The winquake results might've something to do with some kind of incompatibility issue with the +'s?
The CTX/non-CTX might be another incompatibility that's roughly similar. How did the non-CTX perform in winquake compared to the CTX model?

And what motherboard are you running with?

Tetrium wrote:

The winquake results might've something to do with some kind of incompatibility issue with the +'s?
The CTX/non-CTX might be another incompatibility that's roughly similar. How did the non-CTX perform in winquake compared to the CTX model?

And what motherboard are you running with?

old k6-2 core scores 46fps in winquake, which is 2fps slower than cxt and seems reasonable.
my board is TMC mvp3 with 2mb cache, and i wanna know if that superpi performance of k6-2-300 is inline with other people's.

Well, write allocation is a feature which usually should enhance performance, but depending on the memory access patterns it's possible that it reduces performance in this case. I suppose writing to random memory locations would be an example were it would make things slower. Don't know if this applies to SuperPI.
Write combining should never make things slower, but it's possible that this breaks stuff. (ie. corrupted graphics)
Write allocation is not automatically enabled. If you have a BIOS that supports the chip it will enable WA. If you want to try you can get one of the K6-2 specific tools and just disable WA on the CXT and benchmark again.

idspispopd wrote:

Well, write allocation is a feature which usually should enhance performance, but depending on the memory access patterns it's possible that it reduces performance in this case. I suppose writing to random memory locations would be an example were it would make things slower. Don't know if this applies to SuperPI.
Write combining should never make things slower, but it's possible that this breaks stuff. (ie. corrupted graphics)
Write allocation is not automatically enabled. If you have a BIOS that supports the chip it will enable WA. If you want to try you can get one of the K6-2 specific tools and just disable WA on the CXT and benchmark again.

thanks for your advice, but i am not familiar with k6 tweaking tools, can you give me some info or links, especially something that can work in win2000/xp?

It's a long time ago I tried K6 tweaking, but there are still tools around.
Here you can find the SetK6 tool: http://www.heise.de/ct/ftp/98/19/224/
On this old page there is some documentation in German about this: http://www.oocities.org/dvd3500/k6opt.htm

i downloaded the setk6 tool, which has both dos and windows version which claimed to work in winnt, unfortunately neither version seemed to work.
when i ran "setk6 /on:256M", both versions displayed "enabling write allocation". but then, they still detected write allocation to be disabled, and their benchmarked memory performance remains unchanged.
i ran it with k6-3+ cpu, maybe it doesn't work with the + cpus?

Do you have a memory manager loaded that blocks privileged opcodes? (like EMM386?)

elianda wrote:

Do you have a memory manager loaded that blocks privileged opcodes? (like EMM386?)

i tested the windows version in win2000, and the dos version in win98 dos mode(f8 menu, not dos window), without emm386. they didn't display any error warning, but after enabling they still report write allocation to be disabled.

You could try some other tools.
http://www.k6plus.com/phpBB3/app.php/dl_ext/?cat=3
As I said, it's a long time I did that stuff myself.

i swapped in a k6-2-550, and the setk6 tool works! it detects WA to be enabled at boot, and the reported memory performance is about the same as k6-3+550. i can also disable WA with it, and it displays a much lower memory performance.
so i assume that the k6-2+/3+ cpus would force WA enabled and can't be disabled.

further benchmarks indicated that superpi does run a lot faster(by about 15%) with WA disabled, and winrar is also 4% faster. however other benchmarks suffered considerably without WA, especially in quake2, where k6-2-550 is exactly the same speed as k6-2-300!

I ran tests at 3x100 using a Jetway 542C Aladdin 5 motherboard and I can confirm that the CXT core is slower in SuperPi than the Chomper core, in my tests the difference is 5%.

The CXT core seems to be ~7% faster when it comes to MPEG video decoding (S3 Trio64V+ 2MB PCI) otherwise the Chomper and CXT cores seems pretty evenly matched in my tests (0-3% difference).

Write allocation was automagicly enabled by the BIOS with both CPUs, or that is at least what the patched BIOS notes implies. The memory test in Speedsys seems to favor the CXT core with some margin.