My new rig sucks at SETI! [Archive]

markodude

07-10-03, 04:18 AM

Hi People
Ive just benched my new rig, its a P4 2.4C 800 at 3.2ghz on 1066FSB, Twinmos DDR400 at 5:4 (DDR424) Cas2.5-6-3-3.

I benched it using the ocsetiteam.com wu and it got a 2.10hr time, which is very poor IMO, sheesh even Athlons are whooping me! ;)

Why do P4's suck at seti!? I was nearing these sort of times a year ago on single channel Athlon systems.

Demont

07-10-03, 07:39 AM

Oh wow 180 degrees difference from some other threads. But to answer your question, yes p4's do need lots of clock speed to get those fast work unit times. Check out the benchmarking page on ocsetiteam.com though... it can be done. (but then many people have AMD systems over 2.8 and over 230 fsb, not many people have Intel systems higher than 4.2! So hopefully it will be taken back soon away from its temporary holder :D Congrats for now, hipro)

Of course, you can pick up the slack with hyper threading, by enabling 2 sessions of SETI at once within SETI Driver or similar. Then you will get slower work unit times, but it will crunch two at once... giving you 16-18 WU's a day maybe. But that doesn't really help you on the benchmark... buy a prommy and cross your fingers, thats all I can tell you. :)

markodude

07-10-03, 07:53 AM

Im gonna go and ask the guy I sold my 2400+ and 8k3a for it back!
Its gonna take a lot for me to get any faster on this rig - I am faster than the quickest production Intel system available(3.2/800/DDR400), and still its not fast enough. It means getting the solder out to mod the VDimm voltage and like you say, a prommy or watercooling with TEC to get the CPU higher.

Burning Phoenix

07-10-03, 08:20 AM

I think thats an awesome time. Besides not a single AMD cpu can outproduce the WU's that P4c will with HT and runngin 2 instances.

markodude

07-10-03, 08:29 AM

This is true, I was only running one process - I take it I should run 2 in order to crunch most efficently?
Whats you 2.4C/P4C clocked to BF? I need to get me a maze4 and some water! Done the benchie yourself?

Burning Phoenix

07-10-03, 08:37 AM

MInes only at 3.0 becuase i leave it unmonitored for weeks and i trust a lower overclock to be stable with SETI. 2 of my systems are sitting against a wall with only a CAT5 cable and power supply to it nowhere near a monitor, keyboard, mouse and such.

Mictlan

07-10-03, 08:44 AM

Besides, the main advantage of using now a PIV is getting it crunch two WU at once using HT. That way even if you get higher times/WU you can surely crunch more of them. A top of the line Athlon system is currently getting 12 WU per dar (@2 hours/WU), and most of the users of a PIV are getting 16 Wu per day. Also PIV run cooler and with less power than Athlons so you operations cost will be lower by a tad ;)

fruitcake173

07-10-03, 08:45 AM

How can we do the benchie correct with HT, i mean i can do 2 wu's in 3 hours that would mean 1 wu in 1,5 hour but if do a single wu it would probably take me about 2 hours also.

Burning Phoenix

07-10-03, 10:10 AM

I think we should try the SETI bench with 2 instances and figure the average to see what we can obtain. Then place a spot in the benchmark page to note the sytem benched was with HT enabled and 2 instances.

Mictlan

07-10-03, 12:03 PM

It is not possible to use the benchmark to properly evaluate the impact of HT in your crunching. If you use the logical approach of crunch two benchmark WU at the same time an then take the time it took to crunch them and divide it by the number of instances, you are wrongly alterating the results as each WU took the given time not half of it......If you wanted to know how fast you download a WU from berkeley (or how many WU each day you'll produce) you could do that.

Is like trying to say that if you have 20 CPUs dedicated to SETI and manage to pop a finished crunched in each one in 5 hours, then you are crunching each WU in 5 hours/20 WU: 15 minutes......you could get a new WU to crunch and drop a 5 hour-WU each 15 minutes, not the other way ;)

LandShark

07-10-03, 02:03 PM

to properly bench a P4 "C" CPU, you should disable HT via BIOS, and only crunch 1 WU at the time using the benchie WU.

actually, 2:10 from only a 3.2Ghz P4 isn't that slow, but you should be able to shave a couple more mins by tweaking the machine. look at my old entry, my 1.8a @ 3.42Ghz could only manage 3 mins faster than yours w/ 200more MHz!! that means your higher FSB & memory speed is helping you!

jd & TC had sub-2hr entry w/ their P4C but about 300 MHz higher than yours. they bench 1WU at a time w/out HT enabled!

plus the advantage of a P4 "C" isn't how fast it could finish a WU!! it's how much WU it could finish in given time!!

07-10-03, 02:07 PM

You know what would be interesting if it could be done is to see how many of the benchmark units your system could complete in a 24hr period running HT and 2 clients. The HT needs to be addressed in the benchmarks somehow. For instance my 3.0C @ 3500 will complete 1 work unit in about 1 hour and 57 minutes - but if I run two clients at the same time it only adds about 40 minutes to the overall completion time for both units.

Kunaak

07-10-03, 02:49 PM

I have 3 AMD machines here, 2 with XP 2100's, 1 with a XP 1800.
all KT-333 and 400's with cas 2-2-2 ram, 7200 RPM drives, and fancy videocards.
they seem to be averaging about 2:30 hours per project, but my Xp 1800 machine isn't on seti, just the 2100's.

I also have a 2.4C with a Asus P4C800 D. 2 IBM 60 gig deskstars on raid zero and R9800 Pro.
this PC with one work unit was averaging 2-3 hours per project with the screensaver version.
then I got some tips to get the package thing from burning phoenix and run that. now, this same PC running 2 seti projects at the same time, gets 2-3 per project, but no screensaver version, but it's twice as productive.
this one PC almost does more then my 2 XP 2100's combined.

none of my machines are overclocked at the moment.
my prometeia is useless till I can get a P4 Kit for it, and I am not much a fan for air cooling, and leaving my PC's alone overclocked for hours on end.

but if anyone wants to do some testing of hyperthreaded and none hyperthreaded benchmarks, just tell me what you want and how you want it done, and I will give it a go :)

Burning Phoenix

07-10-03, 03:08 PM

I still think the way i mentioned is the best way to obtain the correct speed a HT system can produce WU's. Only one instance is placing a handicap on an HT enabled sysetm. It seems THG has also decided this was the way to obtain scores:
http://www.phoenixhideout.com/overclockers/pics/thgseti.jpg

SkyHook

07-10-03, 03:14 PM

Mictlan

07-10-03, 04:10 PM

That graph from TH was the subject of a lengthy debate on how to show the perfrmace gained with SMP and/or HT systems. I think that SETI, as a single thread process should be addresed only in a single thread enviroment.

The only way to clearly show the advange of a SMP or HT system should be the amut of WU the system can complete in a given set of time (on day for example).....This "let's just run two instances at the same time and the total time then is divided between 2" is not the correct way to show the gain in WU completed IMHO.

Demont

07-10-03, 04:44 PM

Originally posted by Mictlan
That graph from TH was the subject of a lengthy debate on how to show the perfrmace gained with SMP and/or HT systems. I think that SETI, as a single thread process should be addresed only in a single thread enviroment.

The only way to clearly show the advange of a SMP or HT system should be the amut of WU the system can complete in a given set of time (on day for example).....This "let's just run two instances at the same time and the total time then is divided between 2" is not the correct way to show the gain in WU completed IMHO.

Yes it should... the problem with THG is that they did not inform anyone about their testing methods there. I don't think THG's undocumented (and incorrect, as Mictlan argues) testing methods will really be good premise for your argument BP. :)

I think its great that Dual and HT machines can pump out lots of units... but a speed benchmark is a speed benchmark... I agree with Mictlan here... its not correct. Not only are there more variables... but how is our friendly neighborhood benchmark updater supposed to know if you actually did two at once? I say keep the daily output per computer and speed benchmarks seperate with no overlap.

I'm sure this arguement has been going on forever within SETI circles regarding AthlonMP/Xeon systems and now of course the P4's w/ HT.

markodude

07-11-03, 02:20 AM

One thing very apparent with P4 systems is memory settings, so have you tried getting an overclock with your system that can run more aggressive memory timing? Specifically getting your CAS down to a "2". I think it has been demonstrated in the past that even giving up a bit of FSB in favor of aggressive memory timing has been able to shave time off benchmarks.
Yeah thanks, but unfortunately my RAM wont run in cas2 mode at DDR424, even with 2.85v, looking at doing a vDIMM modification though.
I have benched the system memory bandwidth and this is the highest I can get it to in Sandra, if I run 1:1 and DDR466 sandra shows it as being slower than 5:4 DDR424.....?

Thanks

markodude

07-11-03, 02:21 AM

BTW The P4C-800 also seriously undervolts the CPU at full load, I have set 1.675v in BIOS and I get 1.55 under full load, so maybe my CPU can go a bit further yet ;) Dying to get under 2hrs you see!

LandShark

07-11-03, 03:07 PM

07-11-03, 04:21 PM

Originally posted by markodude
BTW The P4C-800 also seriously undervolts the CPU at full load, I have set 1.675v in BIOS and I get 1.55 under full load, so maybe my CPU can go a bit further yet ;) Dying to get under 2hrs you see! That doesn't sound right - what psu do you have? My P4C800 is set at 1.6 volts and it runs about 1.58 under full load according to mbm. While idle it often goes up to 1.62-63.

Burning Phoenix

07-11-03, 05:19 PM

My P4C800 also undervolts severely. I have mine set in Bios at 1.675 and according to MBM it's 1.55 at idle. I have a 350W ANtec PSU.

texasfit

07-11-03, 07:45 PM

Originally posted by Burning Phoenix
My P4C800 also undervolts severely. I have mine set in Bios at 1.675 and according to MBM it's 1.55 at idle. I have a 350W ANtec PSU.
All three of my Asus boards undervolt. It is just how Asus has been setting up their bios for some time. My P4C800 runs at 0.045vcore under bios setting while running 2Xseti instances. The Abit boards seem to all overvolt.

I wish that I had thought or seen the point about setting HT to disable when running the seti bench. Missed that but I still got 2.0189 with HT enabled on my P4C800 rig. Dang Nabit - might have got under 2 with HT off. :rolleyes:

I think that if you wanted to see the real value of HT that you would have to do as mentioned by a couple earlier. But I don't think that a 24 hour period would be correct to establish the number of wu's completed. I think that you would have to take say 20 wu's and run them until complete then divide by the number of min/sec it took to complete all wu's. (ie: 2500.5 minutes / 20 = 2.08375) This would be of no real value other than to compare to say another system like a AMD 2700. Just to see how many more minutes/hours the 2700 takes to do 20wu's.

07-12-03, 12:02 AM

psyshack

07-12-03, 12:29 AM

theres no reason to some for the stuff that mass. profuction boards will do!!!! and how the cir. really works... I hate the thought of landing thermal probes back on slug and doing all my own thermal stuff.... But the marketing hype of the cpu manufactures and there thrmal internal temps is not looking good,,,, you have the intels amds saying we are right and then the mb producers seting on there backsides and we set here back at square one trying to figure out whos right....

its so stupid

psy

Horribleron

07-12-03, 03:15 AM

My understanding is the the i386 WINNT client was optimized for the Pentium 3 CPU, which Mhz to Mhz is problably still the fastest cruncher there is for plain x86 class machines. Too bad they only made them up to 1.4 Ghz (Tullatin C). The Athlon family is very similiar in design and thus crunches well too. The P4 on the other hand, is a radicaly different design thus it doesn't process the P3 optimized client as well as it should. Some time back I tested a early P4 1.4 Ghz against a 1 Ghz Athlon T-bird, and the T-bird kicked it's butt. Same clients, and over a weeks time the T-bird finished about 30% more WU's.

Now, if there was a P4 optimized client, then all of that would change. But if you look at the list of clients to download, there isn't one. However, there is a Intel IA-64 Itanium client, and I'm sure it rips through a WU faster than I can run to the fridge for a beer. Better yet, there's even a Linux IA-64 client! Now you all know what the Intel team (and problably the HP team) is running.

Anyone got $10,000 to loan me so I can get a Itanium server?

Horribleron
Rank #54

markodude

07-12-03, 03:38 PM

Well I just setup my new cruncher tonight - P4P800 with a 2.4C. So far it's running well at 3400 with 1.6 volts. MBM is showing about 1.585 under full load with bios set to 1.6. I am using a 420 watt psu though. The 5 volt line is staying around 5.1 under full load. I've never had any problem with my P4C800 either. As I said it often overvolts while idle, and usually runs just a hair below the actual setting while under load. I wonder why our boards are so different?
I think I will try a new PSU anyuway....just in case. and maybe some more vcore when I get better HSF than the 3.06 one (which is at about 46c load, not bad considering its about 25c ambient).

I have been crunching with 2 Processes and my WU per day seem to have increased significantly so thats cool! I may even beat my GF4MX440 record on 3dmark with this rig.....:D

Greg M

07-12-03, 05:46 PM

Originally posted by Horribleron
The P4 on the other hand, is a radicaly different design thus it doesn't process the P3 optimized client as well as it should. Some time back I tested a early P4 1.4 Ghz against a 1 Ghz Athlon T-bird, and the T-bird kicked it's butt. Same clients, and over a weeks time the T-bird finished about 30% more WU's.

Now, if there was a P4 optimized client, then all of that would change.
The P4 was the first CPU from Intel to take a step backward in the amount of work it performed per clock cycle. The only way to optimize a client for it, would be in leveraging SSE2 for increased floating point performance. Intel has nobody to blame but themselves for this.

The Pentium M mobile chips are very good crunchers with processor efficiencies in the 1.x range and 1MB caches. I'm evaluating a Dell D600 right now with a 1.6GHz Pentium M that's completing high AR's in 1:39, and low AR's less than 2:20, as I recall (I'll verify early next week and update once I can get to the SETIQueue with the actual stats). The Pentium M core architecture should be where Intel puts their energies for the future, but they've go too much invested in P4 marketing (it's too difficult for them to admit MHz is not the best way of identifying real performance).

The successor to the Banias Pentium M, Dothan, is due this fall. It will release with 2MB of cache and a speed of 1.8GHz. I can already hear the sizzle...:burn:

The cheaper Itanium 2 "Deerfield" CPU's coming out late this year should also be interesting.

LandShark

07-12-03, 08:49 PM

Originally posted by Greg M

The P4 was the first CPU from Intel to take a step backward in the amount of work it performed per clock cycle. The only way to optimize a client for it, would be in leveraging SSE2 for increased floating point performance. Intel has nobody to blame but themselves for this.

The Pentium M mobile chips are very good crunchers with processor efficiencies in the 1.x range and 1MB caches. I'm evaluating a Dell D600 right now with a 1.6GHz Pentium M that's completing high AR's in 1:39, and low AR's less than 2:20, as I recall (I'll verify early next week and update once I can get to the SETIQueue with the actual stats). The Pentium M core architecture should be where Intel puts their energies for the future, but they've go too much invested in P4 marketing (it's too difficult for them to admit MHz is not the best way of identifying real performance).

The successor to the Banias Pentium M, Dothan, is due this fall. It will release with 2MB of cache and a speed of 1.8GHz. I can already hear the sizzle...:burn:

The cheaper Itanium 2 "Deerfield" CPU's coming out late this year should also be interesting.
all i can say is tooooooo much nice toys on the way, but toooooooo toooooooooo less $$ for me to play w/ them.......:cry:

markodude

07-14-03, 02:11 AM

Yeah the Pentium M was a great option, I had one a few days after launch but gave up finding a board that would work in a desktop PC, and not spending £1000 on a laptop!! I have now given it back, but if I work out how to build a cruncher with one perhaps I will get another sometime soon....