Page 12 of 16

Posted: Sun Feb 04, 2007 4:43 pm
by Jim
Statement 1, is false. The with cache runs all wound up faster on the second run than the no cache runs at the same speed.

Statement 2 is true; but the Sandra runs only show a consistant pattern for a limited number of times. How many times that is depends on how much RAM is uncached at 3rd level. Stedman did not have a large amount of uncached RAM at 3rd level; that is why there were no gains in the Sandra runs.

Statement 3 is true. So do most other progams.

Statement 4 is false. The best runs were achieved with the 5.5 multiplier. I would suggest that the anomalous 397 run was probably due to the manner in which the tests were run. To confirm or debunk that supposition, I would suggest rerunning the tests in reverse sequence. I.e. Instead of running the 5.5x100 first and the 6x100 last, do them in the opposite order.

Statement 5 leads me to wonder why. A long time ago I reported a gain in Everest Memory Write speed from 177 to 201, Or 13.6% as a result of running Everest after Sandra had run several times. I also reported rising Sandra Memory Bandwidth scores on the order of 14.76% for INT MMX and 12.86% for Float FPU.

Statement 6 : Answer : Only one possible thing; - increased concentration of cache hits!! Apparently at both 2nd and 3rd level.

Posted: Sun Feb 04, 2007 6:28 pm
by DonPedro
Jim

the "mystery" is a two-fold:
first that sandra gives higher results with consecutive runs.
second, "mystery" about why benches run after sandra runs show better results.

the data stedman provides states clearly that the mystery is not cache-related, can not be explained by the amount or even absence of mainboard cache.
a) he expresses in clear words that he experienced no higher results with consecutive runs of sandra. quote: "The Sandra scores for all of the above fell by about one or two points during the 3x runs rather than increase"
b) whether cache was enabled or not, superpi showed the effect of the mystery (gain in speed when run after sandra).

so your statement, that my statement 1 (the mystery is cache-independent) is false, is blatantly false. the explanation you give is misplaced, because it just says that superpi likes cache. which is my observation #3. at least you admit this one is right, thank you! :)

my observation #4 states in clear and visible letters that "SANDRA does not like cache when 5.5 multiplier is used". I suggest you look again at the SANDRA numbers at 5.5x multiplier with cache and compare it to the respective ones without cache. all 4 four numbers w/o cache show clearly and unambigous to anybody who has eyes to read that they are higher than their respective counterparts. so your statement, that my statement 4 is false, is blatantly false.

re my statement #6 "what is it that makes it happen?"
"it" is the mystery as explained above in my first paragraph. according to what stedmans data shows your answer "increased concentration of cache hits!!" is just otherwordly. there can't be any concentration of cache hits when cache is disabled (the mystery appears also when cache is disabled), the sandra scores did not increase whether cache was enabled or not (the alleged mystery does not dare to show up at all) and how superpi is able to profit because of "increased concentration of cache hits!!" when it is run without consecutive repetition is a miracle.

I have noted it in my last post but I gladly repeat it again. would you mind to follow my invitation to open a new thread on this issue? I will follow you over there, if you don't mind. don't you think it would work better this way, somehow off-topic here, isn't it? has nothing to do with tweaking the sis530 and running it at 133mhz ...

thank you! :)

Posted: Mon Feb 05, 2007 6:32 am
by KachiWachi
Has anyone here reported this to the test program developers...to get their input?

After all...they write the test...maybe they can shed some light on what is going on.

Don't forget to invite them to review the data you present here.

Thanks.

Linux on Gigabyte 5smm

Posted: Tue Feb 06, 2007 2:12 am
by larrystotler
Greetings. I have been looking over some posts on this forum over the past week or so and finally decided to join when I saw that these are relatively new posts.

I have a Compaq Gigabyte GA-5SMM motherboard like you guys have been discussing. It's a v1.2. Thanx to info I got here, I was able to successfully update the BIOS to the Gigabyte version. Much better configuration. I have a K6-3/400, 2.4V installed.

Currently, I have the system setup with the following:

K6-3/400 running at 467Mhz using 3.5x133fsb, 2.5V
128MB PC133 SDRAM
ATI Rage 128/16Meg PCI card, onboard disabled
Intel e100 PCI Ethernet card
20GB Maxtor
DVD-ROM

I don't use Windows, so some of your testing methodology and tweaking is lost on me. I currently have openSUSE v10.2 installed as well as MPlayer/Mencoder. I have the following tests done if anyone is interested. This is a re-encoding, first pass using mencoder, of an episode of Transformers Energon. Command line setting was:

mencoder test.copy -alang en -oac copy -ovc xvid -xvidencopts pass=1:turbo:cartoon -ofps 120000/1001 -o test.v1

-alang en Output english
-oac copy Copy Audio stream
-ovc xvid Encode using XviD codec
-xvidencopts pass=1:turbo:cartoon First pass, turbo(slightly faster), optimize for cartoons
-ofps 120000/1001 use 60 frames per second

Results

k6-3/400 100fsb, 4x100 6.35fps baseline
k6-3/400 66fsb 6x66 4.96fps 22% slower than 4x100
k6-3/373 124fsb 3x124 6.63fps 4.4% faster than 4x100

The first set of tests had the motherboard's L2 cache turned on. When the FSB was increased to 133, the system would lock-up during boot with the motherboard cache enabled. The kernel would halt the system.

k6-3/400 100fsb(no l3), 4x100 6.19fps 3.5% slower
k6-3/400 133fsb(no L3) 3x133 6.97 9.76% faster 4x100
k6-3/466 133fsb 3.5x133 7.70 21.25% faster than 4x100

I've also re-encoded the Shrek2 movie from the DVD first pass(2nd pass is currently running). First pass gave me 10.60fps, which is really good considering the speed of the system. I usually get around 11fps on my Dual Xeon Dell(500Mhz, 100Mhz fsb) for movies. Re-encoding of the audio track to mp3 using mp3lame codec is usually around 130fps.

I haven't really tweaked my memory timings yet. I did have to use a better heat sink and fan to run stable at 467. I'd be interested in going higher, but I doubt this processor could handle 4x133 @533Mhz. A K6-3+ probably could, but I only have regular K6-2s.

The system is installed with no X server and as few processes as possible. I chose mencoder because to me it's a real world app unlike a lot of synthetic test results. Using mencoder is very processor intensive, and the faster fsb helped out a great deal. Increasing the fsb from 66Mhz to 133Mhz showed a 40% increase in speed using the same internal processing speed, even tho the L3 was turned off on the 133fsb test.

Posted: Tue Feb 06, 2007 1:19 pm
by larrystotler
2nd pass mencoder of Shrek2 gave me: 5.01fps.
3rd pass mencoder(mp3lame audio) gave me: 131.71fps

Posted: Tue Feb 06, 2007 2:55 pm
by lazy_kalabo
@larrystolter and others

if you are looking for a faster X+ cpu, here a site with some offers:

http://www.futurecb.com/Boutique/

keep looking for keywords!

Posted: Tue Feb 06, 2007 3:52 pm
by larrystotler
Wow! They've got K6-2+ processors, 500, 550, & 570Mhz for $7.44 each! Or $6 each for quantities of 2 or more! They are listing 100+ of the 500s and 10,000+ of the 550 and 570s. Check this page:

http://www.futurecb.com/Boutique/Search ... RO+DEVICES

That's CHEAP! much better than eBay and these are probably BRAND NEW!

I may have to look into this.

Posted: Tue Feb 06, 2007 4:04 pm
by larrystotler
Whoops. Not 550, 533.

Here's the list:

500ACR
533ACZ
570ACZ

So, which would be the best overclocker? Too bad they aren't K6-3+'s, but K6-2+'s are good.

Posted: Tue Feb 06, 2007 6:27 pm
by Jim
The best overclocker in that list would be the K6-2+ 570 ACZ.

Posted: Tue Feb 06, 2007 6:55 pm
by larrystotler
Jim wrote:The best overclocker in that list would be the K6-2+ 570 ACZ.
What kind of speed would be practical? 600? 650? 133x4.5? 133x5?

Posted: Tue Feb 06, 2007 7:36 pm
by Jim
Not sure cause I don't have a K6-2+ 570 ACZ, but the ACZ has the highest voltage and temperature rating of any K6 +, and the 570s have a rep for being able to go higher than the 550s. An article that I read at Ace's Hardware which was about K6-3+550ACRs said "All samples tested were able to clock 660MHz". How high you can go w/ the K6-2+570ACZ is probably more dependant on your mobo than it is on the processor. Considerations to take into account are 1) The Higher the FSB the better performance you will get. 2) At any given FSB, the 5.5 times mutiplier will give better performance than any of the others. 3) Be careful, or else you will overreach and blow something up; (See posts by "Saintly Cobra"). It is best to approach it conservatively, go for gains incrementally, bit by bit checking operating temperatures as you go. If things get too hot, either back off or improve your cooling. 4) The highest speed I have ever heard of reached by a K6-anything, in stable operation, (watercooled), was 744MHz. 5) I have heard of people getting 690 MHz aircooled, but rest assured they had better than stock cooling arrangements not just for the processor; but also for some mobo components.

Posted: Tue Feb 06, 2007 7:49 pm
by DonPedro
welcome larrystotler! :)

nice to read from new faces and from someone who engages his k6-system with linux.

maybe I misunderstood what you tried to say, but don't you think it is the ram which is clocked at 133 speed that makes the sis-system fly? or is mencoder really raw cpu-power dependent and the data it works on fits perfectly in 2nd level cache?

the k6-2+570 will make 600 (4.5x133) without any problem. if you would like to go higher than the next setting would 5x133=666 which would need some adjustment to the hardware I guess. I imagine that it could be possible if you decap the cpu and use a very good heatsink and fan.

I will finally go for the 666 by myself. I can't try it for now because my soyo-board does not make 133mhz, won't even boot. so I have to wait until an asus p5s-vm I recently ordered will arrive.

Posted: Tue Feb 06, 2007 8:11 pm
by Jim
If I were trying to get max out of a K6-2+570ACZ, I think what I would shoot for is 5.5x124 = 682MHz. But I'll say again, approach it slowly and carefully.

Posted: Tue Feb 06, 2007 9:03 pm
by larrystotler
DonPedro wrote:nice to read from new faces and from someone who engages his k6-system with linux.
I've been using Linux full-time since 1999. I use W2k sometimes depending on the need for a windows app or for my son's games. I also use 98SE on some of my really old laptops.
DonPedro wrote:maybe I misunderstood what you tried to say, but don't you think it is the ram which is clocked at 133 speed that makes the sis-system fly? or is mencoder really raw cpu-power dependent and the data it works on fits perfectly in 2nd level cache?
Mencoder seems to be more fsb speed dependent than processor speed when comparing similar speed procs. I had a Celeron Mobile system that used SDRAM. When I moved it to DDR, it almost doubled in speed overall. The RAM speed is important, but only if the processor can actually make use of it. It irritates me to no end to see systems that have a 400 or 533Mhz FSB require PC 2700 DDR. There's no reason for it and almost no benefit from it. I still haven't figured out exactly how much cache is neccessary when using it. Using the Celeron Mobile(P4 based) which has 256k L2 isn't really any slower than the P4's 512k cache. I did some tests with a K6-2 vs a K6-3 at the same speed, and the K6-3 was at least 20% faster. But I'm what I'm unsure of is what the size of the data set is and when more L2 becomes irrelavent.
DonPedro wrote:the k6-2+570 will make 600 (4.5x133) without any problem. if you would like to go higher than the next setting would 5x133=666 which would need some adjustment to the hardware I guess. I imagine that it could be possible if you decap the cpu and use a very good heatsink and fan.
I will finally go for the 666 by myself. I can't try it for now because my soyo-board does not make 133mhz, won't even boot. so I have to wait until an asus p5s-vm I recently ordered will arrive.
I have a couple of Super 7 boards, but I think that my Gigabyte is the only one that will do 133. I have a PCCHips M599LMR board that looks a lot like the Gigabyte, but it has onboard ethernet and stuff. I don't think it has the 133fsb. I'll have to fire it up and check. I also have:

AOpen AX59pro ATX
Tyan S1590S AT
Elitegroup P5SD-B AT(doesn't work)
Shuttle Hot 591P AT(doesn't work)
SPAX ATX(only does 83Mhz fsb)

I probably have more if I keep looking.

Posted: Wed Feb 07, 2007 8:07 am
by DonPedro
@larrystotler

it is the chipset that provides the 133mhz setting. afaik the m599lmr uses the sis530, so the only thing you got to know is how to set the jumpers for selecting the various fsb/sdram speeds (100, 112, 124, 133). whether the pcchips board can stand it is a question of manufacturing quality.

in most cases you won't succeed with onboard-vga enabled. only super7dude who has a ga-5smm was able to run the board at 133mhz with onboard-vga enabled. all others who tried it failed.

you wrote: "Using the Celeron Mobile(P4 based) which has 256k L2 isn't really any slower than the P4's 512k cache." if these cpus have the same core and only differ in cache-size than I would say that this hints towards mencoder's data to be worked with fits almost into 256k cache.

so far our bench adventure has revealed that programs which rely only on raw cpu-power show no improvement when one switches from 6x100 to 4.5x133. cpu and ondie-cache are both run at 600mhz. the results are equal. just look at the many charts we have produced so far earlier in this thread. the benches that have grey background will show this effect.