hyper threading analysis

[zabomb4163]

ok.....

new test. run superpi 1million with 3dmark 2001 running. post the resulting superpi results. i will post amd64, p4, and amd xp results tomorrow.



2ghz amd xp
2.6ghz p4c
3000+ amd64

sorry i dont have more rigs :(

[Mini]

Ok... Lets do that.. But hey lets do a 4MB test on superpi... Im not home... 4MB and 3dmark01... at same time...

But hey matbe it doesnt perform better? Maybe it just smooths everything else up...

I would rather do a SuperPI and a game... Than do a superPI and Benchmark...

[Anemone]

This should be interesting :)

Thank you to owners of both!

[Soulburner]

I think you people with the complaints of performance hits with HT on need to stop running Win2k.

[Mini]

Np... I always wanted to see a test like this.. I hope that he has winXP otherwise it might be a bit "Unfair" some might say???

[zabomb4163]

(please do not comment on the scores. they are not fully tweaked. i am looking for a % difference in the scores)

2.6C P4 - 3dmark 01 looping Car low
super pi - 1 minute 21 seconds
super pi without 3dmark - 1 minute 1 second
25% difference

AMD XP - 3dmark 01 looping Car low
super pi - 1 minute 50 seconds
super pi without 3dmark- 1 minute 1 second
44% difference


-i would appritiate it if others would do test in the same manner. I do not want to hear "feels faster". i absolutely hate subjective measurements. i will update my list with every machine i have at my disposal.

[freecableguy]

How are you supposed to do that. As soon as I try to switch to start SPI 3dmark2001 quits saying it lost focus. Also, what size SPI? 1M?

[zabomb4163]

Originally posted by freecableguy
How are you supposed to do that. As soon as I try to switch to start SPI 3dmark2001 quits saying it lost focus. Also, what size SPI? 1M?

yes 1 million

hit start for super pi then click benchmark for 3dmark01. remember tell it to do car low 3 times.

[LilGator]

Originally posted by freecableguy
How are you supposed to do that. As soon as I try to switch to start SPI 3dmark2001 quits saying it lost focus. Also, what size SPI? 1M?

1M for sure. I would open both 3DM01 and Super_Pi. Open the dialog to start the 1M test, and as soon as your are ready hit start, and immediately hit bench on 3DM01...

Then once 3DM01 finishes (or just quit it after 2mins in his case) check out what score Super_Pi left...

I think it's a ridiculous test, as we don't know how much Super_Pi hurt the 3DMark scores. For all we know, the P4 could be giving a better Super_Pi result at the expense of a massive 3DMark score plummet...

The processing power has to come from somewhere, nothing comes from nothing...

[zabomb4163]

LilGator, heres a thought......run the test on one of your amd64 rigs and post the scores.

there are too many opinions on both sides that are not based on anything.

[zabomb4163]

AMD64 3200+
seconds without 3dmark01 car low looping - 41 seconds
seconds with 3dmark01 car low low looping - 1 minute 15 seconds
45% difference

[LilGator]

How often in real life are we going to run Super_Pi and 3DMark01 at the same time, and piss in our pants over how bad it was ?

If you bench, A64 wins hands down anyway.

If you want to compare everyday apps multitasking, then do so... I really don't see the point in all of this. And like I said, nothing comes from nothing, P4 ain't got dual cores...

Show the 3DMark01 results as well, otherwise, you don't know where all the processing power is being used...

[WesM63]

Originally posted by LilGator
How often in real life are we going to run Super_Pi and 3DMark01 at the same time, and piss in our pants over how bad it was ?

If you bench, A64 wins hands down anyway.

If you want to compare everyday apps multitasking, then do so... I really don't see the point in all of this. And like I said, nothing comes from nothing, P4 ain't got dual cores...

Show the 3DMark01 results as well, otherwise, you don't know where all the processing power is being used...

Well said ;)

[saaya]

xp@10x200 (256kb L2 cache)
dfi infinity
512mb 2.0-4-4-4-11-13-15
geforce 4mx 440-8x@400/675

auperpi:
1M 52s
2M 1m 56s
4M 4m 13s

3dmark2001:
8382




superpi 1M+3dmark2001:
1m 08s (-30.7%) 7561 (-9.8%)

superpi 2M+3dmark2001:
2m 43s (-40.5%) 6433 (-13.3%)

superpi 4M+3dmark2001:
6m 10s (-46.2%) 4403 (-48.5%)

since i was running no nature test (g4mx) both, superpi4m and 3dmark2001 finished at exactly the same time!

so while running 3dmark2001 and superpi for the same period of time the system performence was split almost exactly 50/50 between the two apps.


please post your results!

[Dagalidis]

SAAYA This also means same priority on both apps..... :D

[saaya]

Originally posted by Dagalidis
SAAYA This also means same priority on both apps..... :D yes, both are set to normal in windows task manager.

can you please run the benches and post results as well? pleeeeease :D

[Mini]

I will do it as soon as i get home... In about 3-4 hours...

[Dagalidis]

Originally posted by saaya
yes, both are set to normal in windows task manager.

can you please run the benches and post results as well? pleeeeease :D

Not possible .....
My mobo is on RMA procedure (K&M Elektronik).....
Next week will be ready again....:(

[saaya]

Originally posted by Dagalidis
Not possible .....
My mobo is on RMA procedure (K&M Elektronik).....
Next week will be ready again....:(
ah, too bad :(

i didnt know you have k&m in greece :eek:

[Perc]

hey guys i did a few tests for the project. what i wantd to do was run cpu,multimedia,and memory in sandra to compare with the HT on to see if there is any performance hit or not? then i did a test with toast running while suprepi was doing the 2m test then i ran sandra's cpu test. i left it all run till superpi finshed. now ill go back and do all that with HT enabled... here are the screenies...

Btw system was run at 16x250 4000mhz 1:1 1.38v for all tests and no HT .... be back with HT enabled results...


enjoy perc,

[Perc]

OK heres the multimedia test...


perc,

[Perc]

and here is the memory test. with no HT...


perc,

[Perc]

ok last but not least here it is with no HT multitasking just fine....


perc,

[Perc]

ok im back! i ran everything the same way but this time with HT enabled. this should be a decent comparsion between the the HT on or HT off....

first one is HT enabled cpu test. this can tell us if there is a peformance hitt by disabling it or that its not much a differance? lets find out..


enjoy.....

[Perc]

now this is test 2 the multi media test with HT enabled..


perc,

[Perc]

ok heres the memory test with HT enabled..

perc,

[Perc]

last but not least the one with HT running and toast super pi 2m and cpu test all running at the same time. a let the super pi 2m finsh before i took the screen shot so you can sompare the 2m scores on both ht and no ht... ok guys hope you enjoyed this...

ill post back with all the info borken down making it easyer for everon to look at...


peace perc,

[Perc]

ok the differances in the HT being off and HT being on for the tests i ran. here are the results....



HTT OFF
CPU Test 10494
2680/4956


HTT ON
11056
4622/8433



Multi Media
HTT OFF
22589
25943


HTT ON
28373
38192

Mem test

HTT OFF
6030
6035


HTT ON
6062
6059


multi task test HT OFF

CPU
10062
2681/4956


Super PI
3:38 2m test

Toast running the whole time.

Multi media test HTT ON

CPU
10995
3986/8073


Super PI
2:46

Toast also running the whole time.



ok guys there it is. sorry if it sucked but i think it shows that HT does help more then it doesnt. infact i seen NO gain in having HT off....

peace perc,

[saaya]

uh, why did you use sandra?

i dont like sandra as its highly synthetic and has proven to show scores waaaay off of the real world performence.

why were you running toast in background?? that makes it almost impossible to see how the calculation performence was split as we dont know how many resources toast ate up with ht enabled and ht disabled, so its not really possible to draw a real conclusion from the numbers.

please run 3dmark2001+superpi like we did.

[Perc]

only time toast was running was on the last test the multi tasking test that was it...

peace perc,

[PGgagne]

But in Folding wich would preforme better??? HT on or off?

Thanxs

[saaya]

ht on and run 2 clients at the same time

perc, why did you run toast?

[Kanavit]

wow, it looks like hyperthreading indeed is useful especially when mult-tasking! Good work zabomb4163

[Perc]

Originally posted by saaya
ht on and run 2 clients at the same time

perc, why did you run toast?

i told you i only ran it on the last test the multitasking test the others are straight up no htt or htt nothing else running but the test its self man not toast in those first 3 tests HTT on or off. the info is there and there is a diff between the two. i dont knw man just trying to be helpful..


CPU test

HTT OFF
10494
2680/4956


HTT ON
11056
4622/8433



Multi Media

HTT OFF
22589
25943


HTT ON
28373
38192

Mem test

HTT OFF
6030
6035


HTT ON
6062
6059


all the above tests were run with either HTT on or HTT off there was no toast involved. i did those tests to prove HTT made a diff performance wise and that there was a performance hitt when truning htt off. only thing i screwed up in was the multi tasking test cause of the toast running. you mean to tell me all that info above isnt usable? dont understand why theres no toast involed....



later perc,

[saaya]

the other scores are... fine... if you trust sandra. i dont trust sandra at all. sandra used to be a very nice bench, but the new versions are crap and dont reprsent real peformence anymore, not at all.

please answer my question, why did you run toast?

[-=TriX=-]

and why is sandra untrustworthy? Does it give false scores? Becasue it is synthetic? Most benchmarks are. Would some video encoding, 3d rendering, game benchmarks be more legitimate in your opinion?

Perc said he only ran in the multtasking test....so perhaps he ran it to provide a more efficient and accurate hyperthreading test. 3 applications running would be a heavier load than two applications in a multitasking test. I don't get why you keep pressuirng him. :\

[Perc]

Originally posted by saaya
the other scores are... fine... if you trust sandra. i dont trust sandra at all. sandra used to be a very nice bench, but the new versions are crap and dont reprsent real peformence anymore, not at all.

please answer my question, why did you run toast?

beacause i thought it would put a big enough stress load on the system so that when running other apps it would show the system is actualy doing somthing. ohter then that dude i dont know im learning from you guys... rember :D

peace perc,.

[saaya]

Originally posted by -=TriX=-
and why is sandra untrustworthy? Does it give false scores? Becasue it is synthetic? Most benchmarks are. Would some video encoding, 3d rendering, game benchmarks be more legitimate in your opinion?

Perc said he only ran in the multtasking test....so perhaps he ran it to provide a more efficient and accurate hyperthreading test. 3 applications running would be a heavier load than two applications in a multitasking test. I don't get why you keep pressuirng him. :\

the idear was to run two or more apps at the same time and then compare how well they are running. its impossible to tell how "well" toast was running, how many resources it got etc...

so it made the meassurement pretty useless.

and no, i dont distrust and dislike sandra because its synthetical, superpi and 3dmark are just the same, but their results are closer to real world performence, as i said.

and sorry if i sounded harsh perc :toast: i just didnt understand why you ran toast and wanted to know why :D

perc, can you do me a favour and run 3dmark+superpi? pleeeease? :D

[Anemone]

And the car, with all 10 people grabbing at the wheel careened wildly off the road and crashed into the gully.

Any way we could get the described tests back on track?

[Perc]

Originally posted by Anemone
And the car, with all 10 people grabbing at the wheel careened wildly off the road and crashed into the gully.

Any way we could get the described tests back on track?

sure thing slick post your results :rolleyes:

@saaya

yes i will run 3dmark and super pi ... ill do it tonight my time....


perc,

[saaya]

thx :toast:

[Salkcin]

Saaya->

I think you can be pretty sure that Toast runs in highest priority - I mean, it will take all the CPU power it can get... 49-50% (No application can use 100% ALONE since HT was meant for multi-tasking)

All->

Another interesting thing. Would HT be power saving? under normal circumstances you would never use more than 50% CPU power... so, what is the temperature diffrence between 50% and 100% load?

HT is nice... allways CPU power left over - no more hour-glass! feels pretty quick allways :D

[saaya]

lol@hour glass, havent seen it since i got my 1700+ :P

and if you run 3 apps, why should toast get 50% resources constantly while the other 2 apps share the other 50%? that doesnt make much sence to me. :P

[Lithan]

Don't tell me you aren't realizing that when you have three programs all running real-time with high Cpu demands it pretty much puts you back to square one, HT or no HT.

No, really the real world is just like you picture it. Two high-demand apps walking along holding hands in an Intel garden of Hyperthreading, saved from the horrible task managing of our modern cpu's and operating systems, which are no doubt confused when facing the task of running multiple programs at once. It's not like they do that all day, every day; now is it.

[Salkcin]

saaya->

the programs run in a priority... lot of folding software runs "idle" wich is the lowest priority. Anytime any other software needs CPU power the folding would give up the power it is using... Toast would run high priority while the others would run normal. That's why Toast should get 50%...

...I'm not sure Toast runs in the highest, but I think so. Having a CPU stability tester (toast/prime95) running with 1 other application (benchmark) would test if the HT CPU really performs "200%" - 2x 3,2GHz (each 100%) for example, but the truth is more like 150% effiency of HT.

[Lithan]

Originally posted by Salkcin
but the truth is more like 150% effiency of HT.


Did you just say that HT raises a cpu's IPC?

[Salkcin]

Lithan->

IPC?

Hang on here - what I meant was... imagine you would have 2 CPU's (one phycial and one logical - Hyper-Threading)... each CPU potentially capable of performing 100% like the speed they run - thats "200%" CPU power in theory.

Like having a 3.2GHz with HT - "each" CPU should perform like a 3,2GHz at full speed, but when one of them is full loaded the other one will not perform like a 3,2GHz at full speed so the overall CPU effiency would be like 150%. If the CPU would get 100FPS in a game with only the one of the CPU's full loaded (by the game) then if you ran 2 clients of the game at the SAME time you should get 100FPS in each client, but the truth is more like getting 100FPS in the one client and 50FPS in the other?

Hope what i wrote was'nt a mess :D

[Lithan]

Well FPS isn't a good example. But if your saying one program will work at 100% and another at 50%, you are wrong. HT can not ADD performance to a cpu. Ideally both programs will work at 50%, so it will be like a dual 1.6ghz rig. Of course I believe that HT works at other divisions than 50/50. So if one program demands real-time attention, but only requires 800mhz, the other which wants all 3200mhz will wind up getting 2600mhz (minus the mhz chewed up by all other running applications.) That's how I would figure it works at least. It sounds good, and at times has it's uses (mainly for a low-maintenance server environment when multiproc systems can't be afforded) but all in all, I consider it completely useless in a desktop environment. (Come on when people are complaining about choppy music when they are gaming, they are lucky I don't hit them and call them an idiot for needing HT to fix that.)



IPC means Instructions per Clockcycle.
Increased IPC means a cpu does more work per cycle.

[Salkcin]

What I wrote was all theory. What I tried to tell was that Hyper-Threading does'nt double the performance when multitasking. If you run two 1M Super-Pi's at the same time (loading 100% in the task manager - 50% each) they would'nt both finish at same time as if you ran one Super-Pi (loading 50% looking in the task manager). HT does'nt work like two CPU's independent of each other...

...I'll leave it here. I don't seem to express myself clear enough... english is'nt my native languech soooo

EDIT: The one result should be the same as before and the other should be slower instead of the same... even if "both" CPU's are the same frequency and loads the processer 50% each... maybe both would be slower, but the same... can't tell

[Lithan]

I see what you are saying now. You are saying that HT is actually less than 100% efficient. Not that it is 150% efficient.

100% efficiency would mean that running two superpi's at once would finish in the same time it took to run superpi twice (one after the other) on a non HT system.

[Salkcin]

Lithan->

A single CPU is 100% efficient and a HT CPU is theoretecly 200% and that's why reallife performance is like 150% COMPARED to a single CPU...

Lets say it takes 1 minute to run a SuperPi for a single CPU. A HT CPU at same speed should theoreticly complete two SuperPi's running at same time in 1 minute, but will complete them in about 1 minute and 30 seconds where the single CPU would use 2 minutes to complete 2 SuperPi's. Saying the Single CPU's 1 minute is 100%, my 200% and 150% are "true", but I know it's "impossible" for a CPU to load or be more efficient than 100%...

...lets take it like "performance"... performance can be more than 100% compared to something :D

[Lithan]

You are wrong.

A 3.2ghz without HT would finish two runs of Superpi as fast or faster than a 3.2ghz with HT running them both at once. (minus the time it takes to restart superpi after the first run).

And a HT CPU is NOT theoretically 200% efficient. It is 50% efficient times two.

[Karnivore]

Let me just point out this thread is for analasys and results, not debating if it works.

[saaya]

why does nobody post results? come on!

Salkcin,
no cpu reaches 100% efficiency as some parts of the cpu always have to wait for other parts of the cpu to give them new input. ht is supposed to increase the efficiency of a cpu by seperating most of the cpu into two parts, so that if one stage of the cpu is busy the other half of the cpu can still process data. this improves the processor efficiency in certain cases, but in most cases it decreases performence slightly (afaik) as the two parts of the cpu are less powerfull than one cpu. so while the ht enabled cpu might be able to process small instructions faster, complex instructions will occupy one of the two parts, or eventuelly both parts if there are two complex instructions, for a longer time as if the whole cpu would work on it, wich results in a performence loss wich means a ht enabled cpu is not faster or even slower in most apps.

of course you could argue that most apps are not ht optimized, but its a fact that its hard work and not worth it for most apps to be optimized for ht.

running superpi and 3dmark2001, two none ht optimized apps, should show how well ht works though as we are running them at the same time , wich is like running a bench that is ht optimized like sandra, only that this time we are using two tests wich are proven to reflect actual real life system performence, not like sandra.

[-=TriX=-]

had to get the pins on my p4 repaired, should be arriving soon I'll post some results as soon as I can.

[Lithan]

I'd run some benches at stock (dell box), but I'm on win2k... so results probably wouldn't mean much.

[-=TriX=-]

what's the problem w/ win2k and HT? fill me in please, I've been out of the loop for a bit and this is my first P4 HT system :)

[saaya]

win2k doesnt support ht afaik.

[-=TriX=-]

ahhh I see, thx for the info

[saaya]

your welcome :) :toast:

[accord99]

Originally posted by Lithan
You are wrong.

A 3.2ghz without HT would finish two runs of Superpi as fast or faster than a 3.2ghz with HT running them both at once. (minus the time it takes to restart superpi after the first run).

Why don't you test this? I did it on a relatively slow 3GHz Dell and two 1M digit runs without HT takes 100s, two simultaneous runs with HT enabled takes ~68s, an increase in throughput of nearly 50% and an effective time of 34s for 1M digits.

And for the most part, this trend holds everytime you run two CPU intensive applications simultaneously on a HT processor, though 50% tends to be rare, the typical performance increase is around 20%, sometimes more. I recall ~30% for Pifast, 27% for Lame. HT also turned the P4 from an average performer on SETI to a top performer.
Here are some more benchmarks:
http://forums.anandtech.com/messageview.cfm?catid=28&threadid=1180277&highlight_key=y

[saaya]

why does nobody post scores? :/

[Lithan]

Originally posted by accord99
Why don't you test this? I did it on a relatively slow 3GHz Dell and two 1M digit runs without HT takes 100s, two simultaneous runs with HT enabled takes ~68s, an increase in throughput of nearly 50% and an effective time of 34s for 1M digits.

And for the most part, this trend holds everytime you run two CPU intensive applications simultaneously on a HT processor, though 50% tends to be rare, the typical performance increase is around 20%, sometimes more. I recall ~30% for Pifast, 27% for Lame. HT also turned the P4 from an average performer on SETI to a top performer.
Here are some more benchmarks:
http://forums.anandtech.com/messageview.cfm?catid=28&threadid=1180277&highlight_key=y

1. I run win2k.
2. I don't have too. It isn't possible for HT to make a cpu run faster than it naturally can. Perhaps the cpu can recognize it is running similar operations twice and in effect "cheat". The only other possibility is that Superpi is not 100% utilizing at all times, and by running two examples you reduce idle clocks.

[Perc]

Originally posted by saaya
why does nobody post scores? :/

ill post some for yeah :D

perc,

[accord99]

Originally posted by Lithan

2. I don't have too. It isn't possible for HT to make a cpu run faster than it naturally can. Perhaps the cpu can recognize it is running similar operations twice and in effect "cheat". The only other possibility is that Superpi is not 100% utilizing at all times, and by running two examples you reduce idle clocks.
No, HT takes advantage of the fact that there are idle execution resources that are available due to sub-optimal code, cache misses, memory access, etc even for most CPU intensive applications, which can be used if a second CPU intensive application can be scheduled simultaneously. If it was just idle cycles, then a non-HT processor will see a similar performance increase.

And no, the CPU does not cheat because you can do the same thing with two different applicatications. For example running Super Pi and Quick Par together results in Super Pi running at 67% of the original throughput and Quick Par running 58% of the original throughput. A non-HT processor would be 50% for both applications.

[-=TriX=-]

P4 2.4c @ 292FSB 3.5ghz
5:4 RAM Divider 234 DC, PAT enabled CL2-2-2-5

1M SuperPi w/ 3dmark in background 43 seconds
!M SuperPi no 3dmark in BG - 37 seconds

Not much difference, I'll run it again later.

[saaya]

please post all results

superpi 1m 2m 4m

3dmark2001

3dmark2001+1m
3dmark2001+2m
3dmark2001+4m

thx :)

[-=TriX=-]

you want a full 3dmark01 run and a superpi at the same time? I don't think that's a very good test, because the superpi will be done before the 3dmark01 run is over...

Anyhow, ill get to it this weekend. I'm still trying to get some answers in my thread here: http://www.xtremesystems.org/forums/showthread.php?s=&threadid=40614

the pro bios on my 9700 has been causing alot of instability and I lost the original in a format :(

[saaya]

1m will finish around the car low
2m will finish at the end of car high
4m will finish around lobby high

at least thats how it was on my rig.

feel free to run superpi 8m + 3dmark2001, i can add that score as well

[saaya]

please post results!

[Soulburner]

So you guys are trying to do a Hyperthreading analysis, with programs that are not programmed for HT?

How about you use something that is?

How do you know the CPU is even using the HT unit here?

Look in your task manager...if you did not set one of the processes to Affinity 1, they will ALL run on CPU 0...thus making your analysis pretty pointless.

[Iridium192_217]

Originally posted by Soulburner
So you guys are trying to do a Hyperthreading analysis, with programs that are not programmed for HT?

How about you use something that is?

How do you know the CPU is even using the HT unit here?

Look in your task manager...if you did not set one of the processes to Affinity 1, they will ALL run on CPU 0...thus making your analysis pretty pointless.

He brings up good points. I really can't trust any of these results. There are to many unknowns.

[saaya]

we are running two apps at the same time ;)

please correct me if im wrong, but running two apps at the same time DOES make use of HT, thats the whole point of ht, beeing able to run two apps at the same time :P


i think the fact that nobody with a P4 posted any results in here means you all ran the tests and saw no real improvement, so ht is not as helpfull as some people claim it is.