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

Discussion relating to Socket 7 hardware.
DasMan2

Post by DasMan2 »

Oh, a couple of other things to consider. What type of ATA performance do you have with your motherboard?
ATA33 or 66 or higher?

Your hard drive you use ? ATA33 or 66/100 ..does it have onbaord cache or 2mb or more ...and RPM speed level (5,400 or 7,200)?

Do you clean out temp + cache files of your web browser each time you might do a real live internet test on performance?

I assume you have a CTU type program to optimize the CPU performance.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

@dasman

as I have explained in detail in posts above there are no net-speed-tests within the osmark bench-suite. but anyway thank you for your hint at the tcp-optimizer. regarding the disk-subsytem: the sis530 chipset provides an ultra-66 interface. I currently run the win98se installation on the soyo5ssm from a quite old harddrive, 20gig maxtor, 2mb cache, 5400rpm, 33-interface.

back to the story of exploring the soyo 5ssm:

bench-run #3: going from 60hz to 85hz
the new setup of win98se solved also another problem I had namely that I was not able to change screen frequency from 60hz. so the second column of my former bench results shows results when the onboard graphic card is run with only 60hz. so I decided to to run it this time with 85hz, not only because 60hz is a pain to look at (given that I use a good old crt-monitor, using a tft or lcd is another story), but because I it would be interesting to see what memory-penalty would follow. the phenomenon of lower memory performance because of higher refresh-rates applies because we are dealing here with an onboard graphic solution, where video ram is part of main memory. a stand alone graphic card would not show this behaviour.

the bench results show clearly how terrible such an onboard solution is. almost all benches show lower results. superpi bench 1m is hurt most: almost 100 sec slower! whereas cpumark99 is almost unimpressed (in my opinion the most imperfect benchmark ever if one thinks it shows how fast a system is).

bench-run #4: going from 16bit to 32bit

since the benches of #3 showed that using the onboard graphic card at a higher frequency in fact lowers performance I wanted to see what happens if I even change color depth to 32bit. this should really cut into performance and my assumption was proven to be right. just see what happens at col#4: have you ever seen a k6-system that produces an everest write score of 48 and latency of 721! large objects sort goes from 130 to 101, fast fourier std from 462 to 310, circles go down from 229 to 93, rectangles from 96 to 53 just to mention some. to sum it up: performance can be up to halfed when going from 16bit to 32bit color depth. to give you the possibility to see how bad the sis530 chipset performs I have included the bench results of a k6-2-300 and a k6-3-400, both on an ali-5 system. if you for now just compare the k6-2-300 results with the sis-results up to bench run #4 you will realize that the k6-2-300 on the p5a is a terrific performer, or to be precise, the ali-5 chipset's performance is just outstanding (except of course for the benches that depend solely on the raw cpu-speed. I have colored those benches that show no memory dependance with darkgrey/lightgray background.

bench-run #5: switching from onboard vga to a geforce 6200 pci add-on card.

with the new graphic card the situation gets really a lot better but still the results stay far behind what I expected. to be honest I am disappointed. memory performance on the sis530 board is just terrible even when the bottleneck onboard vga is removed. when comparing this col#5 with col#8 (k6-2-300) it becomes clear how bad the situation is. the k6-2-300 beats the k6-3-450 like I would not have imagined that could be possible (at least in some benches). also another interesting thing becomes obvious when we look at some graphic intense benches: paired with the gf-2-mx the ali-system beats ..... out of the sis/gf6200 system. this shows the importance of using a grapic card's driver that was optimized for k6-cpus. the gf6200 (300mhz core, 200mhz ddr ram) with a todays driver just can not stand up to the technically "poor" gf2-mx (183mhz core clock, 175mhz sdram). a very good indicator on how well ("efficient") the cpu-cache-ram implentation of a given chipset works is the "MetaMark to Mhz ratio" of "hot cpu test". this bench number gives a "ratio" on how many bench-points per cpu-mhz are produced in this cpu-memory intensive test. given that the k6-2 has no on-die 2nd level cache compared to the k6-3 it is just fantastic how well the ali-5 chipset transformes the cpu-power into performance! but be aware that this ratio-number does not tell whether a system is faster than another! it is an "efficiency"-indicator!

bench-run #6: introducing some tweaks via wpcredit (many thanks go to stedman for providing the pcr-file!)

up to this bench-run timing parameters were set to default, ie 3337. the bios of the soyo5ssm does not give much opportunity to tune the memory timing parameters. I was not able to lower the timing parameters through the means of the bios because the system won't even finish the first boot screen. so my hopes were that I would be in the position to alter the settings via wpcredit. but this hope was quickly set back to reality when I realized that the headroom for improvements regarding timing paramters is very small. the only memory timing parameter that did not freeze the system at the very moment I turned the corresponding bit was the "ras precharge time". so in this respect I only succeeded in going from 3337 to 3327. fortunately I found some other tweaks that brought "substantial" memory improvement. here comes my complete wpcredit test-log:

wpcredit-tweaks:
without any tweaks: 3337; 150/80/425 (everest mr/mw/latency),
onboard vga "disabled" (no driver installed)
d-link 530tx lancard, geforce 6200 driver 82.16 installed
register-offset: bit
+++ 52: bit 4 sync/async cpu-dram setting to 1 (sync)
189/93/333
+++ 52: bit 7 dram 1t setting to 0t
197/94/322
52: bit 6 setting single read allocation from 1=enable to 0=disable
no influence
52: bit 5 graphic window decoding setting to 0(=1clock)
no influence (as expected, since onboard vga is "disabled")
55: bit 3 setting res bit for write dram cycle to 0 (=disable)
no influence
+++ 55: bit 2 setting res bit for read dram cycle to 0 (=disable)
251/93/261
55: bit 0 res bit graphic window range setting to 0(=disable)
no influence (onboard disabled)
56: bit 7n6 refresh command, setting from 10=7 cycles to 11=6 cycles
no influence
56: bit 5n4 setting queue depth from 11=12 to 10=8
no influence
+++ 58: bit 7n6 setting ras precharge time from 01=3t to 00=2t
253/96/261
!!! 58: bit 5n4 setting ras to cas delay from 01=3t to 00=2t
!!! crashes and system reboots !!!
does also crash and reboot when setting 00 is used without any tweaks done before!
!!! 5D: bit 3 setting cas latency from 1=3t to 0=2t without any other tweaks done before
!!! system freezes, had to hard-reset !!!
!!! 5D: bit 4 setting dram write retire rate from 0=x-2-2-2 to 1=x-1-1-1, executed without any
tweaks done before:
!!! system freezes, had to hard-reset !!!!
5E: bit 6 setting best ras precharge time from 0=disable to 1=enable (no other tweaks)
no influence

as you can see, memory performance really goes up: mr +103, mw +16, latency -164ns
the sad thing is that the scores still are no reason to jump for joy and the memory write score is just a joke! write allocation (with no memory-hole at 15-16mb) was activated according to k6speed.

for convenience I included a "cmp" column that gives the performance gain as a percentage. it shows that tweaking with wpcredit gives some good improvements of up to 28% (sort95 in neumeier-bench4), but also the graphic intense benches profit a lot (image rotate/resize 24%).

bench-run #7: raising the fsb/sdram speed

fsb/dram speed could only be raised up to 112mhz. higher speeds did not pass the boot screen or would not even show the boot screen. the culprit could be the onboard vga that can not turned off completely (electrically via jumper). the cpu-speed is 2mhz slower (448=4x112 compared to 4.5x100), but that difference should not have an impact to consider here.
again it were the the graphic benches that profited most. some benches showed no better results although I had expected they would. I have colored their scores' background with orange. I have also included a "cmp" column to give the gain in percentage numbers.

to give you a better understanding on how "bad" the sis-chipset performs I have also included in column #9 scores for a run with a k6-3-400 paired with an ali-5 chipset. these benches were run with just some wpcredit tweaks I was aware of back then (a year or more) when I gathered the scores. so the ali-5 scores are actually lower than they could be. when you compare these results with the sis-scores be aware that the ali-system is 50mhz slower, but manages to just destroy the sis-system here and there (circles, mazes, richedit, fern fractal, maze threads, rectangles, lines, image rotate, lorenz attractor, nbody open gl). even the k6-2-300 gives the k6-3-448 on the sis-board a very hard life on some benches.

one note at the end: I think that the osmark-memlatency test is to be interpreted as higher is better. I think they made it that way (via 1/xxx conversion or alike) to make it compatible with osmarks feature of bar-graphing the results and compare them visually with other systems. the bigger the graph the better. but I don't know for sure.

one more note: I have changed the thread title to better reflect what is discussed here in this thread (it evolved from problem solving to tweaking / benching) and possible newcomers to this thread could make a better choice from start.
Attachments
Soyo5ssm bench results and comparison.png
Soyo5ssm bench results and comparison.png (46.62 KiB) Viewed 18382 times
DasMan2

Post by DasMan2 »

A lot of work to show performance of the SiS 530 chipset.

I still have this preception of the SiS 530 chipset (without onboard video) being a capable internet performer when ATA 66 is activated from my experience. This is when TCP Optimizer is used to insure top end cable connection

More feedback and some chipset adjustments are needed. :)

Sure would like to get me a PCI 6200 Nvidia unit for my AGP-less SiS530 ATA66 Mhz chipset motherboard.

Thanks DonPedro for the repsonse and posting.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

@DasMan2

I agree with you that using a machine based on a sis530 chipset for normal office use is possible and "makes sense". in fact I will give this soyo-based machine to my sister in the near future and she really will be happy with it. but once you are accustomed to how snappy an ali5-solution feels like you can't go back.

if you want to give your machine the speed-boost by using a pci-graphic card I heartily would like to advice you to look for a gf2-mx solution if you are with nvidia-based graphic cards. go for one that features
- 64mb sdram (not ddr!!!)
- because the sdram models features 128bit memory access (but not all of them! be aware!) and this proved to be the better/faster pairing then choosing ddr/64bit
- 180mhz core / 175 mhz memory clock or better. the gf2-mx cäme in various flavours with core/memory clocks low as 143/143!!!
- choose low driver version, I would suggest 12.90 or 14.10

gf2-mx cards like the one above are available on ebay from $10 to $20.

if you like ati-cards then look no further than radeon 8500, 9100, 9000 (pro), 9200, 9250.

going for the gf 6200 pci is expensive (Euro 65 and up) and brings some problems. the biggest problem is that directx8/9 games just don't run. this seems to be a driver problem that I was able to observe on ali-5, via mvp3 and also now sis530 chipset boards. also using a gf6200 with a sis530 is really not smart, because you just don't get out of the card what it otherwise could do paired with an ali5 or via solution.
Last edited by DonPedro on Mon Nov 27, 2006 7:46 pm, edited 1 time in total.
User avatar
Uranium235
Senior K6'er
Posts: 183
Joined: Mon Sep 12, 2005 9:59 pm

Post by Uranium235 »

DonPedro, that was truly a very detailed analysis of the SiS530 performance on the SY-5SSM. :) I have also reached some of the same conclusions based on my own testing with the Asus P5S-VM and Gigabyte GA-5SMM.

1) The performance of these boards is dismal at the 100FSB. I believe a major contributor is the small 512k L2 cache which will only cache 64MB of RAM.

2) The onboard video performance is lame and a resource hog. ANY PCI video card would be better.

3) The inability to fully disable the onboard video, by jumper, prevents the system from being fully stable at the 133FSB. The P5S-VM does have a jumper to do this and is the only SiS530 board that I know of that has a disable-video jumper.
When I tested the GA-5SMM at the 133FSB, the onboard video displayed many artifacts until the system crashed. It may be possible that the AGP bus on the SiS530 chipset does not have a 2X divider even though the PCI bus has a 4X divider.
One extreme way to disable the video may be to find and sever the wire trace to the chipset that controls the onboard video.

4) To achieve any real level of performance, the SiS530 needs to be run at the 133FSB with the L2 cache disabled and the onboard video disabled.
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

@ Peter : What kind of SDRAM were you using for those tests? Good quality RAM does make a difference when you are trying to tweak memory timings.
Superpuppy 3
K6-3+ 450 ACZ (6x100)
DFI K6BV3+/66 Rev B2 (2 Meg) w/ 2x28mm Chipset Fans
2x256 Meg PC 133 Hynix SDRAM
1x 20G Maxtor (7200)
2x 80G Maxtor (7200) Ducted w/ 2x486 Fans Mount
52/24/52/16 LG CDR/RW/DVD
8/4/3/12/24/16/32 LG Super Multi
ATI 9000 aiw Radeon AGP
SB Audigy 1 MP3 Sound
CMD 649 IDE Controller
NEC USB 2 Card
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

@uranium235

could you do us a favour and send your asus p5s-vm through the benches as above? I would be glad to work your results into the data table .... naturally this would involve "some work" ;) because what is the interesting thing to compare here is that your board seems to be able to run at a) 232 memory timings b) can really disable the onboard vga and can therefore c) be run at 133/133mhz. so just 3 little bench-suite runs ;) would be nice: 100/100, multiplier 4.5x (to make your numbers directly comparable with the soyo's results), mem params 232, onboard enabled. then the same but with some pci-card. and of course what everybody would like to see, the crown of all settings: 133/133, 4.5x, 232, pci-video card.

@jim
the memory used was a high quality brand, infineon, specs 133mhz @ 222. the module works just perfect in any other board.
User avatar
Uranium235
Senior K6'er
Posts: 183
Joined: Mon Sep 12, 2005 9:59 pm

Post by Uranium235 »

DonPedro, I'll be glad to run those benchmarks on the P5S-VM. I'll need some time to set it up because I removed the P5S-VM from my "test rig" for some experiments I've been doing on an ALi 7, Powercolor P561A, which has been a disappointment....:(
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

AH, you got avail of one of those ali-7 boards!?
tell us more about it! features, 133/133 available? and why not also run the benches on this board? it does not matter if the scores are good or bad, but it is interesting and revealing anyway, because I have not heard or read of anybody here in this forum to have run any benches on it .....

I also would like to mention one more time, that running the "official" osmark via its respective button is a pain when run with the onboard vga because lorenzattractor and nbodyopengl will take forever to be run. just untick these two and also the 3 benches that are in the very last line (those that are a little "offset" to the right bottom of the bench-box). you can save hours and we will not loose any important information. then set the run counter to 3 and let it go. also "priority" of osmark should be set to default = normal (right top).

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

Post by Stedman5040 »

Peter,

Just in quick reply to your request here are some benchmarks for my tweaked Gigabyte GA-5SSM

K6-2+/550 @ 100 x 5.5
3x 128 MB pc100 SDRAM Hyundai (2-3-2-6)
PCI Matrox Mystique 4MB
PCI VooDoo2 12MB
Compac NIC 3120
Matrox 10Gig ATA66 HDD
Compac CDROM

Quite a basic system but here are the results

Everest

MR_______________________________295
MW______________________________122
ML_______________________________229

Superpi

256k____________________________72
1M______________________________368

Hotcpu tester pro lite 4.2.2

Total score______________________945
Metamark to MHz ratio____________1.72

Integer instuctions_______________175
Floating point instructions__________68
MMX instructions_________________372
3D NOW instuctions______________337

Memory fillrate memcpy method___151
Memory fillrate C++ method_____147
Memory rep mousd______________151
Memory FPU 64-bit______________149
Memory Assembly 32-bit_________145
Memory fillrate assembly 64-bit___147
Memory fillrate MMX 32-bit_______150
Memory fillrate MMX 128-bit______149

Insertion sort__________________7
Shell sort_____________________239
Merge sort____________________456
Heap sort_____________________212
Quick sort____________________998
Large objects sort (quick method)_188

Fast fourier transform (standard)__378
Fast fourier transform (optimised)_401

getting a little bit faster for SIS530. Maybe in a few weeks we might see some more improvements :wink: I must retry the 133 fsb on this board. There is also the option of 124fsb with the 4x divider. Although this slows down the PCI bus it might be worth seeing the results if it is stable.

Didn't have time to run OSMark tonight so I'll try it another night.

Stedman
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

thank you stedman,
once you got the osmark scores I will work your data into the comparison table for better viewing.
User avatar
Uranium235
Senior K6'er
Posts: 183
Joined: Mon Sep 12, 2005 9:59 pm

Post by Uranium235 »

DonPedro, here is a link that will give you a little information about the Powercolor P561A. I haven't been able to translate the language but you can understand the specs and the benchmarks in the review.

http://www.SPAM/testy/ ... _ArtX.html

The board does have a 133FSB but I can't get it to work. The onboard 3D looks good but is slow. I think I'll start a new thread for this when I'm ready.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

yes, the aladdin 7 should support 133 fsb and 133 sdram speed.

but this is not the only good news!

I just searched through the archive of a german pc-magazin and was double lucky, because they tested an aladdin 7 board AND report some very interesting stuff!

the really good thing about the aladdin 7 is, that by doing some tricks it succeeds in featuring 128bit memory access!!!

this can only be achieved when 2 ram-sticks indentical in size are used!

they run their memory test program, comparing 2 different ram settings on the aladdin 7 board and also tested the memory speed of an ali-5 board.

just by using the 128-bit mode gives a tremendous boost in memory performance :)

I have compiled the data for your convenience. don't give up that board easily, if you succeed in getting the 133 sdram speed working this board may prove to be a superperformer!

EDIT: I found someone who also wants to try the 133mhz thing with this chipset. in his case it is the pcchips m587lmr board and he is asking for help here. his post is a year old but maybe he found out something useful.
Attachments
aladdin7 comparison.png
aladdin7 comparison.png (3.17 KiB) Viewed 18246 times
User avatar
Stedman5040
Veteran K6'er
Posts: 271
Joined: Mon Sep 05, 2005 4:22 pm

Post by Stedman5040 »

Peter,

Here is OSMark results for the GA-5SMM with the cpu @5.5x100 and the rest as stated in previous reply above. External cache is on. I think some of the tests take a hit because of the video card combo. Do you want me to carry out any more tests with this combination?

Test, Score, Time(s)
==== ===== =======
Grid Blast, 243, 7.680
Grid Blast: Floating Point, 238, 9.078
Fibonacci threads test, 283, 5.861
n-Body, 171, 56.077
Pi threads test, 161, 43.202
Trig Curves, 437, 5.437
Trig Curves2, 260, 0.626
Plot Lines, 474, 17.319
Random Dots, 459, 33.537
Mandelbrot threads test, 228, 21.324
Filled Circles, 233, 9.251
Maze, 240, 31.080
Fern Fractal, 221, 10.304
RichEdit, 39, 138.302
Draw Ellipses, 142, 40.262
PiTest, 314, 23.199
Dhrystone threads test, 277, 19.636
Whetstone threads test, 153, 23.850
Maze threads test, 426, 42.069
Draw Rectangles, 30, 235.275
BandwidthBP64 threads test, 100, 21.911
Memory latency threads test, 220, 33.372
Orthogonal threads test, 171, 12.540
Identical threads test, 139, 6.572
Draw Lines Test, 185, 32.124
JPG Decode, 7392, 4.127
Image Resize, 195, 15.748
Image Rotate, 214, 29.117
MP3 Encoder, 25, 200.695
Web Page Load, 140, 79.105
Zip File Compress, 127, 118.877
AES Encrypt/Decrypt Thread Test, 4, 297.609
File Copy Test, 137, 80.240
Lorenz Attractor, 22, 2,454.830
N-Body OpenGL, 6, 4,024.060
PngOut Thread Test, 74, 718.066
7zip Thread Test, 69, 333.477
Upx Thread Test, 61, 65.480

*** Test Ends ***

It seemed to take forever. Anyway I will see if the machine will get up to 133fsb.

Stedman.
DonPedro
K6'er Elite
Posts: 578
Joined: Wed Jul 27, 2005 2:11 pm

Post by DonPedro »

thank you stedman for your contribution! :)

I entered your data in my table. before I post that table again I need some information: what was the operating system, what driver version of the matrox card, and at what screen res / color depth were you running the tests.

in case that you are using win98 I have a tool for matrox cards that set up the lfb addresses. I don't know if this applies only to agp cards but maybe it helps with pci cards too. just let me know and I will send it to you via email.

what I found interesting while looking at your data: there is something wrong with the score that you got for the JPEGdecode-bench. it is abnormally high, about 80 to 90 times higher than expected. please run this test again (3 times) and tell me what you get.

also some numbers are astonishingly high or low but not in that magnitude (comparison number is taken from k6-3-450):
AESencrypt-decrypt (4 instead of 7 or higher)
mp3encode (25 instead of around 50 or higher)
pngout, 7zip and upx "should" also be twice as high if not higher
both fastfouriertransform benches are way lower as expected given that all other benches of hot cpu show that the memory performance of the ga-5smm is way better than that of the soyo-board.

the matrox card has its limits (nbodyopengl and lorenzattractor), I would not expect it that low (that is the reason why your osmark-bench run took so long ...) but on the other hand on some benches it gives the gf6200 a hard time.

for the moment the only problem is that we can't really compare our systems since the cpus differ in 100mhz. I will run the benches with an k6-2+-550 on the soyo board again. then we get closer to see what difference is left that can be explained because of better memory performance and or graphic card's capabilities.

if you don't mind you might also run the suite again with 4.5x100 and 4x112. just untick the opengl-benches, these are not that of importance in our chipset investigation, and your bench time will be m u c h shorter.
Post Reply