Page 2 of 3

Posted: Tue Aug 21, 2007 3:49 pm
by DonPedro
socket7!,

just in case you find a place where such a "decoupler" - thing is available please post! I am interested ....

Posted: Tue Aug 21, 2007 6:13 pm
by socket7!
I'm sorry, but i did not find time today to do additional benches. I hope to do it tommorow. I enclose a a small pic of super pi @4.30.

The system is not perfectly stable, but I'm working on it.

And a note about silent serpent: I'm hunting for it for 2 years already. No luck so far :-( .

and just a very quick question: is esdram or hsdram compatible with any socket7 system? I know that vcsdram is not (I tested it).

Posted: Wed Aug 22, 2007 7:21 pm
by Jim
Re your "Silent Serpant", that reminds me of a "Power Supply" that I built from a kit several years ago. Thing had output voltage that fluctuated quite a bit, so I designed my own smoothing system for it.

Basically when you take AC current and use a bridge rectifier to convert it to DC, the resultant output wave looks something like this :

Image

Adding a smoothing capacitor changes the form of the wave to something like this :

Image

What I did was say ok, if adding a smoothing capacitor reduces the ripple; but still leaves some ripple because of "Rise Time", (of the capacitor current), and "Fall Time", (of the capacitor current); what would happen if I put a series of capacitors across the output wherein each succeeding capacitor in the series was 1/2 the value of the preceeding one; with the series running right down to the Picofarad range. I forget how many caps I used, somewhere between 15 & 20 as I recall. The concept was that each succeding smaller cap would fill in the gap, (ripple) caused by the rise time and fall time of the preceeding cap. EG :

Image

KachiWachi knows a whole lot more about this sort of thing than I do, as to how valid my idea was; but this I can tell you : That power supply went from having an output that drifted up and down by as much as 2 or 3 tenths of a volt to being absolutely stable as far as my multimeter could measure. - 1/100 of a volt.

Wonder if the concept could be applied to the CPU power supply ?

EDIT : No Response whatsoever to this post; so I guess I'll have to try it with one of the three 300 watt Antec AT supplys that I bought. Of course since nobody cares, there will be no point in posting the results.

ridiculous things with mvp3

Posted: Wed Aug 22, 2007 7:26 pm
by cskamacska
socket7! wrote:I know that vcsdram is not (I tested it).
Whoah, VCSDRAM? If I may ask you, where did you find that kind of memory? I have been looking for vcsdram a long time now, but could not find any. :?
Did you experience any performance gain from using VC SDRAM?

Posted: Thu Aug 23, 2007 3:45 am
by kalabok
4:30 with 112 fsb ... wow. best performance i have ever seen.

btw - my record till now is 4:27 with 133 fsb on ali7.

greetz from dubrovnik

Posted: Thu Aug 23, 2007 5:46 am
by Jim
Actually he was running 124 Mhz FSB.

Posted: Thu Aug 23, 2007 12:08 pm
by socket7!
I'm leaving for holidays tonight, and as a result, I will not be able to post new benchmarks until 4th.

Info for cskamacska: I bought vcsdram last year from the guy that didn't now what he sells. I would recommend to visit a good Polish auction site: www.allegro.pl.

Posted: Sat Aug 25, 2007 4:41 am
by kalabok
jup,

sorry for my first statement, it is 124fsb indeed. still a very good result!
i would be very interested to see the superpi result with no l3 enabled.

Posted: Sun Sep 09, 2007 2:43 pm
by socket7!
just a small update:

my mem scores with everest 2.2 are (read/write/latency):
@616: 352/89/203.3
@624: 390/99/193,4

scores are not comparable with different version of everest.

I also managed to get 4.36.427 in SuperPi, but @616, L3off and wpcredit tweaks.

Posted: Sun Sep 09, 2007 5:31 pm
by DonPedro
socket7,

now that you managed to run the benches with version 2.20 it is far more easy to rank your setup: your mem-write scores seem to be abnormally low especially if we take into consideration that you run your system with overclocked fsb-speed.

I also see with astonishment that everest-numbers "obviously" don't need to be in line with good (low) super-pi numbers.

strange ...

Posted: Sun Sep 09, 2007 9:32 pm
by Jim
He did say that he doesn't use Write Allocate.

Posted: Mon Sep 10, 2007 2:41 am
by DonPedro
@jim

the more it is interesting because superpi depends heavily on memory performance and constantly read/writes to the memory to build its outcome. still super7's system manages to pass super-pi at blazzing speed.

Posted: Mon Sep 10, 2007 2:08 pm
by Stedman5040
Haven't got to sub 4m30s yet. Got the Epox MVP3G2 up and running at 617MHz (5.5x112) with 384 Mb of CL2 Hynix ram @ 2-2-2-5 with wpcredit settings. WA on and 1Mb L3 cache on

Everest 2.20 MR/MW/ML 332/173/193.5
Sandra 2004 Memory Bandwidth 226/224
Superpi 1M 4m33s

So so close. Maybe 115fsb and 632MHz will crack it if I can get there with this K6-III+/450

Stedman :)

Posted: Mon Sep 10, 2007 4:37 pm
by socket7!
I did some more benches today, so here they are (everest 2.2):

the following are with L3 cache enabled and @616MHz:

- no wpcredit tweaks, no Write Allocate:
288/87/241,9

- no wpcredit tweaks, Write Allocate enabled:
288/140/241,4

- wpcredit tweaks, Write Allocate enabled:
290/162/239,1

and now with L3 cache off and still @616MHz:

- no wpcredit tweaks, no Write Allocate:
287/87/242,7

- no wpcredit tweaks, Write Allocate enabled:
288/139/241.8

- wpcredit tweaks, Write Allocate enabled:
348/179/203,0

That score is signifficantly higher, because of offset 6C set to 08. I can't do it with L3 cache on.

and now interesting part: disabling Data Prefetch (I use CPUCool instead if K6speed) lowered latency score in everest (WA disabled):

291/127/130

and two benches done @624 (for obvious reasons with L3 off)

- wpcredit tweaks, WA enabled:
390/168/183,4

- wpcredit tweaks, WA disabled:
390/99/183/4

- and again both dataprefetch and Write Allocate are disabled:
326/99/118,3

I was surprised to see latency @118.3, so I decided to check Super Pi again. Screen is below

I don't think there is much room for further improvements here. My FIC 503A doesn't offer FSB higher then 124 so I'm stuck @624 maximum. What I'm going to look at now, is how to reduce mem latency.

Posted: Tue Sep 11, 2007 6:04 am
by Jim
It would seem then what is critical for good results in "SuperPi" is not high memory read and write speed, but low latency; which opens the question of : What is SuperPi measuring? Is whatever it is important for good overall performance? Is Memory Read and Write speed important for overall performance? Which is more important to good overall performance, (not just in "SuperPi but in general usage) : read and write, or latency?

NOTE : I Don't have SuperPi, and have never used it, so I know nothing about it. If it relates to calculating the value of Pi then I suspect it is a highly artifical benchmark for most purposes other than intermidable number crunching.