Rising Results in Sandra's Memory Bandwith benchmark

Discuss software and how to tweak more performance out of your system.
User avatar
Stedman5040
Veteran K6'er
Posts: 271
Joined: Mon Sep 05, 2005 4:22 pm

Post by Stedman5040 »

I didn't feel any need to complain, because no slight was intended. I was merely reporting results found and would seriously like to find out why this sequence of events occurs. We do need some more data to work this out to which I shall endeavour to do asap.

besides all of this we have learnt quite a lot about optimising the SIS chipset and I hope more will be found to optimise it further.

After all already we have gone from a chipset that is just a laugh to something that actually works.

I'm all for open debate as it does seem quite refreshing.

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

Post by DonPedro »

@jim and others who are interested

re cacheable area of k6-3: 4GB?

I went through the pages you have recommended about how cacheing works and also collected information on the k6-3 cache capabilities from various sources on the internet.

I then put all these information together and applied it to accordingly re 4gb of ram, K6-3 256kb cache.

since the information and calculations pcguide gives and carries out are only done for 1-way-associative caches I could have made a mistake, but according to the spreadsheet calculations I could have done it right. I checked it by the two numbers (grey background) that must be equal if the calculations are done correctly, but to get sure I contacted the author of pcguide-webpages and asked for help ....
Attachments
k6-3 cache decomposition.png
k6-3 cache decomposition.png (16.78 KiB) Viewed 22094 times
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

Peter, you have the most amazing way of twisting peoples words. Re machine code, I may not know as much about computers as you do; but I did spend a year at Humber College studying programming. That was a long time ago, but I learned to write software in PL1, Cobol, and Assembler. I know the difference between Machine Code, (000 & 111) and High Level Languages. I also know that when a program is compiled it gets translated into machine code. I also know about loops. ("Do while" in PL1.) All I was pointing out is that what seems rather simple becomes rather complex when you write software. There are a heck of a lot of loops, and loops within loops within loops written into programs; and loops are repetitive operations which are conducive to cache optimization.

Regards your program A and Program B; you seem to miss the point. The OS controls the operation of the machine, not the programs the user gives the OS to run. On SP2 I had 128 Meg of 768 Meg of RAM cached at 3rd level. That meant I had 640 Meg UNCACHED at 3rd level. Windopes loads first into UNCACHED RAM. That meant I had enough uncached RAM to swallow WIN98-SE whole. i.e. The whole of windows was loaded into RAM that was Uncached at 3rd Level That meant that all of those parts of the OS required to run Sandra, Everest, or Super Pi, could only be cached in 2nd level cache. Running Sandra several times in succession, apparently results in either ALL or at least a high percentage of of those parts of the OS required to run Sandra winding up cached in 2nd level cache. THAT benefits any program which requires those same OS components in order to run. Because it is OS components we are talking about, It is the OS which decides if the cache memory ought be reset. And if the OS finds the components it requires to run program B already in cache, as a result of running program A, it would be a very poor OS that chose to overwrite them with garbage from Program B, when it has already reserved 3rd level cached RAM for user programs, ( in order to speed their operation), by loading itself into uncached RAM. I am beginning to think you are deliberately spouting nonsense in order to goad people into running tests for you so that you can collect the data for some purpose of your own.

Re "Cheat Bench" that is not the first time you have made derogatory comments about this particular method of obtaining results. Where followed by a "smile" it means you are kidding; but you have not always done that.

Re you cache chart, it looks accurate to me. I could be wrong however because I have not checked the math. Bottom line is as soon as possible I am going to fire up SP2 w/ or w/out a new mobo, (I have 4 new ones, [1x Rev 103, 2x Rev 104, & 1x Rev 105]; but the old one still works w/ 16 colours), and I am going to run some tests w/ it using Norton's "System Doctor" cache hit tracker. It will mess up the results to some degree, (hopefully not so much that they lose their characteristics), so will running my "Corel Capture" software to take pics of the cache hit tracker results; but I KNOW you are wrong.

EDIT Reviewing old posts has only confirmed my beliefs.

I Posted these results :
Having installed three 256 Meg sticks of Hynix HYM71V32635HCT8P-K, (PC133 CL2), in Superpuppy 2, I decided to see if I could get 112 MHz FSB stable.

Superpuppy 2 : ASUS P5A-B Rev 104 w/ K6-3+ 450 ACZ @ 5.5 x 112 MHz @ 2.3v. w/ mobo cache enabled. (First attempt; - has gone well, will try to reduce to 2.2v. later).

Everest :
Memory Read = 361 MB/s
Memory Write = 177 MB/s
Memory Latency = 182.8 ns

Sandra :
Performance Rating (Estimated) = 740
Arithmetic = 1521 D ALU MIPS ; 731 W FPU MFLOPS.
Multimedia = 3544 INT MMX ; 4217 FLOAT 3D NOW!
Memory Bandwidth : (Run repeatedly on account of rising results)
INT MMX : 210 ; 213 ; 229 ; 231 ; 238 ; 241 ; 240.
FLOAT FPU : 210 ; 216 ; 228 ; 234 ; 237 ; 237 ; 237.
Cache & Memory Bandwidth : 1164 - 2k ; 1179 - 4k ; 1183 - 8k ; 1164 - 16k ; 1120 - 32k ; 1036 - 64k ; 915 - 128k ; 630 - 256k ; 268 - 512k ; 236 - 1Meg ; 236 - 4Meg ; 235 - 16Meg ; 237 - 64 Meg ; 236 - 256Meg.

Rerunning Everest AFTER Sandra gave the following results :
Memory Read = 365 MB/s
Memory Write = 201 MB/s
Memory Latency = 184.3 ns

All tests run under WIN98-SE ; w/ Power Tweak 2's tweaks ; w/ : CAS Latency = 2T ; RAS to CAS Delay = 2T ; RAS Precharge = 2T ; RAS Active Time = 6T ; Row Cycle Time = 9T. - & my usual mess of 10 taskbar items running.

Said results were obtained w/ these tweaks :

Settings, = Customized Powertweak 2.

Powertweak 2 Processor Settings :
Write Allocation = Enabled.
Write Combining For VGA Framebuffer = Disabled.
Write Combining For VGA Banked Memory = Disabled.
Data Prefetch = Enabled.
Weak Write Ordering = Disabled.

Powertweak 2 ALi Chipset Settings :
Use M1541 internal TAG RAM = Disabled.
Use M1541 internal MESI = Enabled.
Cyrix Burst Mode & K6 Write Allocation = Enabled.
Cyrix Linear Burst Order = Disabled.
L2 TAG Output Delay = Enabled.
CPU Single Read Cycle L2 cache Allocation = Enabled.
Dynamic Write Back = Enabled.
Pipeline Function Control = Enabled
Write Combining For VGA Fixed Framebuffer = Disabled.
Write Combining For Linear Framebuffer = Disabled.
SDRAM CAS Latency = 2 Clocks.
SDRAM RASJ Precharge Time = 2 Clocks.
SDRAM RASJ Cycle Time = 9 Clocks.
A.G.P. Side Band Addressing = Enabled.
A.G.P. Transfer Rate = 2x.

Stedman posted the following results on an MSI (also ALi Chipset).

MS5169 Ver4.0 (ALI AladdinV)
K6-III+/450ACZ @ 550 (5.5x100)
3x256Mb high density PC133Ram (records 768Mb)
Intel i740 AGP
Creative SB128
Fujitsu 6.0Gig HDD
24x cdrom
Win ME

The reason I chose to run these benchmarks on this board was because it would support the high density SDRAM so I could really beef up the memory as per comments by JIM. I run ctu at start up to enable write allocate etc and without any further tweaks from wpcredit get the following from Everest and sandra2004 (Same as Jim).

Everest read/write/latency
333/156/200
Sandra memory bandwidth
190/193

Aladdin 5 chipsets do seem to give reasonable scores from the off.

By wpcredit tweaking the dram to give timings of 2-2-2-5 with everything else as default

everest scores give 334/163/200
Sandra gives 196/198

By wpcredit tweaking mixed dram command interval to 3T/5T we get

Everest

333/176/199

Sandra2004

198/196 rising to 208/206 after repeat tests

A further wpcredit tweak enabling fast dram read gives

Everest 351/183/191
Sandra 212/210


Stedman also posted the following results :

K6-III+ @ 550 (5.5x100)
MS5169 Ver4.0 (Ali chipset) 512k L3 Cache
1x 256mb SDRAM High density PC133
Intel i740 graphics
Creative SB128
24x CDROM

Run1.

Default settings as booted up no tweaks other than usual CTU start up. Memory timings at 2-2-4-5

1. Everest MR/MW/ML 331/153/200
2. Sandra start of run 185/177
3. sandra end of run 178/176
4. Everest MR/MW/ML 328/145/202

Run2.

Memory timings changed with wpcredit to 2-2-2-5

1. Everest MR/MW/ML 333/161/199
2. Sandra start of run 183/179
3. Sandra end of run 180/179
4. Everest MR/MW/ML 329/142/198

Run3.

As above but with mixed dram set to [1] with wpcredit

1. Everest MR/MW/ML 331/161/199
2. Sandra Start of run 189/183
3. Sandra end of run 182/180
4. Everest MR/MW/ML 329/141/198

Run4.

As above but also with Fast Dram read enabled with wpcredit

1. Everest MR/MW/ML 351/168/188
2. Sandra start of run 190/185
3. Sandra end of run 188/186
4. Everest MR/MW/ML 345/150/189

Please note that with all the above tests the Sandra readings go down with successive tests and end up with poorer Everest readings the second time around. Just look at that drop in Memory write (more than 10%)


Set up 2 as above but with

3x 256mb SDRAM high density PC133


Run1.

As run 1 above.

1.Everest MR/MW/ML 332/155/200
2.Sandra Start of run 192/183
3. Sandra end of run 204/200
4.Everest MR/MW/ML 332/176/204

Run2.

As run4 above ie all tweaks

1. Everest MR/MW/ML 352/170/188
2. sandra start of run 195/191
3. Sandra end of run 212/208
4. Everest MR/MW/ML 352/182/192

Clearly more memory changes the trends in this testing. In this case sandra bandwidth results go up and the subsequent Everest results also go up

So as Jim has noted on many occasions before lots of ram improves Everest MW results after repeated testing with Sandra memory bandwidth benchmarks. Has anyone seen the opposite effect as seen above with lower amounts of memory??

Stedman also posted the following results :

Here are some figures using the Ver1.04 P5A board. The set up is as follows

K6-2+/550 @ 617 (5.5x112)
512Mb pc133 cl3 Hynix ram (2x256mb)
Asus P5A (ver1.04) (cache on)

WD204 20GB 5400rpm HDD
Voodoo3500 AGP graphics
Hercules muse lt sound card
Windows 98se

Everest results from straight boot up with no memory tweaks

MR 380
MW 190
ML 173

2004 Sandra memory bandwidth

ALU 217 rising to 233
FPU 214 rising to 229

Beats the Aladdin7 reported results on Sandra by a full 10% or more.

Everest results after Sandra shows no gains

MR 380
MW 192
ML 170

Not bad results and pretty cool to see the K6-2+/550 at 617MHz . This seems a bit odd to me however as Asus report in the MB fsb settings frequencies of 105, 110, 115 etc and not 112. I'm getting the 112 MHz from the 110MHz setting. It is also nice to see that the onboard cache is running at 112MHz.

CONCLUSIONS :

1) Cache ON works better than cache OFF.

2) Having a large amount of uncached RAM improves the performance of the machine when all else is equal.

3) Rising Sandra Memory Bandwidth results are associated w/ large amounts of uncached RAM. Falling Sandra Memory bandwidth results are associated w/ little or no uncached RAM.

4) Higher FSB speed combined w/ the 5.5 multiplier allowed the K6-2+ to outperform the K6-3+ in Stedman's tests. It also out performed my K6-3+ @ the same multiplier setting and FSB setting, which may indicate that 512Meg is the optimum amount of RAM to install on an ALi board; BUT the Everest result did not rise after first running Sandra. My K6-3+ showed higher final Everest results; and higher final Sandra results. (Tweaking aside, - my machine was bios tweaked and also Powertweak tweaked though not w/ WPCredit. Stedmans was apparently Untweaked!!!!!)

5) Everest resuts did not show a gain after running Sandra numerous times w/ the K6-2+ because : That processor does not have enough 2nd level cache to allow the OS to cache all of the OS data points required to run Everest, and therefore the Cache contents did not optimize.
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
lazy_kalabok

Post by lazy_kalabok »

sorry, jim,

but i have to make a statement how i do see those numbers concerning your statements:

1) there is no evidence for better cache ON performence since no cache OFF measurements were made (in the project above)

2) hmm ... hard to agree for me. ok, let us define system performance as the numbers u get in everest. so stedmans k6-3 benches @550 give 331/153/200 for 1X256mb ram and 332/155/200 for 3x256mb ram for run1 - quite similar numbers. as well as for run4 - have a look on it. ! its unimportant what sandra makes with those numbers afterwards! only talking about system performance, there is no gain for me - sandra is no tweak! it does make something with the data stored in ram/l3/l2/l1 - but it does not making your system faster as ram timing, bank interleave or system fsb. so ram amount does no change ...

3) here the numbers obviosly claim you are right - as far as i understand what u mean

stedmans numbers are amazing, thats true. even outperforming the aladdin 7 - with no tweaks ... omg. why the results are not rising after sandra - maybe its because the smaler l2 cache. maybe because of smaler ram amount (512mb) ... maybe because of that mild evening at the end of march ...
User avatar
Stedman5040
Veteran K6'er
Posts: 271
Joined: Mon Sep 05, 2005 4:22 pm

Post by Stedman5040 »

Ok folks,

Got the Asus P5A v1.04 up and running with a K6-3+/500 cpu. The set up is as follows and is pretty much as before.

K6-3+/500ANZ
Asus P5A v1.04
Voodoo3 3000 agp graphics
3x256mb Hynix pc133 cl2 ram
Hercules muse xl sound
Realtek nic
Maxtor 20gb HDD 5,400rpm
CDROM drive

Kicked the machine off at 550Mhz (5.5x100) with the onboard 512k cache on.

Everest with no tweaks gives 329/161/215
Superpi 256k gave a result of 61s followed by 1m which gave 313s ?

I then ran Sandra2004 memory benchmark, which gave

190/188

After repeated runs this only rose to a maximum of 196/192. a miserly 2% difference.

Running Superpi 1M after the Sandra runs gave 310s. Clearly nothing like the improvement found with the SIS530 chipset.

However, I powered down and restarted and ran Superpi 1M before any thing else and then got a higher figure of 346s. Clearly something is a little odd here. I ran it again and got 323s. I need to investigate this a little further.

Anyway I then powered up the cpu to 617mhz (5.5x112) with onboard cache enabled and ran superpi256k and got a creditable 54s and followed that on with Superpi 1M which came in at 273s !! Again I tried Sandra to enhance the Everest scores which started out at

Everest 371/181/192 (not 190 MW as seeen before)

Sandra started at 216/212 and rose to a maximum of 226/217. again not a big gain. Running Superpi afterwards did not give any improvements to the scores.

Clearly the ALI chipset and Superpi with Sandra is not acting completely in the same way as the SIS530 chipset.

Stedman
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

Tony: Some time ago I got a fair bit better results w/ Everest after Sandra than I did before. This weekend I had SP2 back up w/ a new board. (i.e.different P5A-B Rev 104 bought used), and did some test runs to see about these rising results. The machine is not really running properly, (the old video card is now in SP5; and the Radeon 8500LE replacement is not being found so far), and I am having trouble running my registry checking software to fix things up. Not sure if the issues I am having are hardware related, (Mobo), or software related, (file corruption).

Using an old ATI PCI video card w/ windows supplied drivers, w/out the Soundblaster Live 5.1 card installed, (though the old drivers were), I ran some tests @ 5.5x112 using Everest and Sandra. Though there were rising results, they did not show as dramatic an improvement, either in the rising Sandra results, or in the Everest results after as those I was getting before. I am not sure what to attribute the difference to. Does a PCI Video card put more of a load on system resources than an AGP card? Was it my screwed up software install? Is my new board NFG? I don't know yet; but I will find out. (The board I used is the worse looking of the two used replacement boards I bought, - did not want to wreck the better one by running it at 5.5x112 w/out chip cooling.)

BUT, I did observe a rising % level of cache hits w/ each addidtional Sandra Run.
1st Everest Run
Start Cache Hits = 13%
Finish Cache Hits = 18 %
1st Sandra Run
Start Cache Hits = 21 %
Finish Cache Hits = 28 %
2nd Sandra Run
Start Cache Hits = 28 %
Finish Cache Hits = 32%
3rd Sandra Run
Start Cache Hits = 34 %
Finish Cache Hits = 37 %
4th Sandra Run
Start Cache Hits = 37 %
Run Failed - Invalid page fault.
5th Sandra Run
Start Cache Hits = 42 %
Finish Cache Hits = 47 %
6th Sandra Run
Start Cache Hits = 49 %
Run failed - Invalid page fault
7th Sandra Run
Start Cache Hits = 53 %
Run failed - Invalid page fault. Note Cache Hits @ 55% at time of failure.

The initial Everest scores were
Memory Read = 381, --- previous best was 361 for before Sandra.
Memory Write = 186, --- previous best was 177 for before Sandra
Memory Latency = 187.3 --- best ever was 182.8 before Sandra

The Sandra Scores were also out of character.
1st Run Int MMX = 223 --- previous best for 1st run Sandra was 210
1st Run Float FPU = 220 --- previous best for 1st run Sandra was 210
2nd Run Int MMX = 224 --- previous best for 2nd run Sandra was 213
2nd Run Float FPU = 220 --- previous best for 2nd run Sandra was 216
3rd Run Int MMX = 226 --- previous best for 3rd run Sandra was 229
3rd Run Float FPU = 223 --- previous best for 3rd run Sandra was 228
4thRun No Results
5th Run Int MMX = 230 --- previous best for 5th run Sandra was 238
5th Run Float FPU = 226 --- previous best for 5th run Sandra was 237

Final Everest results were not taken in the cache hit tracking set of runs, but the overall results were not dissimilar to another set taken earlier the same day, w/ out Cache Hit Tracking, wherein the final Everest Results were LOWER than the initial ones.
So either the changes in hardware or the screwed up software install are affecting the results. Initially higher, (Much higher in Everest results), but then LOWER both in Sandra and Everest as the test set goes on. (In spite of measured HIGHER Cache Hits). This is gonna take more work to figure out than I thought, - mainly because it is going to take a lot of work to restore SP2 to being fully operational.
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
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

@ Kalabok : Regards point 2). You are quite right in stating that Sandra is not a tweak tool. It is merely a benchmark tool; but it is a benchmark tool that is capable of reporting, (in an indirect manner), a rising incidence of cache hits. Now when you think about it, any program that is composed of highly repetitive sequences of events, is going to generate more cache hits than one that is not. If you want a real world example; they have supercomputers at work calculating the value of Pi. That task is about as repetitive in its operational steps as you can get; and the cache hits generated doing it are probably near or at 100%. So what the rising Sandra results show, (and the same is true of the Everest results after Sandra), is what is possible given software which generates a lot of cache hits; - and that is all it shows.
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
lazy_kalabok

Post by lazy_kalabok »

jim,

after testing some k6-2+ and III+ on the aladdin7 (no l3 cache) im beginning to believe, sandras effect only exists if there is any l3 cache. enabled or even disabled ... with every amount of ram installed i get same sandra numbers - after varios benching.
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

That is right. Reason is L2 cache is optimized to the maximum degree possible just by loading the program where no L3 cache exists. It is only where L3 Cache exists that some things may be overwritten in L2 cache and recached in L3 thereby allowing somethin g else space in L2 that we see gains.
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
User avatar
KachiWachi
K6'er Elite
Posts: 507
Joined: Wed Sep 21, 2005 10:53 am
Location: Pennsylvania, USA

Post by KachiWachi »

What happens when you turn Write Allocation off?
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 - ??? ;)
lazy_kalabok

Post by lazy_kalabok »

one paradoxon - stedmans benches show clearly better benchmarks through sandra even if the l3 is disabled ...
... ill try out the k6-2 with only l1. lets see, what happens.
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

Answer to both of you: I don't know! I am not a computer expert. Peter knows more about them than I do. I am just someone w/ a better than average reasoning capacity; and that was the conclusion I arrived at. The cache hit tracker shows I was right, for whatever reason. If you want to engage in testing to determine exactly what is causing this effect WELCOME aboard!! It would be helpfull however to have a board that has L3 cache, and is capable of supporting a lot of uncached RAM to test w/.

EDIT: @ Kalabok, have not yet checked which of Stedman's results you are refering to. But is it possible that they were obtained w/ a K6-3+ w/ a larger 2nd Level cache? In that case it may be possible that some unnecessary OS components cached @ 2nd Level may get progressively overwritten by the benchmark program components and thus again show better results; though not as much better as you would get w/ 3rd Level cache present and enabled.
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
lazy_kalabok

Post by lazy_kalabok »

yes, im refering to stedmans results in the sis thread. actually the everest numbers are not rising ... but the pi times do decrease. hmm ...
the sis530 chipset is able to cache 64mb with 512k l3 onboard. lets forget about the caching ability of the k6-III. stedman installed three times more ram than the system is able to manage - and gets faster after sandra.

he wrote on 3. feb:

"... My Sandra scores were about 214/211 and they did not move at all during the four membench runs. Still it is very interesting that the score should be so significantly better after the sandra runs. ... . I didn't see a change in my Everest scores either. It is all a mystery to me but it got the SIS530 below 300 seconds which must be a minor miracle."

ok, so far the sis530 chipset.

for my mvp3 system i do get better everest and unchanged pi scores, with cache enabled or disabled - it does not matter. note that everest is slowing down when varios programs are running - its a plus of about 5% closing sandra. my system had 50% of ram uncached (256mb with 512kb cache). quite similar to jims superpuppy 3 analysis.

for the ali5 chipset stedman and you provided a lot of data. sandra, everest and pi are rising for the p5a and the ms5169 has a complete other behaviour. have a look on your chipset revision - on the asus site there are 4 ali V revisions, each with different cacheable area for the l2/l3.
and there are still cache off measurements missing. maybe this will be a hint for a solution of this miracle.

ok, finally the ali7 - results are like rocks. not impressed by sandra at all - mem bandwidth, everest and pi stay the same. but they differ from cpu to cpu. ill post some results next week - getting mr results above 550 for the cyrix, lol. but thats another question.

see ya
Jim
K6'er Elite
Posts: 1745
Joined: Wed Jan 21, 2004 7:10 pm
Location: Toronto

Post by Jim »

K, something new to report re "Cache Hits". Was doing some work on a photo w/ Corel "Photo Paint" a few days ago and had Norton Utilities 2000's "System Doctor" applet running in the background. Was surprized to see cache hits running at 70% for an extended period of time.
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
Post Reply