The CDT and copywaza lab

[mrlobber]

For CTD and A64, try MAXMEM=104 and three files of 136MB

Now where does this come from?

[massman]

- For A64, maxmem=104 seems to be the best (not tested, but everyoneuses 104).

- Now, Kevin explains that the files are 632MB big as it's maxmem 600 and 32M run, so I reckon you should use 104 + 32 = 136MB on A64 to have the same effect.

[STEvil]

How does 600MB + 32 Million digits = 632MB? That's always confused me..

[KTE]

Here, here, let me show you why I think Super Pi is sooo inconsistent and crazy.

Finally resolved the ATi GPU BSOD issue (temporarily) so installed Crucial Ballistix DDR2-1066 which is stock bootup like this, messing about does 700 at 2.4V loose timings from 5 mins of quick testing, booted and loaded windows at 1440 with 2.4V but BSOD on opening Memset. It runs 600 4-4-4-4 PL5 very stable at 2.35V like so: http://img86.imageshack.us/img86/808/32006004444nc6.png Pity I can't run 3600-600 on this setup.



Now, Super Pi reacted VERY different today with this RAM (same everything else) than with the other on other days. It was VERY slow compared to the Corsair RAM at the same settings, same board and everything which sounds wrong but its fact. Actually, even at slower Corsair RAM settings Crucial RAM used was just not even close to Corsair RAM results. Which makes me say "WTF"

ALL settings and hardware apart from RAM were exact same, even the procedure after bootup. Have a look at this, I chose a slow one to show you how bad, compare 1M:

No tweak/No Maxmem/No LSC/Pagefile 384-384/Realtime/Core0+1/Blue theme (usually fastest for me)/white background/waste services&processes disabled/1024x768 res.

Corsair RAM -= 3605 - 563 - 5-5-5-8 PL 7 (slow subtimings): http://img218.imageshack.us/img218/5...eat3600ed2.png

13.906s

That was consistent with that RAM. Now wait till you see todays best.

CDT/Maxmem 600/LSC/Pagefile 384-384/Realtime/Core0+1/Silver theme (usually not fast but today it was fastest)/white background/waste services&processes disabled/1024x768 res.

Crucial RAM -= 3605 - 563 - 4-4-4-4 PL 6 (faster subtimings): http://img116.imageshack.us/img116/598/cdt1miq3.png

14.453s



Explain? That was repeatable by the way.

This is why I consider it a waste of time. Honestly, there's just too much inconsistency with SPi beyond user control to test it and compare, because most iof its happenings are beyond user credit or understanding. Keep all this in mind when you see my 32M times now, they will probably be very slow compared to what they should be (for a reason I don't know) but best manageable today at those settings.

===================

More tests: Stock vs CW vs CDT (32M)

Settings Used for All Main Runs: (lol)


CDT Start File Movement:


Maxmem 600/LSC/Pagefile 384-384/Realtime/Core0+1/Silver theme (usually not fast but today it was fastest)/white background/waste services&processes disabled/1024x768 res.

Stock: http://img116.imageshack.us/img116/8...413m34snq0.png
3605 - 563 - 4-4-4-4 PL6: 13m 34.782s

CW 3.8GB (good 540/550): http://img159.imageshack.us/img159/1...413m30sxw7.png
3605 - 563 - 4-4-4-4 PL6: 13m 30.234s

CDT (bad 520/555): http://img159.imageshack.us/img159/3...513m27src2.png
3600 - 562.5 - 4-5-5-15 PL6: 13m 27.875s

CDT (85% good 525/537): http://img225.imageshack.us/img225/4...413m15sug0.png
3605 - 563 - 4-4-4-4 PL6: 13m 15.668s


I don't really want to show you the one that was best.

Also, if anyones "thinking" then I've already tested and compared Core 0, Core 1 and Core0+1 results, as well as results when you set Affinity of every program to Core 0 and Super Pi to Core 1 or vice versa. There's no difference in running both cores as normal or any of these. Thanks.

[dinos22]

Superpi is very consistent if system is stable

also make sure your OS is not ed up because that's the other thing that can be messing your setup

[T_M]

pagefile is 384Mb???
Why not 512-512 as directed?

[mrlobber]

CDT (85% good 525/537): http://img225.imageshack.us/img225/4...413m15sug0.png
3605 - 563 - 4-4-4-4 PL6: 13m 15.668s

I don't really want to show you the one that was best.

Now this is something... This evening I think I'll be able to run close to your settings except at PL=7 instead of PL=6 because my Anus doesn't want to run PL6 over 420 fsb

I see in your "85% good CDT" run, you have the system cache much lower than in the "bad CDT" run. This is a very interesting behaviour I've never managed to observe because whenever I start doing the CDT copying stuff, system cache always rises close to its maximum (just like the same way if I were doing the CW stuff) and then just remains there. You have it by ~15Mb lower - could you tell please which is the moment you observe it dropping? Or, better, could you provide the available memory / system cache balance after each of the CDT copy steps? (that is, together 3x3 records?) And which is the drive you have your pagefile on? C:\ or F:\?

Btw, care to show us your best run?

Right now, with 515x7 PL7 1:1 (my Anus doesn't want to run PL6 over 420fsb with any divider ) and timings very much like yours I'm at 13m11.6s - I've managed to squeeze out a couple of seconds at these settings by improving my copywaza, but CDT still remains uncooperative with me

[mrlobber]

Ok, some more interesting stuff I just put together.

Here's a comparison of 5 runs (all done on ~3600 Mhz on P35 chipset and 450+ fsb at PL6 or PL7):
1. KTE 85% CDT:
http://img225.imageshack.us/img225/4...413m15sug0.png

2. Zeus run here:
http://ocxtreme.org/forumenus/showpo...&postcount=406 (as far as I understand from his post he does not have CDT working (at least previously he posted here he doesn't )

3. OPB's uber CDT4 run:
http://ocxtreme.org/forumenus/showpo...&postcount=360

4. My own best CW run at modest fsb and memory settings:


5. Gandalfones run here:
http://www.ocxtreme.org/forumenus/sh...&postcount=283

In Excel they look like this:


Let's discuss this one.

1. Gandalfone's, my and KTE's runs are deadly similar except of KTE's last loop (I have a theory why is this, but more testing is needed ) That's why I'm very interested in seeing his best CDT run - whether he got that last loop output "down" or not... and whether the pattern remained the same. As for Gandalfone - with the same memory timings, he's just on a way higher memory frequency as me - and this obviously pays dividends - but can we say we both used the same type of tweaks even if we have never spoken to each other?

2. Zeus and OPB's runs are even more deadly similar except that OPB in each loop is 0.7-0.8s faster than Zeus is.

3. Now if we "cut" my, Gandalfone's and KTE's graphs from point 4 and move them 2 points ahead we actually get the same Pi loop output pattern as OPB and Zeus.



Thus, the length of first slope seems to depend on memory timings, I suspect, mostly on tRCD (RAS to CAS write delay, and the pattern then should be proportional - the higher tRCD, the longer the slope - you can search for a couple of tRCD=2 runs and see it yourself ). But, if it is so, the question is why a memory parameter so heavily impacts one part of Superpi, while continuing similar loop output pattern later?

Edit: to further illustrate my point, I include one Miskos 600x6 run as well as the best 32M @ 3600 Mhz DDR2 run (only by 0.1 secs better than Gandalfones run above though ) known to me by Buttertoast (except of course OPB's uber run).
Buttertoast: http://ocxtreme.org/forumenus/showpo...&postcount=355
Misko: http://ocxtreme.org/forumenus/showpo...&postcount=445



You can see Misko is using tRCD=2 (and his pattern later anyway changes to the one of OPB), while Buttertoast obviously is using some different tweaks then everybody else pictured so far (as effective as the others though).




The most obvious thing from these graphs is another one though: CDT "effectiveness" indeed doesn't actually come in the last part or any other parts of 32M output (could this imply that CDT doesn't "optimise" anything related to HDD - assuming "spikes" happen when Superpi writes/reads temporary calculations to/from the HDD?). The CDT gain should be consistent all over the place and this makes the CDT tweak so godlike (displayed in 1M calculation as well) : where and what kind of balance in OS might give you such a consistent boost to a number-intensive calculation?

[KTE]

Very nice mrlobber. I agree so can't really say much more.

TBH I'm messing with AMD's X4 now trying to get its problems sorted because the C2D was boring me quite a bit with SPi inconsistencies. I much prefer to use a C2 lappy or a Dothan given the option.

Also one bit I left out is, I have to uninstall my VGA each time I test now which I do hate quite a bit. Blame ATi drivers for this. I also swapped my PSU because they needed the other with a Seasonic S12II 330W now (hi-pots) and its drops 0.2V on all rails under load, which is bad since the rails are very low to start off (was reading 3.1V on 3.3V rail under load recently).

Times are quicker with higher FSB for me and on ASUS boards (especially P5W DH->very quick in everything). Tried tests at 400, 450 and 470 very early with this, probably the first thing I did. With the Corsair running PL 6/7 made no difference in 32M for me. With the Crucial it makes a good difference. But I can't change PL through the OS and at bootup with the same BIOS settings, randomly, it may choose PL5 or 6 or 7 at different times.

Like I said, you can't really analyze Super Pi accurately because it is so day to day and OS dependent regardless of timings/clocks. One day you try your best and get 14M quickest at x settings and at another, with x+1 settings messing about you get 13m 57s, so you just look at it and stare at the screen knowing you did jack different to get that. Beyond user control or understanding. That leaves me very frustrated.

It's only once you can get this problem fixed that you can start analyzing a tweak, because then the OS effect each time is "controlled/fixed" and not a variable affecting the times.

This is why you see timings of CAS 5-5-5-5-10/5-5-5-5-7/4-4-4-4 PL6/7/8 get close by times. Also, in 32M memory clock seems to make a good amount of difference for me but not in 1M/2M.

I ran the best CDT straight after that day and it drops seconds even further. But I'm not testing it on DS4 anymore, moving to Abit board now. Maybe it'll hit 500 FSB there.

I have to make a general observation: the Corsair XMS2 is running quicker than this Crucial Ballistix at the same settings/timings/OS, even in EVEREST, which puzzles me greatly.

Dinos22: I find 1M very incosistent with a perfectly new OS. Depends on day to day. One day it's uber quick and the next its uber slow at same settings with no other difference from a user perspective. I also find 32M similar, differing on day to day but it's more much consistent and replicatable.

T_M: Something I've explained many times in this thread. 384-384 or 512-512 makes no difference in 4M-32M from my testing posted early so I choose 384-384 because it's better at 1M for me and so I don't have to reboot to change it time and time again.

[KTE]

Sorry left out what I change affinities with when testing. This is how:

I see in your "85% good CDT" run, you have the system cache much lower than in the "bad CDT" run. This is a very interesting behaviour I've never managed to observe because whenever I start doing the CDT copying stuff, system cache always rises close to its maximum (just like the same way if I were doing the CW stuff) and then just remains there. You have it by ~15Mb lower - could you tell please which is the moment you observe it dropping? Or, better, could you provide the available memory / system cache balance after each of the CDT copy steps? (that is, together 3x3 records?) And which is the drive you have your pagefile on? C:\ or F:\?

Cache doesn't start dropping after CDT for me (until I run an application) but memory does if I leave it too late. The trend I look for is to start SPi after it's memory has reached it's maximum but before it starts dropping. If it starts dropping or is still going up, I get slow or slower times. If it's reached the top and also balances then you have a perfect combo and the times should be quick. If it's reached the top but it's not balanced, time should be fast but not the quickest by any means. Balancing actually does make a very big difference if both these conditions are met. You can start SPi while it's going up and you'll have slow times, I've had up to 16s slower in 32M this way after CDT.

The memory/cache differs for each run with me. Sometimes it's at 500, sometimes at 510, 520, 530, ...all the way to 560 each.

When each 632MB file inside the 1.85GB RAR starts to copy over the cache falls to very low but as the copying of the file comes to an end, the cache rises very quickly to above 520MB and the memory rises slowly to around 350MB. But as soon as the cpying finishes the memory jumps to plus 500MB. Usually a ratio of the cache. For example, if my cache was 550, the memory would come very close to that, but never higher (with CDT) and if the the cache was at 520, then memory would come close to that but never higher for me. Typically, with a half decent CDT memory will start falling once you get 5-10MB near to the cache size. With a good one, it'll rise to fully balanced and stay there. This is factroring in Task Manager memory usage, because when you close Task Manager, you will gain at least 6MB-10MB of memory.

These are my experimentational findings, not the final word. That belongs to OPB.

Pagefile is on C:\.

CDT isn't working in 1M for me.

[saaya]

very interesting guys!

[massman]

Great work Mr.Lobber and KTE !!

The memory timings are indeed more important than some users tend to realize. Once again, we should ask someone to explain the working of the memory exactly to know what's really important.

[Zeus]

Nice bit of effort you've been putting in there mr Lobber, that sure made some interesting reading.

Scary how similar the curve of my run looks like OPB's superrun, iirc i was using the CDT tweak at that time, although i doubt it was done 100% right.

The lenght of the fist slope that you suspect memtimings are responsable for are interesting even though i have my reservations on that thought.

Memtimings can't be compared when there are so many variables such as memspeed and chipset imo.
It's the combination of the memorytiming at that certain speed and how the chipset reacts to that, P35 could possibly react different than X38, which makes one chipset much more senstive to a certain memory (sub)timing than the other.

I for one, found out the hard way that 400FSB on a 965 chipset is enough to get sub 13 minutes 32M at 3.6GHz as long as the memspeed is very high.

Try the same on a P35 chipset and you won't be able to do that, P35 needs more FSB to get things done.

Pretty much the same applies for memorytimings i reckon.

Still, your scientific approach with those graphs visualizes what's actually happening which is great to see.

KTE, i agree with you on the inconsistency of the superpi output results, especially in 1M runs, one day gettin sub 14 seconds at 3.6GHz is a snap, the other day i'm really struggeling to get there.

As a final word on the CDT tweak, the 1.8Gb rar file has got to be on the root of your drive, so no maps or submaps involved, this might impact the effect of this tweak.

[massman]

Oh, I forgot to post this here:

http://youtube.com/watch?v=3oFmSuswKDk

[massman]

AND this one

____________________

don't think too much though, very easy
** you NEED to do exactly like this, NOT create 3 x 632mb files and put into cdt folder,which is TOTALLY WRONG.












************************************************** *********
墨西哥硬體測試者Kain的測試結果,原文轉貼:

This was done yesterday whenI had little time to prove again CDT.

I was one of the only who tested @ 3.6ghz before (showed results in CDT theory IV thread).
Now, I just wanted to show difference between Copy Wazaa and CDT @ 4 GHZ.

Hipro5 said to some users on XS (exactly, they were i4memory users I think, which I respect a lot) to do CDT at least with some steps, going from 2 or 3 secs and then 6 or 10 over CW. Because if they just manage to give 0.03 segs more in 32M, it was not even a step.

I really cant say a lot about that since I dont want to start fight.

This is what I get with CW @ 4.056ghz with my Tweaked OS:


Now, this is what you get when do well CDT IV at the same 4.056ghz:


So, now you see, is 7 segs now a step or not?
Principle is not just copying and doing the things OPB says with his language (sorry bro, you already now that ), but understanding whats all about.

Now, if you can do that @ 4.056ghz what can you do when going up to 5 ghz or more like OPB did when he was banned and doubted about the scores?

As you see, time you gain from CW scales when you have more and more frecuency.
So, as Ive said, OPB was @ more than 5 ghz, so, now you know how is it done.

Also, dont be afraid of trying more frecuency or other CPUs. I see a lot of results @ XS for example, but all with different CPUs or not een at the same frequency.

************************************************** ***

[Misko]

Great work guys

[Zeus]

Massman, are you sure Kevin is ok with your copy/paste post?

[massman]

How does 600MB + 32 Million digits = 632MB? That's always confused me..

Actually, 32M = 33,55 Million digits = exactly 32MB.

[massman]

Massman, are you sure Kevin is ok with your copy/paste post?

Checked the entire post for warning, didn't found any, I hope he's okay

[KTE]

massman, nice gains.

If you knew me, you might get mistaken that the guy in the vid was me (system info). Did you make that vid?

[Eldonko]

You should be asking Kevin for permission before taking a post from a limited access section of ocx.

[massman]

massman, nice gains.

If you knew me, you might get mistaken that the guy in the vid was me (system info). Did you make that vid?

Nope, I didn't make it, OPB himself, I think.

You should be asking Kevin for permission before taking a post from a limited access section of ocx.

I have no acces to the limited sections of OCX, this was quoted from the public forums of Coolaler.

[mrlobber]

Everyone seems to have run away to test it all?

[kiwi]

:o new info here I will try again to compare but I see there is a need to run at higher clocks to see better boost?

CW 3.8GB (good 540/550): http://img159.imageshack.us/img159/1...413m30sxw7.png
3605 - 563 - 4-4-4-4 PL6: 13m 30.234s

What other tweaks are you running? Time seems kinda slow for your timings. See attachment (run with CW 4gb and pf 512MB)

[KTE]

What other tweaks are you running? Time seems kinda slow for your timings. See attachment (run with CW 4gb and pf 512MB)

Nothing that I've not mentioned, I mentioned it all in the post. Times are fast and slow on different days and by that I mean VERY different. Yesterday my best below 3600/562 with CDT was this at slightly slower clocks (tRFC45): http://img526.imageshack.us/my.php?image=13m13syl8.png

3595 - 562 (same everything else): 13m 13.343s

Today stock runs were on 512-512 PF and lower tRFC (40):-

3605 - 563 - 4-4-4-4 PL6 tRFC40

White theme: 13m 31.328s http://img507.imageshack.us/my.php?image=slowag2.png
Silver theme: 13m 32.672s http://img153.imageshack.us/img153/4797/slow2tx6.png

And CDT 545/545 balanced (v.good) was 13m 24s (never captured ss but trust me on that). That is very slow compared to yesterday but same setup except for PF.

Check this out for 1M.

Last week:
No tweak/No Maxmem/No LSC/Pagefile 384-384/Realtime/Core0+1/Blue theme (usually fastest for me)/white background/waste services&processes disabled/1024x768 res.

Corsair XMS2 -= 3605 - 563 - 5-5-5-8 PL 7 tRFC42 (slow subtimings): http://img218.imageshack.us/img218/5...eat3600ed2.png (this wasn't a one off them days)

13.906s

Today:
CW/Maxmem=600/LSC/Pagefile 512-512/Realtime/Core0+1/White theme (fastest)/white background (6KB)/waste services&processes disabled/1024x768 res.

Crucial Baliistix -= 3605 - 676 - 5-5-5-5 PL6 tRFC45: http://img153.imageshack.us/img153/7307/1m675cc4.png

13.953s

And this isn't just the slowest time today, this was actually the fastest at those settings today, usually it was stuck at 13.969s.

So yeah...

And here's a quick comparison of themes for you @ 3200 - 500 - 5-5-5-15 (all OS on Core 0, SPi on Core 1):

White theme: 15.765s http://img524.imageshack.us/img524/4...0555151du2.png
Blue theme: 15.765s http://img507.imageshack.us/img507/2...555151myd6.png
Olive theme: 15.781s http://img526.imageshack.us/img526/6...555151mdf3.png
Silver theme: 15.797s http://img525.imageshack.us/img525/7...0055515qt3.png

I'm going to lose me hair trying to figure this out while OPB can sit back and do this