Lowest 16k time in SuperPi

[zakelwe]

Now that it can measure to thousands easily and seems to be cheat free at present who can do the quickest 16k run ?

I just did a 0.344s time as a starter with a very slow machine. Can anyone break 0.1s ?

Regards

Andy

[2fink]

omg, my p4 @ stock needs 0,625 seconds (without real-time tweak) and 0,453 (with tweak).
maybe i should oc it to get under 0,4s :P

edit: p4 3,0ghz (26x115) /realtime-tweak /ram @ 154mhz
2-2-2-5 = 0,391 :confused:
2-2-2-6 = 0,375
2-2-2-8 = 0,375
2-2-2-10= 0,375

[zakelwe]

http://www.aocb77.dsl.pipex.com/superpi.jpg ]

Well this is interesting, although the new version measures down to 0.001 seconds when you do very fast runs for 16k then you still get quantised results as with the 1M runs in the original version, ie you jump from 0.188 to 0.172 rather than going in a linear fashion.

I thought this might be due to disk latency so I used a RAM drive and still no change. I then assumed RAM latency so went from 12x243 to 11x265 and still the stepping was seen. Would be interested to know what causes this.

So if it goes 1.88 to 1.72 I wonder what the next step down is ?

Can someone find out ?

Regards

Andy

[snq]

Well this is interesting, although the new version measures down to 0.001 seconds when you do very fast runs for 16k then you still get quantised results as with the 1M runs in the original version, ie you jump from 0.188 to 0.172 rather than going in a linear fashion.

I thought this might be due to disk latency so I used a RAM drive and still no change. I then assumed RAM latency so went from 12x243 to 11x265 and still the stepping was seen. Would be interested to know what causes this.

So if it goes 1.88 to 1.72 I wonder what the next step down is ?You're correct. It depends on the OS you're running. I believe on NT based systems the timer resolution is around 15 msec, on 9x it's more like 30.
I know one way to get around it and get a VERY high resolution timer (as in nanosecs) but I'm not sure it'll be more accurate in the end because it will need to be calibrated, and for calibrating you'll need these inaccurate windows time functions anyway :)

[zakelwe]

You're correct. It depends on the OS you're running. I believe on NT based systems the timer resolution is around 15 msec, on 9x it's more like 30.
I know one way to get around it and get a VERY high resolution timer (as in nanosecs) but I'm not sure it'll be more accurate in the end because it will need to be calibrated, and for calibrating you'll need these inaccurate windows time functions anyway :)

Thanks for the info sng, I doubt I'll be jumping down to the next level anytime soon.

Regards

Andy

[coop]

.015

[snq]

Heh..
good example of processtimers lag :D

[coop]

Would that be the lag due to the time it takes electricity to go from the hd to the cpu? It is what it says it is, :)

[snq]

0.015 secs is impossible unless your system is at least 10 times faster than mine. My system isn't quite the fastest available, but I seriously doubt yours is actually 10 times faster :)
Anyway, the lag is caused by the fact that processtimer does not know exactly when the calculating starts or stops, it starts counting just after superpi starts calculating, and stops counting just after calculating has finished. So there's lag at both ends. But at start there is more lag because superpi is taking all your cpu power. At stop, all cpu power is available so processtimer will find out sooner. So at start there might be a lag of 0.100 sec and at end only 0.010 sec, and processtimer will show you a time that is 0.090 sec too little.
My best with 16k is 0.186 iirc, with the same config the best 1M time was 38.8 sec. The WR superpi 1M is somewhere around 20 secs right? About twice as fast as my time... So lets say the fastest stuff available clocked to the max you could get half of my time, that would make 0.100 sec or lower an extremely impressive time.
But 0.015 is hardly possible :)

[coop]

I am fairly sure it times the process. It is not linear, Super Pi calcs for incremental test. It takes more time for 2m than 1m x 2, or 1m x 8 for 8m. Thus it would take less for 16k compared to 1m. My 1m is only 31", but it takes me 1'12" for 2m, and 5'40" for 8m.

[snq]

I know it's not linear when comparing 1M to for example 2M. But speeds should be pretty much linear when comparing 1M on one machine to 1M on another. And the same goes for comparing 16k :)
Try running 16k with my modded superpi and you'll get the actual time it took. You can't get a more accurate time than that. And let us know here, I for one am curious and want to finally have proof of my theory that processtimer is lagging :) Download here (http://superpi.radeonx.com/) :)

[snq]

Actually I suspect that something else caused you to have a time of 15 msecs.

You didn't press the refresh button :D

No delta value shows in your screenie and over here when I start up superpi it already shows a 0.015 sec time, which happens to be the timer frequency on most or all xp/2k systems ;)

[coop]

You are correct, I tried it with the test between the 16k and 1m, and the process timer corresponded with the pi time (except for the decimal val). When I got to the lowest it was .234. .015 I think is how long it took to come on. Thank you also.

[snq]

Yea, 0.015 is what it takes to load superpi I guess.
Or a bit less or more, but its rounded to 0.15 because that's the timer frequency.
You should be able to get better than 0.234 tho imo. I found that running the same test twice gives better results the 2nd run. On the larger tests it doesnt have as much impact but on the small tests such as 16k it gives me a big difference.

[coop]

I just did one with your version and it was faster, .220. That is a really nice program you have there.

[RAndomaN]

can i post here, can i post here... hmmm!
http://www.daredevil.nildram.co.uk/pics/superpi16K.jpg
Thats what i managed to do a superpi 16K in with my A64 @ 2.6Ghz

[coop]

http:// http://img135.exs.cx/img135/9204/16k1871eh.jpg (http://www.imageshack.us)

[wwwww]

my rig at my 24/7 speeds.

EDIT: oh i thought u wanted lowest as in slowest...i made a proggy which does a hole load of processing in the background - oh my copy only says 0 sec still - might be time for a update....

[sabrewolf732]

wow i get .625.... a 2.43Ghz sempron 3100 270MHz ram 2.5-4-4. in 1million I get like 38s.

[Ailleur]

0.219
Cpu only at 2500mhz with all the crap you see running in the background

http://www.alienhax.us/uploads/spi16k.JPG

[[XC] leviathan18]

0.297 after three consecutive running win xp pro amd athlon 3500 @ stock

[HARDCORECLOCKER]

:slobber: New high score :party: :

http://img168.echo.cx/img168/4473/16k2nm.jpg

:toast:

[coop]

nice, notice how this ver. of SPi "jumps" to certain time segments for lack of a better term. Maybe it ends on the exact time anyway. (Could you have run .150 but the next stop was .156?)

[HARDCORECLOCKER]

nice, notice how this ver. of SPi "jumps" to certain time segments for lack of a better term. Maybe it ends on the exact time anyway. (Could you have run .150 but the next stop was .156?)

:D Don't know mate, just did it for fun. Take a look at the right down corner - done while NORTON running in background....... :p:

:toast:

[TMM]

:cool:
24/7 speeds (sig)
http://i2.photobucket.com/albums/y3/TMMTMMTMM/PI_16.jpg

[Sephious]

Settings are what the sig shows.

http://members.shaw.ca/sephious/spi.jpg

[nordik]

0.219 w/ stock fx-53

[Magnj]

http://i4.photobucket.com/albums/y132/Magnj/settings.jpg

[coop]

still at it http://img235.echo.cx/img235/2392/16kfst7oy.jpg (http://www.imageshack.us)

[i found nemo]

????

[cpulloverclock]

http://membres.lycos.fr/cpulloverclock79/Informatique/3031.gif