how to set the cpu to pci latency to zero?

Discussion relating to Socket 7 hardware.
DasMan2

Post by DasMan2 »

Kalabok I have 2x K6-III 450+ and 2x K6-2 450+ all ACZ , each will do the same as the others in overclocking to at least 600Mhz.

Where a motherboard will allow 605/617/620/632 speeds they all will do it.

I have posted once at 672 briefly then crashed.

I have since then acquired better memory and understanding to turn off the external cache to get higher stable top Mhz speed and will try again today to get higher speed of 617 ,.... I did @620 Mhz on the EP-MVP3C AT motherboard with external cache off (5 x 124 bus).
DasMan2

Post by DasMan2 »

Using the EP-MVP3C motherboard I have been able to run @ 133 x 4.5 = 600Mhz. SDRAM is Normal / Dram set @ 2 CAS in BIOS.
Using Everest 2.2 test program achieved scores of 272 Read / 109 write/ 267 ns delay. Full OS Win 98 S/E updates + CTU program . No WPcredit tweaks.
Hard drive is a 7200 rom 2 mb cache Maxtor. SDRAM were mixed sticks of 133 CAS 3 + Cas 2 , both still ran at CAS 2 with 133 bus setting. :)

All other attempts at the various 124/ 112/ 100 bus speeds achieved lower scores. The higher bus speed had the bigger effect on the scores than even the higher over-all Mhz speed to my results on this M.B.

Should I turn off the DATA Prefetch in the CTU program I can get a lower ns delay by quite a bit but other scores lose out 10 - 20%.

There is also something about running at least two sticks of SDRAM for this Motherboard to get stable performance vs running only with one stick. Have yet to try three sticks and get results from that.
lazy_kalabok

Post by lazy_kalabok »

lol,

unbelievable, 630 mhz with an 450acz ... so due to jims statement ive got a 450acz as well and an III+ 350acr or what? :x ö,and why amd is selling same cpus with different labels and therefore prices? ....
DasMan2

Post by DasMan2 »

These types of overclocks for the AMD K6-2 / III 450+ CPU's have been know for years. The real test for stable overclocks imho, is being able to run games without crashing for at least a couple of hours straight.

The read/ write /ns delay scores should also be related to frame increase or program running boosts etc. I would rather run the CPU @ 550 optimized than some high Mhz that achieves very little bonus performance results.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

kalabok,

here comes the comparsion table I promised some time ago giving insight what your setup's osmark scores are worth.

from left to right (with rising cpu-speed) I have included benchruns taken on various system setups I have recorded over the last 1 1/2 years.

empty cells in the header section mean "same value as to the left".

your setup's scores are expressed as a percentage number adjacent to each of my setups. red coloring is means under-par performance.

benches that rely only on raw cpu-power without any or almost none impact of memory speed are colored with darkgrey (no memory influence) or lightgrey (a little bit memory speed influence).

what deserves extra comment is that your gf4mx-4000 is a bottomless awfull performer. even the k6-2-300 setup with a gf2-mx annihilates your opengl-numbers (lorenz attractor, nbody opengl).

on the other hand your score for webpageload is extraordinairy high.

also it should be noted that except with the exception of your bench-run all other runs are performed with osmark set up at "normal priority" while your setting ("high") might raise your scores a certain amount. how much I don't know.

EDIT: one more thought: if you have time and will please run the bench again but with at least 256mb ram. win-xp (that you used) has certainly a hard time to get along with only 128mb ram. also choosing another graphic card would help in graphic intense benches a lot. and once you are there you probably could also run some other benches like me and stedman are doing as you can read in the sis530 ..... thread. I think everybody would appreciate that. stedman was able to get his sis-board to run with 133/133mhz too and it would be of great value to see 2 contenders in this class of kings! ;)
Attachments
133mhz bench cmp kalabok.png
133mhz bench cmp kalabok.png (34.42 KiB) Viewed 10310 times
Last edited by DonPedro on Sun Dec 03, 2006 6:34 pm, edited 1 time in total.
User avatar
kalabok
Senior K6'er
Posts: 147
Joined: Mon Nov 06, 2006 11:40 am
Location: germany

Post by kalabok »

wow donpedro,

im impressed. i thought i was crazy enough to build up a rusty AT computer, but looking on your efforts makes me feeling somehow normal.

first of all, some corrections - i used a k6-2+. i ran osmark with my III+ at 533 mhz, the results seem to be about 5% less. if you are interested in these, i can send them to your adress. but there are some conventions to make before running osmark - program priority, colordepth ram used etc. please tell us which buttons to push to be sure, the results are realy compairable.

my results chaded grey/leghtgrey do not realy differ from yours. there are some 3% difference, assuming that benches relying on cpu power only depend on cpu speed.
as you found out, my video performance is just crap. yes, i was wondering why the lorenz attractor bench was running almost 45 mins, getting results of two digits, weird ... but i was happy it ran fine. nearly all drawing benches are worse, so ill replace the graca in future.
ive got some 256mb ram more - i will try to run the benches you need. please tell me which programs to run under which conditions. ill try my best.

kalabok
User avatar
kalabok
Senior K6'er
Posts: 147
Joined: Mon Nov 06, 2006 11:40 am
Location: germany

Post by kalabok »

now something about getting the mvp3c2 at 133mhz:

i found out, that its impossible to get the system run at 143 or 150mhz using cpucool. the problem seems to be the frequency divider, which only allows dividing 133 by 3. even setting the winbond w83194R-58 chip to provide 133mhz system and 33 mhz pci frequency by I2C crashes the machine. it seems not to support the /4 divider.
if there is somebody who can tell me about frequency divding technique and software solutions - please let me know.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

kalabok wrote: .... looking on your efforts makes me feeling somehow normal.
:)

regarding the processor used I will correct the chart tomorrow.
of course any addtional setup's bench scores are welcome! you may post it all along here although the topic of this thread would not suggest to find that kind of intormation here. I think we should open a "133 fsb / 133 sdram: how to setup / tweak / bench results" thread to make it more visible to all who are interested. nobody would expect such information hidden in this thread.

I think I will start such a thread tomorrow where we first should manage to get straight on what benchmarks to use, how to set them up, what kind of information has to be gathered etc. for a possible first try on what benchmarks to use and what information on the system setup is to be taken note please take a look on the bench result charts I posted in the sis530 thread. the format and data positions on the system setup I used is just a suggestion and I am open to any comments, refinements, downscaling, enrichments etc.

as I 've tried to explain in my former post the bench's single tests that are shown with grey background are only cpu-power dependend (mhz) and of course this is true for any results in that row.

for now good night, here in vienna it is 3:41am and I think this is bedtime, isn't it? ;)

you will read from me tomorrow when I open the "133mhz" thread. there we will discuss how to proceed.
lazy_kalabok

Post by lazy_kalabok »

hello donpedro,

ive got an idea concerning your huge project in determing benchmarks of systems and comparing them. there is a powerful mathematic tool called factor analysis which is able to determine the number of variables the benchmars are depending on. its some game whith matrix data created by the benches we use - for ex. osmark. what we need is reliable data from one system - as you said youve got a lot of data from yours, its possible to try ...
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

mmh, need more information on the use of factor analysis here.
if we feed the factor analysis-algorithm with enough appropriate data, what will be the outcome? what will it tell us? what can we do with it?
lazy_kalabok

Post by lazy_kalabok »

donpedro,

factor analysis is able to determine the crucial setups of your system which influence the results in osmark benches. it is able to predict the number of variables responsible for system performance and even identfy them.
as i mentioned recently, you will find a linear dependence of system speed to osmark results of benches shaded grey - raw cpu power operations. the seem to depend only on one variable. others, as gridblast or mp3encode - maybe on some 2 or three as fsb, ram installed, internal cache ... even priority ... it would be interesing to prove how many and what they are.
maybe its possible to give an answer to the huge difference in bench results of video performing operations ... actually i dont know. maybe im just too optimistic concerning my poor knowledges in math. but ill try.

what i do need is data. if you could send me the excel file i will mess around a bit and tell you about the results.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

kalabok,

interesting idea to go "scientific"! ;)
I goggled a little bit and found some interesting sites explaining the idea and also offering appropriate software. but I see some problems looming: since this tool is based on statistic's laws first we need an awful lot of data to make it work, I would say a hundred bench-runs of each bench respectively if not more. the next thing would be that we would need this data for each chipset (ali, via, sis). and third how do we know how many different parameters (cpu version (with/without on-die cache, cpu speed, fsb/sdram speed, amount of ram, memory timing parameters, mainboard cache on/off, bios settings, graphic cards and their respective parameters: gpu / memory clock, memory bit-width), graphic card's driver versions etc are allowed and how the number of parameters that we know that they are influencing the bench score would demandeven more bench-runs to make the statistics work?

also what I learned is that the tool's task is to find the smallest number of parameters that are "enough" to explain the situation. if that is true then is it really of help to us?
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

@kalabok,

I just updated the chart with your correction.

if you (or whoever) wants to contribute to the bench-score collection I would suggest the follwowing.

a "complete" bench-run consists of
- superpi (256k can be left out) 1mb
- everest (version 1.52, 2.01, 2.20 give the same results)
- hot cpu tester
- osmark (just leave all settings in the program on "default": normal "priority", normal timer resolution); screen resolution should be 1024x768x16 (but I don't think that some higher setting like 1280x1024x16 or 1280x960x16 would change much since most graphic benches are fixed in size independent of screen resolution.

for better interpretation of one run it is always a good idea to make some more runs but with only one change in setup at a time. for example turning cache on/off, going from 550 to 600 by changing the multiplier from 5.5 to 6 (2), changing the fsb/ram clock from 100 to 133 and so forth. changing 2 or more parameters from one bench-run to another makes it hard if not impossible to tell what is responsible for the difference (if any) of any score between two runs.
Post Reply