• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Strange P3-S Tullie memory benchmark :(

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

Falkentyne

Member
Joined
Apr 9, 2001
Ok this is my P3 1.4-S @ 1575 (FSB 150) on my BH6 1.1

CPU scores look just fine.
4342 mips and 2116 mflops,, just around an AXP 1800 (slightly lower mflops but slightly higher mips).

But memory benchmark is a different story.
(anyone know why I can't post screenshot attachments??)

RAM Int buffered iSSE bandwidth: 1013 MB/s
RAM Float buffered iSSE bandwidth 635 MB/s <------???

now, in the lower field, it says the int efficiency is 84%(estimated),
while the Float efficiency is 53% (estimated).

What does this mean? What is 'estimated'? Is that the true performance? Why is the float score so low?

I remember getting a not quite so similar (actually, it was 490 for float, and about 581 for INT, FSB was also at 150) float score on my 600e @ 900, also FSB 150.
However I never paid attention to those scores back then
The mobo is an Abit BH6 1.1 (BX)

Is this supposed to be normal?
 
What is it that exactly causes this phenomena? Everyone says that BX-chipset has better memory bandwith than for example i815 or VIA694-chipset. But anyway it scores lower in f.e. Sandra... Or am I mixing things up totally?? :rolleyes:
 
well the Via Apollo Pro 133 has in comparison to the BX Chipset bad Memory Bandwith.

But, the Via Apollo Pro 133A is nearly as fast as the BX, in SiSoftSandra they got same Scores, but anyhow in 3D Marks 2001 BX-Boards got about 5-10 % higher Scores ??

For example my TualCeleron 1,1 @ 1,54 ( FSB 140 ) with ATI Radeon 8500 ( 275/275 ) get 7600 Marks when on an BX Board someone else got 8200 Points with same setup...

Brosken how is your Score on 3D marks ?
 
If I open up my case and run at 1.65v & 143fsb and speed up my Ti4200 a bit more I get about 9800 3D marks (2001). Tried and tried to get it above 10k but no luck... =) At this current setup I get maybe 1000 marks lower...

The reason I wonder is just that in Sandra2002 my memory bench (current setup) is about 950MB/s for both Int&float. But when I run these same parts in my old BH6 it showed about the same int-speed but just above half of the float-speed...
 
Hopefully this will work.....

sanb1.jpg


sanb2.jpg
 
Those numbers look dismal too me.

I just ran the bench on a 815ep board with a 1.26 P3-S at
default speed, and scored 950/940. 89/88% efficiency.
mem timings at 2-2-2 7/9 at the moment.
(Got it at default speed at the moment as I am about to
power down and toss my new 1.2 tB1 in and void the warrenty)

At 150-155 it gets 1100 or so on both tests with about the
same efficiency numbers... 90-91% if I am running at 2-2-2 5/7.
 
Well, the numbers aren't the problem; the efficiency is.
The question is *WHY* is it so low?

I'll send a support email to sandra and ask them about this.

And notice, if you read above, others have had similar bad "float" scores, or low efficiency on BX boards, with Tullies, too.
 
Those memory scores are what you would have expected to see in the old Coppermine days when the processors didn't have data prefetch, due to the fact that the Tualatin cores have data prefetch as part of their design you 'always' find the ALU and FPU scores are almost spot on with each other.

Basically, it looks like data prefetch isn't working on that particular combo.
 
Hmm? On my coppermine, my INT benchmark with an older version of Sandra was at most about 100 points higher than my float. I very faintly remember getting about 580 on the INT test, and 490-500 on the float test.

So that doesn't seem like it corresponds with your conclusion.
Now, with the newer 2002 version, I'm getting 1013-1080 for INT and 600-700 for float. (Sandra Professional)

Well, I just tested something else just now.

I downclocked my CPU to 952 mhz (90 FSB, just where the PCI 1/3 divider kicks in, since that's somewhat close to 900 mhz on my older CPU, except that was 150 FSB). I don't know what difference the FSB is making, however, my results were pretty strange:

RAM INT: 652 mb/s
RAM Float: 377 mb/s

INT scores seem a bit higher than on my 600e @ 900 (652 vs 580) but float is much lower (377 vs 490). But again, the sandra versions were 1-2 years apart.

I then clocked back to spec (133 mhz FSB / 1400 mhz), and got 919 mb/s int, 557 mb/s float. INT efficiency was only 86%; maybe I have too many other things loaded. Float was 52% (it's never above 54%).

I'm going to test this on my roommate's abit BF6....
How is data prefetch supposed to influence the memory scores?

Ok, I came back from testing.

Well, my roommate's Abit BF6, on his 500E @ 733 (fsb 133), he gets 994 mb/s in INT, and 553 mb/s in float ! His INT efficiency is higher than mine, about 94 %, but his float efficiency is the exact same, STILL MUCH LOWER THAN HIS INT !, about 53 % ! His float score per FSB seems to match mine.

This is with the 'new' Sandra (well, 2002 version), professional.

I WILL SAY RIGHT NOW, that the OLDER Sandras had MUCH CLOSER float and INT scores !!

my roommate's computer, at FSB 133, and 733 mhz, is destroying my old 600e @ 900's scores, which were at 150 FSB. But I was using an old Sandra version when I last tested that. I got 580 int, 490 float.

So, this DOES look to be related to the chipset itself.
Does Data Prefetch actually affect *only* float scores? or does it only affect raw CPU?

It looks like, at least the new Sandra's memory scores, are VERY reliant on the FSB, and CPU speed plays a negligable part.

Can someone please confirm that (lack of) Data Prefetch is the actual cause of the lower float inefficiency, or rather, that data prefetch "working" is what kicks up the float score? For some reason, I don't think it does... I feel it's something just related to the BX chipset.

To those (especially with i815 chipsets), who are getting the SAME INT and FLOAT scores, just to prove that it is not the result of data prefetch, can any of you get a NON tualatin coppermine, and stick it in your i815, and check your scores? If they are still identical (meaning, no 53% efficiency on the Float), then clearly it's not the result of prefetching.
 
I seem to remember something like this happening when

"in order que depth"

Not being set a certain level.

I never remember my PIII 800e on a BE6-II getting scores that varied that much. Usually at 120MHz fsb they usually got around 800/800 MB/s. If in order que depth was screwed up though I would get almost half that. Your scores sound bad with this but how does it score in other non sandra benchies?
 
Ol'man,
Which benchmarks did you want me to run?

3dmark 2001SE gives a nice score of +10,580 3dmarks, at FSB 150, 1575 mhz, GF4 TI 4600 @ 320/715.

That's about on par (with the 30.xx series drivers, at least), as people with AXP 1800's and 1900's. This, of course, at 1024x768@32, no fsaa, no aniso. I'm beating some XP1800 scores...(and that's with PC133 on my system :)

Oh, and I don't remember the exact scores for sure (lost the printout), but the last time I checked, I recall UT2003 gave me 135 FPS for the flyby, and 54 FPS for botmatch at 150 fsb. Which was also beating a few people with XP 1900's :) Also tested at the plain vanilla no frills 1024x768@32

you want me to run any others?

BTW, on my roommate's abit BF6, in order queue depth is also set properly, and his benchmarks are ALSO similar as mine, float bandwidth almost half the speed of INT bandwidth; efficiency at 53%. (994/533). And he's using a 733 mhz coppermine, and everything's working as it should.

Since the BH6 doesn't support Queue depth settings in the BIOS, I checked with AGPset, and it says CPU Pipelining Queue Depth is set to 8 (maximum), so that's right. WPCREDIT also reports "In order queue depth" set to maximum (the bit is set on). WBXtune (BX chipset tuner) reports In order queue depth set to 4(max). Never heard of a "4" setting before, could be a typo, and WPCREDIT only has "1 and 0" for values which mean "max and min", no values inbetween, and AGPset says it's set at 8 (only other choice is 1(minimum). So that can't be the problem.

Minor update:

I just set the In Order Queue depth, on my roommate's BF6 (in BIOS) from 8 to 1, and his RAM score dropped to 326 (INT), and 223 (float), the INT being roughly one third (!) of what it usually is (640'ish), and Float being about just less than half. With it set on max(8), his scores is what I said above, 994 INT, 533 float.
Although that score was 2 days ago, go figure..

I just ran sandra again just now, on his BF6 computer, and he got 794 float, 448 INT at 133 FSB. Notice the score separations? And he's at cas 2-2-2, also. (I wonder how he managed 994/533 when I tested earlier?)

So it definitely isn't that :)

Ol'man:
Remember that the older version of Sandra reported much closer int/float bandwidth scores, although both scores were lower. I remember getting 580 MB/s max for Int, and 490 MB/s for float, in one of my tests, and that was at 155 FSB. I forgot if that was on a 600e @ 930 or a 1000EB @ 1162...I think it was on the 600e...

Can someone with a BX board run one of the newer Sandras? I could check some of the previous threads, but that would take awhile, and no idea which version they are using...
 
Last edited:
Only thing I can thing of is for me to run some benchmarks at 1.4GHz 133fsb and comapre to your and see if they vary. I don;t have a 4600ti so maybe we should stay away from 3d mark;)

Lets figure out which benchies to run and also stay away from sandra memory. Lets check out the cache mem test and compare. How is that?

What other benchies can we do? I have a few. Sounds like we need some FPU tests since it is that area it is giving low benchies on. Hopefully it is a error. I can't get to it right now but will be able to very soon. 2 hours? In the meantime if you get this figure out which tests to do as also I.
 
ol' man said:
Only thing I can thing of is for me to run some benchmarks at 1.4GHz 133fsb and comapre to your and see if they vary. I don;t have a 4600ti so maybe we should stay away from 3d mark;)

Lets figure out which benchies to run and also stay away from sandra memory. Lets check out the cache mem test and compare. How is that?

What other benchies can we do? I have a few. Sounds like we need some FPU tests since it is that area it is giving low benchies on. Hopefully it is a error. I can't get to it right now but will be able to very soon. 2 hours? In the meantime if you get this figure out which tests to do as also I.

Why don't you do Superpi and Prime95 benchmark, they are both highly dependant on memory bandwidth?
 
Honestly I think this is probably just a bug in Sandra that they overlooked. I highly doubt that SiSoft has a modded BX board with a Tualatin as one of their test systems. :p

The only program that seems to show this massive decrease in Float mem bandwidth is this latest Sandra memory benchmark test. While real programs show no such decrease, and some actually increase. In other words, don't worry about it because it doesn't matter. Let's all take a deep breath and say "it's just a benchmark test". ;)

I think benchmarks should start coming with a disclaimer of some sort, like on the television broadcast system tests...

This has been a test of the Computer Benchmarking System. The programmers on our internet, in voluntary cooperation with federal, state, and local networks, have developed this system to keep you informed in the event you want an estimate of your computer's performance. If this had been an actual program, the code you just ran would have been followed by work, gaming, or other goals being accomplished. This test serves the world wide web. This concludes this test of the Computer Benchmarking System. :D

-IronPlasma
 
I finished all the benchmarks, except the prime test, which I have no idea how to do. How can I send them to you, ol'man? Can I email them to you?

The results are very similar, except for the sandra memory benchmark...
 
Well in benchmarks I was actually running a 134fsb which the ABIT defaults to when you run it at 133fsb in the bios. You beat me in prime actually and I slightly beat you in super pi but I am running a 1410MHz or so as also a 134fsb. In prime you kinda beat me a little worse though. THe BX does shine in apps like this and should perform higher in mem dependent benchmarks clock for clock.

I would say sandra is screwed up and not your system.
 
Back