SiS530 @133MHz (GA-5SMM/Soyo 5SSM) setup, benching, tweaking

Discussion relating to Socket 7 hardware.
Post Reply
User avatar
KachiWachi
K6'er Elite
Posts: 507
Joined: Wed Sep 21, 2005 10:53 am
Location: Pennsylvania, USA

Post by KachiWachi » Sat Dec 02, 2006 6:47 am

@KGB -

You do know that Everest 3.5.x now has a 30-day timeout...right?
Moderator - Wim's BIOS

PC #1 - DFI 586IPVG, K6-2/+ 450 (Cyrix MII 433), 128 MB EDO. BIOS patched by Jan Steunebrink.
PC #2 - Amptron PM-7900 (M520), i200 non-MMX, 128 MB EDO
PC #3 - HP8766C, PIII-667, 768 MB SDRAM
PC #4 - ASUS P3V4X, PIII-733, 256 MB SDRAM
PC #5 - Gateway 700X, P4-2.0 GHz, 768 MB PC800 RDRAM
PC #6 - COMPAQ Evo N1020v laptop, P4-2.4 GHz, 1 GB PC2700 DDR
PC #7 - Dell Dimension 4600i, P4-2.8 GHz, 512 MB PC2700 DDR
PC #8 - Acer EeePC netbook, Atom N270 @ 1.60 GHz, 1 GB RAM
PC #9 - ??? ;)

User avatar
KGB
K6'er
Posts: 98
Joined: Mon Apr 07, 2003 7:55 pm

Post by KGB » Sat Dec 02, 2006 11:01 pm

@Kachi Wachi.

I didn't realize after purchasing and registering the Everest 3.5 program, I would have to remove the program a mere 30 days later. I mean, if that were true. One would ask, just what sort of scam is Lavasys is running? :wink:
-K6-2 550 Gigagyte GA-5AX (5.2) 1x512MB GeForce 2 Ti500 64MB
-Pentium III 933 Asus P3C-L 2x128MB GeForce 4 Ti4200 128MB DDR
-Pentium III 750 Asus P3B-F ATi Rage128 GL 32MB
-Celeron 1.2 Asus TUSI-M 2x128MB
-Pentium IV 1.5 DELL 2350 1x128MB
-Athlon 900 Asus K7M 1x256MB 2x128MB GeForce 3 64MB
-AthlonXP-M 2400+ Asus K7V880 2x512MB HIS Radeon HD3850 512MB DDR

User avatar
Stedman5040
Veteran K6'er
Posts: 271
Joined: Mon Sep 05, 2005 4:22 pm

Post by Stedman5040 » Sun Dec 03, 2006 1:49 am

Peter,

here are the OSmark figures for the GA-5SMM at 4.5x133 (600) using the same screen res and color depth as before. The memory is set at 128Mb (2-3-2-6) and I think some of the scores reflect this when comparing the previous set at 100fsb and 384Mb of memory.

Test, Score, Time(s)
==== ===== =======
Grid Blast, 255, 7.311
Grid Blast: Floating Point, 268, 8.058
Fibonacci threads test, 305, 5.442
n-Body, 185, 51.693
Pi threads test, 173, 40.093
Trig Curves, 453, 5.251
Trig Curves2, 275, 0.592
Plot Lines, 503, 16.324
Random Dots, 487, 31.551
Mandelbrot threads test, 248, 19.541
Filled Circles, 252, 8.531
Maze, 258, 28.862
Fern Fractal, 226, 10.082
RichEdit, 41, 131.309
Draw Ellipses, 156, 36.727
PiTest, 337, 21.649
Dhrystone threads test, 300, 18.135
Whetstone threads test, 167, 21.861
Maze threads test, 457, 39.251
Draw Rectangles, 34, 204.131
BandwidthBP64 threads test, 89, 24.664
Memory latency threads test, 173, 42.313
Orthogonal threads test, 114, 18.767
Identical threads test, 92, 9.856
Draw Lines Test, 85, 69.721
JPG Decode, 5140, 5.936
Image Resize, 130, 23.687
Image Rotate, 171, 36.423
MP3 Encoder, 65, 76.678
Web Page Load, 132, 84.180
Zip File Compress, 132, 113.874
AES Encrypt/Decrypt Thread Test, 9, 116.418
File Copy Test, 60, 182.614
PngOut Thread Test, 132, 403.479
7zip Thread Test, 183, 124.673
Upx Thread Test, 127, 31.426

*** Test Ends ***

Stedman

User avatar
KachiWachi
K6'er Elite
Posts: 507
Joined: Wed Sep 21, 2005 10:53 am
Location: Pennsylvania, USA

Post by KachiWachi » Sun Dec 03, 2006 8:13 am

@KGB -

If you buy a license, this doesn't apply.

I was talking about using it as "trial-ware".

Sorry about that. :oops:
Moderator - Wim's BIOS

PC #1 - DFI 586IPVG, K6-2/+ 450 (Cyrix MII 433), 128 MB EDO. BIOS patched by Jan Steunebrink.
PC #2 - Amptron PM-7900 (M520), i200 non-MMX, 128 MB EDO
PC #3 - HP8766C, PIII-667, 768 MB SDRAM
PC #4 - ASUS P3V4X, PIII-733, 256 MB SDRAM
PC #5 - Gateway 700X, P4-2.0 GHz, 768 MB PC800 RDRAM
PC #6 - COMPAQ Evo N1020v laptop, P4-2.4 GHz, 1 GB PC2700 DDR
PC #7 - Dell Dimension 4600i, P4-2.8 GHz, 512 MB PC2700 DDR
PC #8 - Acer EeePC netbook, Atom N270 @ 1.60 GHz, 1 GB RAM
PC #9 - ??? ;)

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Sun Dec 03, 2006 5:58 pm

@stedman

I just finished putting the numbers together (up to the 550mhz runs). I will post the chart on monday with some textual interpretation.

thank you for the numbers at 600/133. what I need to complete the table for the runs at 600/133, 128mb, etc are the numbers for superpi. many thanks in advance!!!

I had a short look on the 600/133 numbers and yes, it seems to me that going down from 3x128 to 1x128 costs a lot of performance in the osmark-benches. sometimes it goes down steeply but other benches show big gains.

since the 600/133 runs needed a lot of change to your setup (+50mhz cpu inncrease, lowering amount of ram, setting cache to off, raising fsb/sdram speed to 133) we are in a little trouble to attribute the difference in score numbers to what made the difference.

so please let me ask you when you have finished the superpi-runs (for the 600/133, cache off, 128mb - run - suite) what we need is a complete run (superpi, everest, hotcpu, osmark) with the settings 6x100 (fsb, ram) with 128mb and another complete run with 3x128. this would make it easier to spot what is responsible for what difference in the scores.

of course if you can also make the k6-3+ work with 600/133 this would also allow us to see what the +128kb on-die cache are worth. but I suggest to apply the baby-step approach.

btw, since I plugged in my gf6200 pci card I no longer can monitor temperatues on my system (the latest bios does not show the numbers anymore and the software within windows does only display voltages, fan speed but not temperatures). does it work on yours and if so how hot gets the cpu @ 600?

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Mon Dec 04, 2006 4:33 pm

here comes the new chart, featuring some more interesting investigations.

first I wanted to know the influence of amount of on-die cache. I did this by running the benches with a K6-3-450 and a K6-2+450. then I proceeded with what influence turning on/off mainboard cache has. this I did by using a K6-2+550. to get an idea on how well/badly soyos' implentation of the sis530 chipset works I have also included the scores for a k6-2+550, mbcache activated, on a gigabyte board in direct comparison. again many many thanks to stedman5040 for his contribution. also you will find a run with a k6-2+600 on the soyo board and to give better insight on what the numbers are worth I have again included the bench scores that turned out on an asus p5a with a k6-2-300 and k6-3-400 respectively.

influence of amount of on-die cache, going from 256 to 128 kb cpu cache, #6 and #11
superpi 1m shows a significant deterioration when going from 256 to 128 kb cpu-cache. everest in turn is not impressed at all. cpumark again shows that it is not designed to reflect performance but is just an indicator of how much cache is there (-41%, a number that is only reflected once in the osmark suite). hot cpu test repeats above results: synthetic benchmarks stay almost the same while application-like benches do indeed show the impact of smaller cache size. osmark shows the k6-3 leading in almost all benches. the most cpu-2nd-level cache hungry applications are mp3encode with a 10% difference, webpageload 12%, zipcompress 30%, pngout 40%, 7zip 15% in favour of the k6-3. some graphic intense benches also show a 5 to 10 % deterioration when a k6-2+ is used.


influence of turning mainboard cache on / off, #13 and #12
superpi and everest just don't care about additional 512kb mb-cache. cpumark likes mainboard cache and gives +1.3 points (+4.5%). hot cpu tester is not very much impressed, shows only a significant better situation in the fast fourier transformation benches (+6 to +8%). osmark is also quite unimpressed, only ellipses (10%), zipcompress (10%) give remarkable better results with mainboard cache turned on. overall it pays to turn on mbcache.

comparing soyo's with gigabyte's implementation of sis530 chipset, #10 and #13
in superpi and everest the gigabyte 5smm solution shows its muscles. how big the influence of setting cas-latency from 3 to 2 is I don't know but it surely has its merits. superpi scores are 6 to 7% better, everest shows whopping +46mb/s (+18%) and +25mb/s (+26%) in memory read and write respectively. hot cpu tester confirms this (+25mb/s in the synthetic memory benches). the sorting routines on the gigabyte board run 10 to 20% faster as well as the fast fourier transformations. osmark results show once again that if your system is not well balanced you loose comparatively and sometimes even wreck havoc. in our case here it once again is the graphic card used that has its difficulties in some benches (gridblastfp -9%, nbody -5%, circles -31%, richedit -26%, ellipses -57%, rectangles -83%, lorenzattractor -91%, nbodyopengl -93%). interestingly in some benches it can even shine: plottrig +18%, plotlines +20%, randomdots +18%, mazes +15%, lines +17%, image resize +160% !!!!!, image rotate +18%. but on the other hand there are graphic card independent performance problems that are really strange because it lacks straight logic: upx -46%, 7zip -53%, pngout -23%, encryptdecrypt -56%, mp3encode -53%. very good numbers are retrieved in ithreads +12%, weppageload +16%, and bandwidth +18%. conclusion: given that the comparison is between two systems of same chipset, almost identical memory tweaks (except cas-latency), same cpu type and -speed, same mbcache setting the results are very surprising - in both directions - and I have no explanation for it (except for the graphic intense benches).


some other observations:
hot cpu test: although the scores for integer instructions, mmx instructions and 3dnow!instructions should be exakt the same for the same cpu clocked at the same speed, stedmans numbers on the gigabyte board (#10) are a tick higher compared with #13 scores. (???)
the k6-2-300's numbers (#8) of hot cpu test's synthetic memory benches are better than all other numbers across the soyo-board columns. this without on-die cache and with no wpcredit tweaks at all on the k6-2 side! also the k6-2-300's everest mem-latency score is the lowest I have encountered EVER. this cpu paired with an ali-board also performes very well when compard with the soyo-scores. the pairing of the k6-2-300 with a gf-2-mx card also produces astonishingly high numbers in osmark's fern fractal, richedit, mazethreads, lines, webpageload, lorenzattractor, nbodyopengl, pngout and 7zip. for the graphic intense benches this is insofar interesting as it shows how much one can be mislead by simple name dropping: nobody would expect that the lowly system by name (k6-2-300, gf-2-mx) would blow away such systems like k6-3-450, k6-2+-450/550/600 each paired with a gf6200. this shows ultimately how important it is to use the "right" chipset and a graphic driver that is/was written for K6-cpus.

regarding the comparison mb-cache on (#12) / off (#13) I must say that I did not expect the setting with mbcache activated to be the winner, although it hapens by just a small margin (by looking at the osmarks scores, which interest me most). so far my experience with ali-boards showed the opposite to be true. this possibly can be attributed to the soyo's very bad memory write scores (see everest) that are too much of a bottleneck and therefore distort the results in favour of the mbcache=on bench-run.

final remarks:
empty cells in the header section of the table means value is equal to the value of the cell to the left.
benches with lightgrey/darkgrey background are highly memory performance INdependent, relying almost alone on the cpu's speed (mhz).
for easy reading I have colored the backgrounds of the worst (red) and best (green) results in this table for each of osmarks benches.
Attachments
sis530 256vs128cache mbache-onoff 550+cmp.png
sis530 256vs128cache mbache-onoff 550+cmp.png (40.74 KiB) Viewed 3882 times

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

Post by kalabok » Wed Dec 13, 2006 10:29 am

hello to you guys,

first of all - thx to don pedro, his investigations are usefull to understand the complex interactions between the parts on a mobo.
i have an interesting remark to add here concerning the k6-2 300 mem latency. after i red your this observation, i decided to try out a pentium mmx on my 133mhz fsb machine. it booted winxp at 266 mhz, a cpu with 16kb instruction and same data L1 cache. the external cache was turned off.

running everest gave numbers, i can not explain:

read 443 Mb/s
write 103 Mb/s (the only "normal" value)
delay 122,4 ns

running cpumark and hotcpu do give normal values, they are testing the cpu power, not the memory. sandra memory benches - about 280.

it seems to me, that the l2 cache onchip (!!!) is slowing down the memory in a crucial way. even with onboard l2 enabled and 115mhz fsb i get 380 read and 145ns delay ... suggestions?

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Wed Dec 13, 2006 6:21 pm

kalabok,

these numbers are VERY interesting! fourhundredfortythree memory read!!!??

just don't stop here, let that sys-setup run the whole bench - suite, please! of course a 266mhz cpu will not reach high scores where cpu-power is a must, but scores will be very valuable for their own sake! who had ever thought of such good memory-read scores on a socket7 system?

also it is wrong that hot cpu tester bench suite does not include memory-bneches, au contraire! more than half of its single benches are memory related. I WANT TO SEE THESE NUMBERS ! :)

and run superpi 1m too.

also run osmark, we will get a very deep insight what such a good memory performance is worth. I am quite sure this setup will beat some soyo-runs lefthanded. go for it! (and if your budget allows try to buy some gf2-mx pci-card on ebay so that the bottleneck of your system is removed ....).

edit: and while you are at it: please run the whole bench-suite with onboard cache enabled too! this will lower synthetic memory bench scores, but will boost system performance (real world applications) a lot. see my table of benchscores I have posted in the "10x K6-2+550 for EURO 99.-" thread and you will understand what I am talking about.

Super7Dude
K6'er
Posts: 77
Joined: Mon Aug 28, 2006 8:05 pm
Location: Brisbane, Queensland, Australia

GA-5SMM at 133Mhz

Post by Super7Dude » Sat Jan 13, 2007 5:29 am

Hey guys,
I have an AMD K6-III+450ACZ at 4.5x133 stable so far. I've tried orthos beta, Everest RAM/CPU benchmarks and memtest with two 256MB PC133 CL2 RAM sticks... current timings are 2-2-3-7 @ 133MHz. It's passed 6 hours of memtest (2 complete passes) and Orthos has been running about 3 hours now in blend mode with no problem so far. Overall the system is quite solid, the only crash I've seen is with 3DMark2001, which runs fine however at 124MHz. Everest RAM bandwidth is 141MB/s with 133MHz CL2-2-3-7. The chip is handling the speed fine at 1.95v core. Will post more benches soon :)

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Sat Jan 13, 2007 11:11 am

thank you for spending your time! :)
could you run the same benches (same version) as seen some posts above in the table?

Super7Dude
K6'er
Posts: 77
Joined: Mon Aug 28, 2006 8:05 pm
Location: Brisbane, Queensland, Australia

GA-5SMM at 133MHz

Post by Super7Dude » Sat Jan 13, 2007 11:29 am

No worries m8! I'm trying to find the benchmark programs in that table. Would you be able to email me the versions used? Would be much appreciated :) Cheers

It's still going btw, Orthos blend 6 hours and counting.... onboard vga, no artifacts at all :o

:edit: thats 9 hours for Orthos, I've taken a screenshot as well :lol:

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Sat Jan 13, 2007 11:45 am

s7dude,

for the osmark go here

for "hot cpu tester pro" go here

for everest 2.20 go here

for superpi go here

Super7Dude
K6'er
Posts: 77
Joined: Mon Aug 28, 2006 8:05 pm
Location: Brisbane, Queensland, Australia

Post by Super7Dude » Sat Jan 13, 2007 12:11 pm

Sorry guys, I didn't realise my BIOS had reset to 3-3-3-8 timings for the RAM (as my screenshot and BIOS just showed). However, it's still stable AFAICS. Will run OSMark and Hotcpu in a bit :wink:

DonPedro
K6'er Elite
Posts: 573
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro » Sat Jan 13, 2007 12:17 pm

s7dude,

did you realize that in the previous pages of that thread we also found some ways to tweak memory performance with wpcredit?

also be aware that running the osmark (official run or 3 runs manually as described in the osmark-thread) will take approximately 1.5 hours. with onboard vga this will take double time if note more!!!

Super7Dude
K6'er
Posts: 77
Joined: Mon Aug 28, 2006 8:05 pm
Location: Brisbane, Queensland, Australia

Post by Super7Dude » Sat Jan 13, 2007 7:42 pm

Yep, OSMark is running just fine atm, about 3 hours in, no failures/crashes yet. I'm patient, will report results after it's done :wink: I'll have a look at mem tweaks soon... I've heard quite a lot about WPCREDIT but am a bit worried that it might wreck my board permanently... is this possible? or will a reset make all the settings go back?

Post Reply