P-II 450/527 BENCHMARKS - NUMBERS & THOUGHTS

[Roy]

A few days ago I posted some results from Ziff-Davis' CPUMark32. Eric Mintz of www.pcbenchmarks.com pointed out that this test has been replaced by CPUmark99.

I downloaded the new test and made some more runs. The results clearly showed the difference between SEL 100/66 settings but was less sensitive to CAS Latency settings on the Abit BX6-2. A review by ZD's PC Magazine's PC Labs explains that the new test measures performance in mainstream business applications. It features a better mix of instructions and accounts for newer cache and memory sizes. It excludes floating point and MMX instructions!

I feel that this tailoring reduces the test's usefulness to overclockers and gamers.
Further, the results posted at pcbenchmarks are incomplete and difficult to believe. They do not show what the default speeds of the processors are. They exceed the range of numbers shown in the PC Labs review by a more than reasonable margin.

For example, my P-II 450 @ 450 tests at 33.5. This is the same number as is shown in the PC Labs table. The various 450 results (from P-IIs and Celerons) in the pcbenchmarks table range from 36.9 to 39.1. How? Why?

SiSoft's SANDRA has a test that seems more meaningful. The Memory Benchmark evaluates the CPU, Chipset, Cache and Memory. It's results in my tests track faithfully with the speeds and settings.

The following results are from a P-II 450 on an Abit BX6-2 mobo, using a single 64Mb CL2 memory module. CPUmark99 tests were run 3 times and showed almost no variance. The SANDRA Memory Benchmark tests were run 10 times and the last 3 results were averaged. This is because in 10 of the 16 sets of runs, the numbers started out higher and tapered off to stable, but somewhat lower, values. I cannot explain this phenomenon but its lesson is clear. Don't accept the first number you get.

UBB does not like tables, so I'll group the figures according to CL setting / SEL 100/66 setting. In each case, the first number is the CPUmark99 result, the second is SANDRA's

CL3, SEL Hi/Lo
450 - 33.5/34.8 - 213/220
464 - 34.4/35.7 - 219/227
504 - 37.5/39.0 - 238/247
527 - 39.2/40.7 - 248/258

CL2, SEL Hi/Lo
450 - 33.6/35.0 - 216/224
464 - 34.5/35.9 - 223/230
504 - 37.6/39.1 - 242/251
527 - 39.3/40.8 - 253/262

I look forward to your comparisons and comments.

[This message has been edited by Roy (edited 02-08-99).]

[jkauf58]

Roy,

I have a PII450, Abit BX6, 128mb PC100 RAM, IBM 10.1 7200rpm harddrive, ATI expert@play98 AGP video card, Win98. I thought that I would write you because of your knowledge on overclocking a PII450. My computer will run fine at 450 and at 103 (turbo enabled). When I set it to 504 @ 112, I receive the following error while surfing the web with Netscape and IE4: This program has performed an illegal operation and will be shutdown. Iexplore caused an invalid page fault in Kernell32.dll at 023f:6ff9d709. Could it be the RAM causing this error? Or could it be a buggy software program? I am looking forward to hearing from you, and possibly a solution to maybe cracking this problem.

Thank you,

Jay

[Roy]

Most (but not all) PC-100 SDRAM should perform OK at 112MHz. Try putting a real load on it by opening several programs at once and bouncing around among them to see what happens. Getting the same type of fault would point away from the browsers.

The invalid page fault message contains an answer, you could try posting a query on the Tech Support board to see if anyone can translate the address.

[lonnier]

And I have pushed this cheap piece of dirty silicon as far as I can. This celeron 300A will only go to 450. Some measures are:

WinBen98.10___cpu_fpu__955__2410
WinBen99.11___cpu_fpu_37.7__2410
Sisoft.Sandra_cpu_fpu_1077__305__mem_250
WinTune98.32__cpu_fpu_1319__525__mem_804

( I plan on getting a amd k6-3 )

[Eric]

Roy, you are correct regarding CPUMark99, it's usefulness is limited for gamers (although not necessarily for overclockers, depends on what areas of performance you are trying to improve).

As for variations in performance... first of all, I don't have any results from a straight P2-450. But the Celeron-A scores slightly higher than the P2 at the same clock speed, so that may account for the higher scores. However, remember that when PC Labs tests systems, they are almost always doing so on systems made with Intel motherboards, which have been setup for stability and not necessarily top speed.

As for variability in the tests... well, this points out the drawback of such surveys, namely that some people may have background processes running on their systems that could affect the results. The positive side is that many more configurations are tested than would be possible in a rigorous, controlled review.

I don't really understand your conclusion regardin Sandra's memory benchmark vs. CPUMark99. Based on your results, they appear to be measuring exactly the same thing. If you line up all the results in order by score, they line up together almost exactly. The highest CPUMark99 you got, 40.8, is 22% higher than the lowest at 33.5. The highest Sandra memory score you got was 262, which was 23% higher than the lowest at 213. Looks to me like they measure the same thing!

Incidentally, I vaguely recall reading somewhere that Sandra gives strange results with K6-2 chips, but I don't remember where I saw it. Anyone know anything about that?

It would be nice if there were a quick synthetic test for gaming performance, but I doubt we'll see it. Instruction mixes for Quake may be very different than for Half-life etc. That's why the frame rate benchmarks are so popular.

Futuremark is soon to release a new version of 3DMark99. This upgrade will have support for both 3DNow! and KNI. I think it'll be useful... and I'll host a comparison list for it when it appears. I wish I had access to sufficient hardware to run the kind of rigorous comparisons I'd like, but I don't. Ah well.

-----------
Visit PC Benchmarks!
www.pcbenchmarks.com

[Roy]

Eric is correct. Neither test is perfect for everyone. The SANDRA tests do, however, show a subtle edge over CPUmark99. The downside is that SANDRA considers the amount of memory as well as performance. It's one of those variables Eric mentions.

My results for a "straight P2-450" may be exceeded in some tests by a Celeron running at the same speed because of the faster L2 clocking and its smaller size. The SEL 100/66 Lo setting helps here because it results in overclocking the L2. (According to Oskar Wu, BX6-2 co-designer.)

We all look forward to each new standard in testing.

[Roy]

Ignore this.

[ZX1100]

It's not nice to raise your Personal posts to the Top unless you have something new to Add. lol, Jim

[shocker]

Hmmm...

My Celeron at 504Mhz gives a CPUmark99 @ 42.3

Let me run it again...

...ok, I'm back and still online 41.5, that can be my low score.

Soyo 6BA+ motherboard running at 2.0 volts.
I have 64MB of CAS2 8ns RAM. Memory is running at the fastest settings with no crashes.

At 450Mhz my scores ranged from 37.9 to 39.4

Weird how different the scores are between systems.

Interesting stats, thanks for the post.

[Roy]

I hope there's a prize for the one who reveals the reason for Celeron's higher marks.

It will be interesting to see if this changes with the 0.18 generation. It's almost here.

[shocker]

Maybe it's the internal cache on the Celeron that is giving the higher marks???

[clocker]

I just ran SandraSiSoft and I am getting

1170 cpu and 420 fpu

AMD K6-2 300 Non CXT o/c to 392 (3.5 x 112)

CPUMark99 I got 29.9 while online....

[This message has been edited by clocker (edited 02-16-99).]

[Roy]

The CPU Benchmark in SANDRA measures only the speed. Almost worthless.

Use the Memory Benchmark. It analyzes CPU, Chipset, Cache and Memory.