Tuning NF7-S and Athlon Mobile 2600+ for images and audio / low energy use

Hi,

I return to this newgroup, after having some good advice some months
ago about building a PC with a Mobile 2600+. My first 'clocking', but I
did a lot of home work ;-)
I've now build it and busy tuning it. In daily use I don't need a
superfast PC (just doing E-mailing, webmastering etc.) and want it to
run as energy efficient and silent as possible.
But I also need a powerplant to work on 8 megapixel images (11 Mb raw
files) and hi-res audio (WAV, 94Khz/32 bit up to several Gb). In that
case I want to put it in high gear, as fast as I can get it running
stable.
I'd like to switch gears (incl. multiplier if possible) on the fly, one
of the reasons for buying a NF7-S.
I'm also writing an article about my experiment, making the point that
a +/- 3 Ghz desktop PC also can be (and should be) very energy
efficient and already can be build with normal everyday parts without
modding. And if I can do that as an amateur, the industry can do a much
better job, incorporating some laptop principles, which earn itself
back in a few months in saved energy.


The components:

Aopen H600A miditower with 350 watt power supply (should be enough for
the 45 watt mobile CPU)
Abit NF7-s v2 latest BIOS 27, got 22, its the 2 color memory slots
version
Athlon XP Mobile 2600+ AXMG2600FQQ4C
2x 3200PC 512 Mb noname
Coolermaster CP5-8JD1F-0L heatsink (nothing special, but quite silent)
Zalman ZM-NB47J passive northbridge cooler
Samsung ATA harddisks SP4002H (40 Gb) and SP8004H (80 Gb), Toshiba
DVD-ROM
Serial ATA switched off in BIOS
Old Diamond ViperII Z200 32 Mb AGP 4x videocard (first upgrade
candidate, but want passive cooling)

Running Win98 with a [vcache] entry in C:\windows\system.ini to allow
for 1 Gb:
MinFileCache=0
MaxFileCache=262144
May need to upgrade some time, but I don't want to waste my new
resources on Windows XP yet (Would like to go Linux if I can find
the right apps for images and audio).
Using 8rdavcore for clock and voltage settings
I'm using a X4-LIFE Inspector to measure total power consumption of the
whole box.

What I've already learned:

2x 256 Mb is faster, but I need the 1Gb.
Reading what others did and finding out myself, it seems hard to get
this combination over 210 FSB and that needs lots of Vcore. My
impression it's due to the instability of the Vcore (caused by the NF7
it seems). Using expensive memory trying to get more FSB is called a
waste of money, while others say: "it's picky on memory, should get
Geil etc."
The real Vcore on the NF7-S is always less than set in BIOS, and that
is not a power supply issue. (Even underclocked 5x100 at Vcore 1,1 both
AbitEQ and 8rdavcore read 1.07 volt, running Prime95).
New BIOS 27 (I got 22) helps to recognize the CPU, but looses any clue
when the multiplier is too low and resets memory timings to more
relaxed settings. Somtimes I get this red "Don't power off while
writing BIOS" screen after saving BIOS changes. Can't find a patern,
maybe it needs to 'rewrite everything' after a few cycles of changes?

My questions:

1. Im confused by the 'FSB/multiplier/running sync' relation.
What does "running CPU and memory in sync" mean for a mobile? The
BIOS has an 'auto' option, on by default. Does the Mobile 2600+ have a
'best' multiplier?

2. I know the 3.2 Gb speed for dual channel is only theory/marketing,
but are values around 1340 Mb/s (what Memtest86 v3.2 tells me) normal
with PC3200 memory?. (2x512MB PC3200 noname).
I've tried a lot at BIOS suggested 'Turbo' timings, but it failed even
at 9,5x200 7-3-3-2 1,575 Vcore.
Anyway, I'm at 11x200 at 1,725 Vcore and 2,9 Vdimm timing by SPD
8-4-4-2,5 and Prime95 running some hours now. Memtest said 2205 Mhz and
1345 Mb/s.
Is this memory crap? PC3200 should do FSB 200 out of the box, not? But
it needs 2,9 volt for this setting and the sticker says 2,5 volt.
Return these sticks?

It also seems to me my the 'clockable limit' went down permanently
during testing, is that possible? The highest I ever pushed it was
11x220 at 2,6 volt (not stable).

Experimenting with speeds (not running all tests) I also got some
confusing results with Memtest, all memory timings 8-4-4-2,5 by SPD
(most of the times equals Aggressive and Optimal):
11x200 gives 1345 MB/s
11,5x200 gives 1343 MB/s
12x200 gives 1342 MB/s
Slower with much higher CPU speed? Is that because the memory times are
getting more and more 'out of tune'? Is the CPU simply wasting cycles,
waiting for memory? What is there still to gain when memory stopped
supplying data any faster?

3. With all the variables the NF7-S with mobile offer, 2 ways seem
possible:
- Maximize CPU speed (Mhz)
- Maximize memory bandwidth (MB/s)
For for working on really large image and audio files, for what should
I go?
Or is there some 'sweetspot' in between?

4. I've found several tools to change NF7-S BIOS values on the fly, but
(still) nothing which does the multiplier. Anyone?


Some results energy-wise:

5x100 at 1,1 Vcore, CPU fan 1050 rpm, chasis fan off running Prime95:
32 degrees C for CPU, 47 watt.
Hmm, that 80 Gb Samsung is quite noisy

12x208 at Vcore 1,750 running Prime95 (did not last):
55 degrees C for CPU, 173 watt.

Current, seemingly stable 11x200 at 1,725 Vcore and 2,9 Vdimm:
47 degrees C for CPU, 137 watt.
This did go back to 127 watt in 2 hours 43 min.: prim. slave 80 Gb
switched off (did not know Prime95 can do that, but 10 watt seems a lot
for a idle HD, must be more to it, I'll have a look in the BIOS later


BTW, does anyone know which percentage of total energy use is used by
PC's?
US figures and others are all fine. Looking all over the place, only
hard figure I could find is 5% in 1992 when the EPA launched the Energy
Star Program.

Thanks,
Boyd Noorda

N.B. all voltages are BIOS settings, not the reported values.

wrote:
Hi,

I return to this newgroup, after having some good advice some months
ago about building a PC with a Mobile 2600+.
[...]
I'd like to switch gears (incl. multiplier if possible) on the fly,
one of the reasons for buying a NF7-S.

If you wanted to change gears on the fly you got the wrong chipset The
NF2 chipset doesn't support PowerNow transitions, so you're limited to
changing the FSB through ClockGen or similar programs.

[...]
1. Im confused by the 'FSB/multiplier/running sync' relation.
What does "running CPU and memory in sync" mean for a mobile? The
BIOS has an 'auto' option, on by default. Does the Mobile 2600+ have a
'best' multiplier?

This just means that the memory bus and the FSB are run in sync. Running out
of sync (say DDR333 and a 200MHz FSB) gives a reasonable performance hit.

[...]

--
Michael Brown
www.emboss.co.nz : OOS/RSI software and more
Add michael@ to emboss.co.nz - My inbox is always open

On Fri, 18 Mar 2005 21:26:26 -0800, joburgnewsdesk wrote:

I'd like to switch gears (incl. multiplier if possible) on the fly, one
of the reasons for buying a NF7-S.

You want Powernow, but it's not supported by the Nvidia chipset. I also
dont' know how well W98 supports it. With my K8, W98 pwernow only gives me
2 speeds, adn not the lowest. Maybe the K7 is better. But I don't know. I
also don't normally boot to windows. 99.9999% of the time I run Linux adn
cpufreqd to autoscale the cpu speed between 800, 1800, and 2000MHz with
vcores of 1.3, 1.4, and 1.5v. I think WinXP does this too. Couldn't say as
I've never seen it.

1. Im confused by the 'FSB/multiplier/running sync' relation. What does
"running CPU and memory in sync" mean for a mobile? The BIOS has an
'auto' option, on by default. Does the Mobile 2600+ have a 'best'
multiplier?

Michael explained running in sync. The best multiplier depends on what cpu
speed you want to accomplish so I guess the best wouldn't be the same for
everyone.

2. I know the 3.2 Gb speed for dual channel is only theory/marketing,
but are values around 1340 Mb/s (what Memtest86 v3.2 tells me) normal
with PC3200 memory?. (2x512MB PC3200 noname). I've tried a lot at BIOS
suggested 'Turbo' timings, but it failed even at 9,5x200 7-3-3-2 1,575
Vcore.
Anyway, I'm at 11x200 at 1,725 Vcore and 2,9 Vdimm timing by SPD
8-4-4-2,5 and Prime95 running some hours now. Memtest said 2205 Mhz and
1345 Mb/s.
Is this memory crap? PC3200 should do FSB 200 out of the box, not? But
it needs 2,9 volt for this setting and the sticker says 2,5 volt. Return
these sticks?

I hoped Michael would have answered this one as he's probably more into
memory than I am, or at least likes to write about it because it is
complex and I don't like giving complex answers with my limited typing
skills (meaning slow). I think you're getting the memory bus speed and FSB
speed confused. If your memory setting is BySPD, then the ram should run
at whatever is in the SPD data of the memory module. Most PC3200 rams spd
data sets the memory bus to 166MHz, not 200, for compatability reasons. If
you want 200MHz memory bus, you have to manually set it. I don't know how
the NF7 handles this so you should check to see if you really are running
the memory bus at 200MHz. But, the 2.9v for the ram is not good at all.
It's likely to burn out quick at that voltage. I personally wouldn't go
over 2.7v on a 2.5v module. there used to be a problem with a lot of
bioses misinterpreting the PSD data too and setting wrong values. I think
this was one of the reasons ram modules SPD data now defaults to 166Mhz.
Too many returns of perfectly good ram not set properly by the bios
defaults.

It also seems to me my the 'clockable limit' went down permanently
during testing, is that possible? The highest I ever pushed it was
11x220 at 2,6 volt (not stable).

If in fact you had the ram set to 200MHz, then overclocking the FSB to
220MHz would also overclock the memory bus to 220MHz unless you changed
the FSB/Ram bus ratio which I'm pretty sure is a setting in the NF7.

Experimenting with speeds (not running all tests) I also got some
confusing results with Memtest, all memory timings 8-4-4-2,5 by SPD
(most of the times equals Aggressive and Optimal): 11x200 gives 1345
MB/s
11,5x200 gives 1343 MB/s
12x200 gives 1342 MB/s
Slower with much higher CPU speed? Is that because the memory times are
getting more and more 'out of tune'? Is the CPU simply wasting cycles,
waiting for memory? What is there still to gain when memory stopped
supplying data any faster?

No, It's because with bySPD, the ram is out of sync running at 166Mhz more
than likley. manually set the ram bus. Loe\wer the voltage to 2.7v.

3. With all the variables the NF7-S with mobile offer, 2 ways seem
possible:
- Maximize CPU speed (Mhz)
- Maximize memory bandwidth (MB/s)
For for working on really large image and audio files, for what should I
go?
Or is there some 'sweetspot' in between?

The sweet spot is to run the ram in sync with the FSB, so leave the FSB
and ram at 200MHz and raise the multiplier for more speed. Now if you
determine your ram can run at higher speeds, then you can raise the FSB up
to that speed.

4. I've found several tools to change NF7-S BIOS values on the fly, but
(still) nothing which does the multiplier. Anyone?

Powernow, which you now know your chipset doesn't support.

--
Abit KT7-Raid (KT133) Tbred B core CPU @2400MHz (24x100FSB)
My server http://wesnewell.no-ip.com/cpu.php
Verizon server http://mysite.verizon.net/res0exft/cpu.htm

Wes Newell wrote:
On Fri, 18 Mar 2005 21:26:26 -0800, joburgnewsdesk wrote:
[...]
2. I know the 3.2 Gb speed for dual channel is only theory/marketing,
but are values around 1340 Mb/s (what Memtest86 v3.2 tells me) normal
with PC3200 memory?. (2x512MB PC3200 noname). I've tried a lot at
BIOS suggested 'Turbo' timings, but it failed even at 9,5x200
7-3-3-2 1,575 Vcore.
Anyway, I'm at 11x200 at 1,725 Vcore and 2,9 Vdimm timing by SPD
8-4-4-2,5 and Prime95 running some hours now. Memtest said 2205 Mhz
and 1345 Mb/s.
Is this memory crap? PC3200 should do FSB 200 out of the box, not?
But it needs 2,9 volt for this setting and the sticker says 2,5
volt. Return these sticks?

I hoped Michael would have answered this one as he's probably more
into memory than I am, or at least likes to write about it because it
is complex and I don't like giving complex answers with my limited
typing skills (meaning slow).

Hehe, yes, I ran out of time too when typing my answer (realised I was
supposed to meen someone 5 minutes ago). Probably should have just saved it
and come back to it later, but it looks like you did a fairly good job of it
Just a couple of comments ...

In this particular case, the bandwidth deficiency is due to the way that
memtest86 calculates memory bandwidth. The 3200 MBytes/sec is the
theoretical maximum speed, though due to bus inefficiencies it's rare to get
over ~90% of this, or ~2880 MBytes/sec. This is the sort of number Sandra
would (should) spit out at you. However this is measuring the memory
bandwidth in a different way.

The Sandra way of measuring it is so-called buffered memory bandwidth. What
this does is effectively just make the CPU pull the data in from memory to
the L1 cache (or write out a continuous stream). This is a highly efficient
way of doing things, and is made even better because you can be doing
calculations on the buffered memory while the CPU is pulling in the next few
cache lines. The way this is done is to use the load/store buffers in the
Athlon to have several outstanding requests to main memory being satisfied
in the background, while processing the data you pulled in just before. Due
to the streaming nature of this method, memory latencies matter very little.

The memtest86 way is less efficient, and tends to vary much more with
timings. Although I'm not entirely sure on the method (too tired to RTFS at
the moment), I'm prettur sure it doesn't do buffering. It does a cache-line
read, and waits around doing nothing until the cache line has arrived. This
means that a lot more of the time the FSB and memory busses are idle waiting
for the RAM to get ready, which gives you a much lower memory bandwidth (and
also obviously one much more dependent on memory latencies).

It's been a while since I ran memtest86, but I do remember that with
high-latency PC2100 (limited to a 133MHz FSB here) I was getting in the
order of 600MBytes/sec, and with tight-timing cas2 memory (actually
el-cheapo PC3200, but even el-cheapo PC3200 can run very tightly at 133MHz)
this increased to about 800MBytes/sec. So doing 1345MBytes/sec at 8-4-4-2.5
probably isn't unreasonable.

As for your sticks, I would be somewhat surprised if they would do cas2 at
200MHz. Most noname stuff doesn't and hasn't been designed to. It's easy to
find out what your memory is rated at. Fir up CPU-Z, go to the about page,
and click on "Memory SPD". Open up the resulting file and look at the
"Timings table" section. This will tell you what your memory is supposed to
be able to do. Mine looks like:

Timings table
Frequency (MHz) 133 166 200
CAS# 2.0 2.5 3.0
RAS# to CAS# delay 3 3 4
RAS# Precharge 3 3 4
TRAS# 6 7 8

If you're lucky, you can run faster than this. The ram above does 2-2-2
timings (at 133MHz). I know there's another tool somewhere that actually
dumps the raw SPD data which you can interpret yourself with a bit of
documentation, or you can post it here and I can interpret it for you.

Given that the noname modules are having trouble (as shown by needing 2.9V)
at 2.5-4-4 timings, I'm guessing that it's actually cas3 RAM with timings
similar to above. I'd be most surprised if they were any WORSE than the
above ram though So if you RAM is actually 3.0-4-4, and it runs fine at
3.0-4-4, you'd have a hard job convincing anyone that a refund is due. But
if it has been advertised as cas2.5 then I think it'd be fairly safe to say
they're refund material, given they don't work (at stock voltage) at cas2.5.

[...]
Experimenting with speeds (not running all tests) I also got some
confusing results with Memtest, all memory timings 8-4-4-2,5 by SPD
(most of the times equals Aggressive and Optimal): 11x200 gives
1345 MB/s
11,5x200 gives 1343 MB/s
12x200 gives 1342 MB/s
Slower with much higher CPU speed?

I think a 0.2% variation is well within the margin of error in the values
reported ... There is some very subtle interaction between CPU speed and
memory speed, but I wouldn't think it's evident in a pure memory blatting
test; you usually need carefully constructed, highly-optimised loops to get
a performance drop with a multiplier increase.

[...]

--
Michael Brown
www.emboss.co.nz : OOS/RSI software and more
Add michael@ to emboss.co.nz - My inbox is always open

Michael, Wes,

Thanks a lot! Will do some more experiments after studying your answers
and get back monday. (I have some extra work to do this weekend).

Boyd

Hi,

After Michaels and Wes comments I guess I have to concentrate on memory
first.
In dual mode the two 512MB PC3200 sticks simply don't run at 200MHz.

Michael wrote:

1. Im confused by the 'FSB/multiplier/running sync' relation.
What does "running CPU and memory in sync" mean for a mobile? The
BIOS has an 'auto' option, on by default. Does the Mobile 2600+
have a
'best' multiplier?

This just means that the memory bus and the FSB are run in sync.
Running
out of sync (say DDR333 and a 200MHz FSB) gives a reasonable
performance
hit.

OK, checked and indeed, both 5/4 and 5/6 ratio lowered memory speed to
about single channel performance with one stick :-(

BTW, when I said the memory timimgs were by SPD, I only meant CAS.
The FSB/RAM ratio was on auto always.

Michael wrote:

Fir up CPU-Z, go to the about page, and click on "Memory SPD". Open
up
the resulting file and look at the "Timings table" section. This will
tell you what your memory is supposed to be able to do.

Both sticks have two 200MHz entries: 8-4-4-3 and 8-4-4-2,5 and nothing
else (no 133 or 166MHz entries). Manufacturer and serial number fields
both only have a load of F's.

This is what the NF7-S does with the SPD and auto setting, all at 11x
multiplier:

FSB Ratio RAM
===============
167 auto 167
167 SPD 167
181 auto 181
181 SPD 217 - CAS becomes 9-4-4-3 (8-4-4-2,5 with all other
settings)
190 auto 190
190 SPD 190
200 auto 200
200 SPD 200
210 auto 210
210 SPD 210

BTW, the 181/217 setting does not clear Memtest at 1,725 Vcore and 2.6
volt memory.
Anyway, it also seems something goes wrong with the SPD setting. Should
not the RAM setting by SPD default to 200MHz, since that's the only
entry?

Wes wrote:

Most PC3200 rams spd data sets the memory bus to 166MHz, not 200, for
compatability reasons. If you want 200MHz memory bus, you have to
manually set it. I don't know how the NF7 handles this so you should
check to see if you really are running the memory bus at 200MHz. But,
the 2.9v for the ram is not good at all. It's likely to burn out
quick
at that voltage. I personally wouldn't go over 2.7v on a 2.5v module.
there used to be a problem with a lot of bioses misinterpreting the
PSD
data too and setting wrong values. I think this was one of the
reasons
ram modules SPD data now defaults to 166Mhz. Too many returns of
perfectly good ram not set properly by the bios defaults.

Well it turns out all my FSB 200MHz experiments did run at RAM 200MHz
and most were at 2,6 volts minimum (NF7-S does not go lower). BTW, it
did not run hot at 2,9 volt.
As long as RAM is at 200MHz, should not PC3200 memory simply work?

It does not at 10x200 (RAM 200MHz) SPD mem timings 8-4-4-3, which is
the normal CPU speed of 2000MHz: 45004 errors in Memtest nr. 5
And not even at 9x200 with still 36066 errors in Memtest nr. 5.
And Vcore is at 1,725 (about 1,68 reported), Vdimm at 2,6 (about right
reported).
So the memory does not even run at 200MHz with underclocked CPU and
overvolted memory and CPU (normal = 1,45 volt).

Could it be both 512 MB sticks differ to much? Both have the same
chips, but I noticed the prints differ a bit.
While both together clearly fail at 11,5x200 (dual mode), they did
better on their own in single mode: 11,5x210 (= RAM 210 MHz also) at
2,7 volt without any errors in Memtest.

Since I'd like to overclock the CPU a bit (that's why I got the
Mobile), I went down from 11x200 and now got 11x194 (= 194MHz RAM)
without erros in Memtest,
SPD mem timings being 8-3-3-2,5

A bit disapointing, I was thinking 200MHz memory should even have a bit
of reserve above 200 MHz. Do I have grounds to return the memory?
Should it do 200MHz also in dual mode?
Would faster memory (PC3500 or more) help a lot with working on large
image and audio files?

Thanks, Boyd

Hi,

After working some hours work on it, I did post a reply here over 9
hours ago, got the confirmation etc, but it seems it got lost ...
E-mailed Google about this.

Anyway, I concentrated on the memory, after reading the comments by
Michael and Wes, the important issues being:

When I said 'memory timings by SPD' I meant CAS only. The FSB/RAM ratio
was on AUTO always.

CPU-Z tells me the memory has only 2 entries for SPD, both for 200MHz:
one 8-4-4-2.5 and one 8-4-4-3 (?!)
Checking how the NF7-S handles this, I found out both the settings AUTO
and By SPD for FSB/RAM ratio did not result in setting the RAM speed to
200Mhz. RAM speed just always follows FSB, with one strange exception:
181 FSB with 217 RAM and CAS changing to a 9-something CAS (I forgot).

I checked both 512MB PC3200 noname sticks alone and found they both run
11,5x210 (RAM being 210Mhz also) at 1,6 Vcore and 2,7 Vdimm. Memtest
was OK. Did not try higher.
But together in dual mode I can get them over 194Mhz FSB/RAM, with
1,725 Vcore and 2,6 Vdimm. At 11x multiplier this means underclocking
the CPU and overvolting Memory (sticker says 2,5 volt, but NF7 can't go
lower) and CPU (1,725 = 1,68 reported, 1,45 is normal for this mobile).
11x194 did test OK in Prime95 for 8 hours, I'm now testing 12x194,
seems to go well.

So I wonder what PC3200 means in this case and if both sticks (same
CPU-Z data and chips, but I noticed the prints differ) don't work well
together in dual mode.

Since I think PC3200 should have a bit of 'reserve', I'm a bit
disapointed it does not even hit 200MHz. I see lots of people doing
210Mhz and more with 2 sticks PC3200 in dual mode. Is that 'only' due
to better matched memory? Should all PC3200 memory do 200MHz in dual
mode (these two 512MB sticks both do more on their own in single mode).


And the main question: working on large image and audio files, could I
benefit from faster memory, like PC3500 or more? My mobile 2600+ can
run much faster before hitting temperature issues. Do I gain overall
perfomance with faster memory or does the NF7-S become the limiting
factor?

Thanks,
Boyd Noorda

Again the message got lost, now trying to insert it as reply to the
first message ...

Hi,

After working some hours work on it, I did post a reply here over 9
hours ago, got the confirmation etc, but it seems it got lost ...
E-mailed Google about this.

Anyway, I concentrated on the memory, after reading the comments by
Michael and Wes, the important issues being:

When I said 'memory timings by SPD' I meant CAS only. The FSB/RAM ratio
was on AUTO always.

CPU-Z tells me the memory has only 2 entries for SPD, both for 200MHz:
one 8-4-4-2.5 and one 8-4-4-3 (?!)
Checking how the NF7-S handles this, I found out both the settings AUTO
and By SPD for FSB/RAM ratio did not result in setting the RAM speed to
200Mhz. RAM speed just always follows FSB, with one strange exception:
181 FSB with 217 RAM and CAS changing to a 9-something CAS (I forgot).

I checked both 512MB PC3200 noname sticks alone and found they both run
11,5x210 (RAM being 210Mhz also) at 1,6 Vcore and 2,7 Vdimm. Memtest
was OK. Did not try higher.
But together in dual mode I can get them over 194Mhz FSB/RAM, with
1,725 Vcore and 2,6 Vdimm. At 11x multiplier this means underclocking
the CPU and overvolting Memory (sticker says 2,5 volt, but NF7 can't go
lower) and CPU (1,725 = 1,68 reported, 1,45 is normal for this mobile).
11x194 did test OK in Prime95 for 8 hours, I'm now testing 12x194,
seems to go well.

So I wonder what PC3200 means in this case and if both sticks (same
CPU-Z data and chips, but I noticed the prints differ) don't work well
together in dual mode.

Since I think PC3200 should have a bit of 'reserve', I'm a bit
disapointed it does not even hit 200MHz. I see lots of people doing
210Mhz and more with 2 sticks PC3200 in dual mode. Is that 'only' due
to better matched memory? Should all PC3200 memory do 200MHz in dual
mode (these two 512MB sticks both do more on their own in single mode).


And the main question: working on large image and audio files, could I
benefit from faster memory, like PC3500 or more? My mobile 2600+ can
run much faster before hitting temperature issues. Do I gain overall
perfomance with faster memory or does the NF7-S become the limiting
factor?

Thanks,
Boyd Noorda