAID 0 mode
Windows XP Pro Edition 64
Antec Sonata II case (All my computers are in Sonata cases!)
I will try to find the time to write a report and a tS7 render time comparison between that config and my "old" one :) |
Post by Emmanuel // Jun 4, 2006, 1:56am
|
Emmanuel
Total Posts: 439
|
Oh! one more thing.
I may have to choose between two fanless ASUS graphic cards :
EN7600GS 512M of DDR2
EN6600GT 256M of DDR3
In other words, a card with a slow ram but more ram, or a card with a fast ram but only 256 Mb.
It seems to me that the video ram is mostly used to store the textures.
In a game, the card needs to quickly load new textures to display. But in a tS scene, textures are only loaded once... but we may need more space for more or bigger texture files, isn't it ?
If so, the EN7600GS 512 DDR2 should be better ?
Or am I missing something ?
Any comments ? |
Post by noko // Jun 4, 2006, 6:31am
|
noko
Total Posts: 684
|
XP Pro 64 will take you beyond the 2gb per program limitation of XP Pro 32bit but after reading a number of reports of driver problems to lack of driver support of hardware I stayed away from it for now and waiting for Vista. I've ranned into the 2gb limitation with TS7 with a highly detailed scene but that is rare. Will be interesting if that works out for you so please let us know.
As more tools become available for the new modeling side or Player, tools more advance then modeler side, I recommend you get as good as graphics card as possible. I recommend a 7900GT (about $300 for 256mb) or X1800XT (256mb version for around $230, 512mb version is around $300, both have better performance then a 7900GT) instead. Better would be a X1900XT which can be picked up for around $400. Not only will the realtime renderer get better shaders and make use of all those pixel shaders but for doing 2 to 10 million polygon modeling will also be possible, that is with the right equipement. If cost is a factor, get a X1800XT 256mb version. Hope this helps. |
|
