Yes there is a driver. For DOS it's a CONFIG.SYS driver but I can't remember the name. Just that I have used it and it worked just fine 😀

as already mentioned, 0.18um k6s(k6-2+ and k6-3+) maps 2.5x to 2x, 2.5x is missing instead.
and old revisions of k6-2 ("26351" numbers in black color than gold) would recognize 2x aswell.

feipoa,

it is k6dos.sys,
look here:

http://k6plus.com/index.php?name=Downloads&re … wdownload&cid=1

Jolaes76 wrote:

feipoa, […]
Show full quote

feipoa,

it is k6dos.sys,
look here:

http://k6plus.com/index.php?name=Downloads&re … wdownload&cid=1

Many thanks. I'll check it out!

I solved my K6 (original K6) issues. Apparently, the 2.2V stated on the K6-300AFR CPU was insufficient at 266-333 MHz. I had to run it at 2.4 V for 266-300 MHz and 2.56 V for 333 MHz. I'm now moving onto K6-2's.

noshutdown wrote:

as already mentioned, 0.18um k6s(k6-2+ and k6-3+) maps 2.5x to 2x, 2.5x is missing instead.
and old revisions of k6-2 ("26351" numbers in black color than gold) would recognize 2x aswell.

Unfortunately, I have a K6-2-300 with "26351" in gold printing. It remaps 2x to 6x and 2.5x is still 2.5x. Does anyone in Canada have a 26351 with black lettering?

Jolaes76 wrote:

feipoa,
it is k6dos.sys

Unfortunately, the k6dos.sys software did not change the multiplier to 2x.

Does anyone else know how to set my K6-2 into 2x mode? I'd really like to include a 133 MHz K6-2 in the Ultimate 686 Benchmark Comparison.

Is there another multiplier which remaps to 2x? Is there any other software which works with the K6-2? The K6dos.sys readme mentions it only works with the + chips. Is the performance of the K6 and K6-2 equivalent at the same clock speed? If I can disable the L2 cache on the K6-2+, will the performance match that of a K6-2? Does anyone know what software disables the cache?

At present, the closest I can get is 137.5 MHz (55x2.5).

If memory serves me right, the old K6-2's have the number 26050. These chips don't remap the 2x multi to 6x

I think k6dos.sys requires/uses the PowerNow! feature of the CPU (i.e. software controllable multiplier).

If you can't get the clock to 2x66 I would still go for 2.5 x 66 and scale it down. At 55 x 2.5 you change a lot on the remaining system.

Oh and maybe also interesting:
http://web.archive.org/web/20090112234050/htt … spec/ct9918154/

I found an AMD K6-2-300 with 26351 in black lettering for $8 shipped. I hope the information provided above was correct. $8 is a lot of money to add one column to a chart, but after comparing the K6 and K6-2 results at 233 MHz, I have determined that the two chips are different enough to warrant needing a K6-2 which will operate at 133 MHz. The K6-2 seems quite a bit faster with memory speed and substantially faster with its FPU. The addition of 3DNow! was no joke.

I'm not sure how I'd easily scale down a K6-2 at 166 MHz to 133 MHz. I'd have to linearise all 97 tests from 166, 200, 233, and 266 to generate values for 133. It is a lot of work considering I was not even going to test the K6-2 at 166 and 200 MHz. I also agree that 137.5 MHz with a different FSB is not the way to go.

I do plan on linearising the data to create a Pentium P55C-300MMX. I feel this is somewhat valuable information if anyone finds a working Tillimock in a desktop board, or finds a way to modify the chip to work. Also, perhaps some rare P55C-233MMX chips will operate at 300 MHz. Mine worked fine up to 262 MHz.

I'm about half way through with the Ultimate 686 Benchmark Comparison. My borrowed K5-PR200 should be here on Monday.

I would not linearize it since the dependency is not linear for sure.
If you have at least 3 measurement points use a quadratic polynom since it is like a second order pertubation using taylor expansion for a unknown function.

elianda wrote:

I would not linearize it since the dependency is not linear for sure.
If you have at least 3 measurement points use a quadratic polynom since it is like a second order pertubation using taylor expansion for a unknown function.

What makes you think it is non-linear and may require a Taylor/Maclaurin series for sufficient precision? If the performance to frequency is anything like it was for 486 CPUs, it is most certainly linear over the small range of 150 MHz I looked at. I'd be happy to send you the data on this if you think you can improve upon the 150, 180, and 200 MHz generated scores from the Ultimate 486 Benchmark Comparison - in fact, that data was already posted and is at your disposal!

If you think you can extract a higher degree of precision such that it is worth your time, I'd be happy to send you my existing 686 data. However, please be willing to commit yourself 100% to this as it will require that I test two more CPU frequencies which I would not otherwise.

The prospect of project collaboration is getting me excited, however let's take a step back for a moment; please confirm your theory with the Ultimate 486 Benchmark Comparison data before I proceed with, potentially, unnecessary tests.

It has been a long time since I've needed to deal with Taylor expansions, or even the majority of Calculus. It just doesn't come up much in my line of engineerig work, in fact, I don't know many engineers who have had to use it. So if it is fresher in your mind, then knock your socks off!

feipoa wrote:

I found an AMD K6-2-300 with 26351 in black lettering for $8 shipped. I hope the information provided above was correct. $8 is a lot of money to add one column to a chart, but after comparing the K6 and K6-2 results at 233 MHz, I have determined that the two chips are different enough to warrant needing a K6-2 which will operate at 133 MHz. The K6-2 seems quite a bit faster with memory speed and substantially faster with its FPU. The addition of 3DNow! was no joke.

Is the K6-2 FPU actually faster on its own (ie FPU tasks without 3DNOW! extensions implemented), or are you specifically talking about the speed boost for drivers/apps with 3DNow! support?

I know the MMX unit in the K6-2 is substantially faster than the K6 and a few other tweaks, but nothing about the FPU itself being improved.

I'd actually gotten the impression that the K6-2 was very slightly slower (around 1%) ALU/memory than the K6 classic running at the same core and FSB speeds. ie, a K6/300 vs K6-2/300/66 or K6 with overclocked FSB, since K6/300s often work fine at 3x100 and occasionally 3.5x100.
The per-clock performance drop was supposedly tied to more conservative I/O timing for the K6-2, to allow for better 100 MHz bus yields, and consequently also a better chance for overclocking beyond 100 MHz.

kool kitty89 wrote:

Is the K6-2 FPU actually faster on its own (ie FPU tasks without 3DNOW! extensions implemented), or are you specifically talking […]
Show full quote

Is the K6-2 FPU actually faster on its own (ie FPU tasks without 3DNOW! extensions implemented), or are you specifically talking about the speed boost for drivers/apps with 3DNow! support?

I know the MMX unit in the K6-2 is substantially faster than the K6 and a few other tweaks, but nothing about the FPU itself being improved.

I'd actually gotten the impression that the K6-2 was very slightly slower (around 1%) ALU/memory than the K6 classic running at the same core and FSB speeds. ie, a K6/300 vs K6-2/300/66 or K6 with overclocked FSB, since K6/300s often work fine at 3x100 and occasionally 3.5x100.
The per-clock performance drop was supposedly tied to more conservative I/O timing for the K6-2, to allow for better 100 MHz bus yields, and consequently also a better chance for overclocking beyond 100 MHz.

Good questions. The FPU itself appears faster, even for 3DNow!-unaware DOS software. For example, ByteCPU 32-bit, FPU on the K6-233 scored 215% of a P90, whereas the K6-2-233 scored 239% of a P90. The 3DNow!/MMX instructions are significantly faster though, for example 3DMark99Max has the K6-233 at 1292 CPU 3DMarks and 3487 for the K6-2-233. The memory throughput also scores faster for 3DNow!-unaware apps, like Cachechk. In Cachechk, the K6-233 scores 278 MB/s and 130 MB/s for L2 and RAM, whereas the K6-2-233 scores 317 MB/s and 184 MB/s, respectively.

Some tests do, however, indicate a slower ALU on the K6-2 for the same clock, for example Wintune98's Integer test has the K6-233 at 591 MIPS and the K6-2-233 at 538 MIPS. Similarly, for Sandra99's Dhrystone, 710 MIPS for the K6-233 and 647 for the K6-2-233.

noshutdown wrote:

as already mentioned, 0.18um k6s(k6-2+ and k6-3+) maps 2.5x to 2x, 2.5x is missing instead.
and old revisions of k6-2 ("26351" numbers in black color than gold) would recognize 2x aswell.

Are you refering to the K6-2 with black numbers on the chip surface as well as with the 26351, or can the chip surface contain the engraved markings? I hope I didn't just blow $8.

Were there any changes in CPU performance between the K6-2 Chomper and Chomper XT?

EDIT: Ok, that was a waste of $8. According to this post on cpu-world, I need one with 26050. The colour of the print is not important.
http://www.cpu-world.com/forum/viewtopic.php?t=18983