K6-III "problem"

[Olof F]

I recently bought a K6-III/400 2.4V and upgraded a daughter's MMX/233
machine. This machine has an ASUS TVP4 MB with 66 MHz FSB, 512 KByte
L2 cache, 128 MByte EDO. The OS is Windows Me.

The resulting system is much faster, memory benchmarks show reads that
are close to twice as fast as with the MMX/233. The MB reports the
processor as K6 300 but DMI says it is running at 400 MHz.

The funny thing is that memory writes are the same speed regardless of
whether L1, L2, L3 or main memory is being accessed. They all report
main memory write times (roughly 109 MBytes/s).

If I were to speculate on what the problem appears to be I'd say that
the processor is set to "unconditional write-through
cacheing" which means that no write will be acknowledged until
it has reached its final destination (main memory).

Does anyone know if this is a K6-III cache option and if there is a
95/98/Me program which will patch the processor to a more
"constructive" setting?

 

[nobody]

I recently bought a K6-III/400 2.4V and upgraded a daughter's MMX/233
machine. This machine has an ASUS TVP4 MB with 66 MHz FSB, 512 KByte
L2 cache, 128 MByte EDO. The OS is Windows Me.

The resulting system is much faster, memory benchmarks show reads that
are close to twice as fast as with the MMX/233. The MB reports the
processor as K6 300 but DMI says it is running at 400 MHz.

The funny thing is that memory writes are the same speed regardless of
whether L1, L2, L3 or main memory is being accessed. They all report
main memory write times (roughly 109 MBytes/s).

If I were to speculate on what the problem appears to be I'd say that
the processor is set to "unconditional write-through
cacheing" which means that no write will be acknowledged until
it has reached its final destination (main memory).

Does anyone know if this is a K6-III cache option and if there is a
95/98/Me program which will patch the processor to a more
"constructive" setting?

Is the BIOS updated to the latest? Later BIOS may (but not
necessarily will) solve it.

NNN

 

[pigdos]

If you're using cachechk, by Ray Van Tassle, this is by design. He doesn't
want to measure the speed of writes to L1, L2 or L3 cache. Here's his
explanation:

Q: No matter how I set the write cache policy (copy-back or write-through),

CACHECHK reports the same speed. Isn't copy-back (AKA write-back)

supposed to be better? What gives?

A: It *is* better, but only when reads and writes are intermixed. By

design, CACHECHK does not intermix operations. It either reads in a

sustained burst, or writes in a sustained burst, so the write policy never
comes

into play. CACHECHK is designed to saturate the memory read/write channel
and

to report the timings in this condition. In fact, in order to be able to

detect and time the various stages of the memory system, it is absolutely

necessary to throw as much cr*p at it as it can possibly handle.



So bascially, you'll never see a difference in terms of memory writes with
his program.



There is a program that allows you to turn on write-combining for K62 and
K63 CPU's, but aside from

that there's nothing you can really do to tweak K63's or K62's.

 

[Scott Alfter]

Is the BIOS updated to the latest? Later BIOS may (but not
necessarily will) solve it.

Also, the K6-III normally runs on a 100-MHz FSB, not 66. If the motherboard
supports it, I'd recommend cranking the FSB speed up past 66 MHz. Reduce
the FSB multiplier from 6 to 4 when you do that to keep the processor speed
where it should be.

_/_
/ v \ Scott Alfter (remove the obvious to send mail)
(IIGS( http://alfter.us/ Top-posting!
\_^_/ rm -rf /bin/laden >What's the most annoying thing on Usenet?

 

[George Macdonald]

I recently bought a K6-III/400 2.4V and upgraded a daughter's MMX/233
machine. This machine has an ASUS TVP4 MB with 66 MHz FSB, 512 KByte
L2 cache, 128 MByte EDO. The OS is Windows Me.

The resulting system is much faster, memory benchmarks show reads that
are close to twice as fast as with the MMX/233. The MB reports the
processor as K6 300 but DMI says it is running at 400 MHz.

The funny thing is that memory writes are the same speed regardless of
whether L1, L2, L3 or main memory is being accessed. They all report
main memory write times (roughly 109 MBytes/s).

If I were to speculate on what the problem appears to be I'd say that
the processor is set to "unconditional write-through
cacheing" which means that no write will be acknowledged until
it has reached its final destination (main memory).

Does anyone know if this is a K6-III cache option and if there is a
95/98/Me program which will patch the processor to a more
"constructive" setting?

I'm afraid it's all a bit hazy already but yes, if the BIOS did not support
the K6-III, there was a program written by Andreas Stiller which would turn
on Write Allocation and Write Combining in the CPU. IIRC on an uncached
write, Write Allocation would read in the cache line being written to and
flush a "dirty" line it replaced. The prog is still available here
ftp://ftp.heise.de/pub/ct/ctsi/setk6v3.zip but I'm not sure it's going to
make *that* much difference.

 

[Roger Hunt]

If you're using cachechk, by Ray Van Tassle, this is by design. He doesn't
want to measure the speed of writes to L1, L2 or L3 cache. Here's his
explanation:

Q: No matter how I set the write cache policy (copy-back or write-through),
CACHECHK reports the same speed. Isn't copy-back (AKA write-back)
supposed to be better? What gives?
A: It *is* better, but only when reads and writes are intermixed. By
design, CACHECHK does not intermix operations. It either reads in a
sustained burst, or writes in a sustained burst, so the write policy never
comes
into play. CACHECHK is designed to saturate the memory read/write channel
and
to report the timings in this condition. In fact, in order to be able to
detect and time the various stages of the memory system, it is absolutely
necessary to throw as much cr*p at it as it can possibly handle.

Good to see a mention of Cachechk! Most excellent prog.
v7 is the furthest he went, isn't it?

 

[Franc Zabkar]

I recently bought a K6-III/400 2.4V and upgraded a daughter's MMX/233
machine. This machine has an ASUS TVP4 MB with 66 MHz FSB, 512 KByte
L2 cache, 128 MByte EDO. The OS is Windows Me.

The resulting system is much faster, memory benchmarks show reads that
are close to twice as fast as with the MMX/233. The MB reports the
processor as K6 300 but DMI says it is running at 400 MHz.

The funny thing is that memory writes are the same speed regardless of
whether L1, L2, L3 or main memory is being accessed. They all report
main memory write times (roughly 109 MBytes/s).

I recommend the Celem Cache Test:
ftp://ftp.sac.sk/pub/sac/utildiag/cct386.zip

It's an old DOS application but it still works well with my AMD K6-2
450. It produces a stepped graphical display of memory performance. I
swear by it.

Although it's for a different motherboard, this page has some tech
notes on the K6-3 which you may find useful:
http://m571.com/m571/m571_amd.htm

- Franc Zabkar

 

[nobody]

Also, the K6-III normally runs on a 100-MHz FSB, not 66. If the motherboard
supports it, I'd recommend cranking the FSB speed up past 66 MHz. Reduce
the FSB multiplier from 6 to 4 when you do that to keep the processor speed
where it should be.

_/_
/ v \ Scott Alfter (remove the obvious to send mail)
(IIGS( http://alfter.us/ Top-posting!
\_^_/ rm -rf /bin/laden >What's the most annoying thing on Usenet?

These things were good overclockers. With good cooling they worked
well up to 600, so the multiplier may stay the same. Extra 50MHz
wouldn't hurt. Wouldn't recommend to crank FSB above 75 though - at
83 most harddrives crap out (or is it IDE controller issue? doesn't
matter - the result is the same - corrupt data).

NNN

 

[Tony Hill]

These things were good overclockers. With good cooling they worked
well up to 600, so the multiplier may stay the same.

I believe you're thinking of the K6-III+ and note the standard K6-III.
The original K6-III had a heck of a time overclocking by more than
about 1MHz. The latter '+' chips were sold at the same clock speeds
but made on a more advanced process and overclocked quite well.

 

[nobody]

I believe you're thinking of the K6-III+ and note the standard K6-III.
The original K6-III had a heck of a time overclocking by more than
about 1MHz. The latter '+' chips were sold at the same clock speeds
but made on a more advanced process and overclocked quite well.

Maybe - now I remember that mine was +, and it worked at 600 from 2000
to 2004 quite stable. Then either the CPU or the mobo developed some
problems. I am more inclined to blame the mobo - it was an older
circa '97 VA503+, and the capacitors could've easily gone bad in 7
years.

NNN

 

[Roger Hunt]

Maybe - now I remember that mine was +, and it worked at 600 from 2000
to 2004 quite stable. Then either the CPU or the mobo developed some
problems. I am more inclined to blame the mobo - it was an older
circa '97 VA503+, and the capacitors could've easily gone bad in 7
years.

I have a PA-2013 which is still going strong.
When my 503+ died it was most definitely my fault, caused by the final
death-throe reaction to years of abuse.

 

[pigdos]

It was/is a great program. I hate to admit it, but I never did end up
sending Ray his postcard.
[/QUOTE]
Good to see a mention of Cachechk! Most excellent prog.
v7 is the furthest he went, isn't it?[/QUOTE]

 

[Roger Hunt]

On Tue, 21 Nov 2006, pigdos typed this :
(Cachechk)

It was/is a great program. I hate to admit it, but I never did end up
sending Ray his postcard.

I did, from the British Museum - a picture of many gold coins.
Trouble is, I was never sure if the address was still his.