Memory performance with high FSBs (for P4)

"Frank Weston" <frank RemoveThis @weston-american.com> wrote in message
news:xdmdndWWxc5SUuOiU-KYgw@comcast.com...
 >
 > "Andy Cunningham" <asc99c RemoveThis @hotmail.com> wrote
 >
 > snip....
  > > Anyone seen benchmarks or have their own experience? Also, how much
  > > difference do the faster timings on lower clocked modules make?
  > >
 >
 > My experience has been that memory timings don't make a whole lot of
 > difference with my P4 2.8. On most bench marks, the difference between
 > 2,2,3,5 and 2.5,4,4,7 is pretty much within the range of error, and in all
 > cases less than 1%. For example, best performance with Pifast and 2,2,3,5
 > is 60.22 seconds, and at 2.5,4,4,7, 60.69 seconds. These differences are
 > within the range of results that I get running in either mode.
 >
 > Memory bandwidth is decreased by running at dividers less than 1:1, but
 > being able to run at an increased FSB makes up for some of this decrease.
 > At 1:1, with my PC3700 memory, I can run stably at about 230 FSB which
gives
 > a MBW of about 5100 + or - 100 according to Sandra. At 4:5 I can run at
243
 > FSB which tests at about 5100 as well, and if I set the divider to 1:1 at
 > 243, I can score over 5400 on MBW in Sandra.....but the system isn't
stable
 > enough to run 3DMK2003.
 >
 > Bottom line for me seems to be that the timings aren't as important as
 > running at a higher FSB with a divider of less than 1:1. Of course, I
would
 > probably get the best performance by running PC4000 at 1:1 and 243 FSB,
but
 > I don't have $400 and change that I want to spend on getting a little more
 > memory bandwidth when what I have now is pretty phenomenal by recent
 > standards. Anybody want a good deal on some lightly used PC3700?

The reviews and owner benchmarks are all over the place on this question,
and every system is different.

At Tom's Hardware there's a review that says it's a good deal faster to run
a 2.6g 1:1 with PC4000 at 250 fsb for 3.25 ghz than it is to run a stock 3.2
1:1 with PC3200. Amost a no-brainer, because they didn't include other
cpu/ram ratios in the mix.

One poster on the OC forum at Anand's went through a fairly elaborate series
of benchmarks and concluded that running CAS 3 PC4000 1:1 at 250 fsb was
slower than running CAS 2 PC3200 5:4 at 200, even though the cpu speed
stayed the same. Very interesting.

As a starting point, you need to consider the cpu and multiplier and what
overclock you're shooting for. Whether you have a 2.4 or a 3.2, the top end
for a P4C is somewhere in the low-to-mid 3ghz range, say 3.4-3.6 on air
cooling.

With your 2.8 running 1:1 at 230x14, you're getting a respectable 3220 and
5100+ in Sandra memory scores at 2.5,4,4,7.

With my 2.4 running 5:4 at 275x12=3300, my PC3700 is running at a slower 220
mhz with tighter timings of 2,3,2,6. My Sandra memory scores are 5800+

It's a tricky equation with too many variables for a one-size-fits-all
approach.

But no matter which path you choose or how much money you spend, the goal
should be the same, maximum cpu overclock while keeping the ram at optimum
mhz and timings.

Running high-latency PC4000 at 1:1 on a 2.4 would be a waste, IMHO. Running
low-latency PC2700 5:4 on a 3.2 would be equally silly. In between there are
lots of combinations that offer various tradeoffs between cpu and ram
performance -- and it's pretty much up to each of us to find the magic combo
to match our hardware and needs.<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: Memory performance with high FSBs (for P4) 

"Skid" <skid2.RemoveThis@comcast.NOSPAM.net> wrote
 >
 > With your 2.8 running 1:1 at 230x14, you're getting a respectable 3220 and
 > 5100+ in Sandra memory scores at 2.5,4,4,7.
 >
 > With my 2.4 running 5:4 at 275x12=3300, my PC3700 is running at a slower
220
 > mhz with tighter timings of 2,3,2,6. My Sandra memory scores are 5800+
 >

Interesting. I'm actually running at 2,2,3,5 at 230 FSB and still getting
less MBW than you are at 220. Both of us are using PC3700 memory, so there
must be some other differences in our setup. I'm running Win2k without
hyperthreading and two 256 Mb sticks (512 Mb) of memory. I'm guessing
you're running 1 Gb of memory and/or WinXP. Either that or we may be
running different versions of Sandra.<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: Memory performance with high FSBs (for P4) 

Well so far there doesn't seem to be any strong indication that the 1:1
ratio is a particularly great performance boost. How good is Geil memory?
I've got my eye on some PC3200 stuff which costs little more than generic.
What sort of speed is likely possible; PC100 RAM very often managed 120 MHz
or even 133 MHz, but is more recent memory built with less tolerance?

"Frank Weston" <frank RemoveThis @weston-american.com> wrote in message
news:9tydnUaukPgQiOKiU-KYgA@comcast.com...
 >
 > "Skid" <skid2 RemoveThis @comcast.NOSPAM.net> wrote
  > >
  > > With your 2.8 running 1:1 at 230x14, you're getting a respectable 3220
and
  > > 5100+ in Sandra memory scores at 2.5,4,4,7.
  > >
  > > With my 2.4 running 5:4 at 275x12=3300, my PC3700 is running at a slower
 > 220
  > > mhz with tighter timings of 2,3,2,6. My Sandra memory scores are 5800+
  > >
 >
 > Interesting. I'm actually running at 2,2,3,5 at 230 FSB and still getting
 > less MBW than you are at 220. Both of us are using PC3700 memory, so
there
 > must be some other differences in our setup. I'm running Win2k without
 > hyperthreading and two 256 Mb sticks (512 Mb) of memory. I'm guessing
 > you're running 1 Gb of memory and/or WinXP. Either that or we may be
 > running different versions of Sandra.
 >
 >
 >


---
Outgoing mail is certified Virus Free.
Checked by AVG anti-virus system (http://www.grisoft.com).
Version: 6.0.516 / Virus Database: 313 - Release Date: 01/09/2003<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: Memory performance with high FSBs (for P4) 

"Frank Weston" <frank.TakeThisOut@weston-american.com> wrote in message
news:9tydnUaukPgQiOKiU-KYgA@comcast.com...
 >
 > "Skid" <skid2.TakeThisOut@comcast.NOSPAM.net> wrote
  > >
  > > With your 2.8 running 1:1 at 230x14, you're getting a respectable 3220
and
  > > 5100+ in Sandra memory scores at 2.5,4,4,7.
  > >
  > > With my 2.4 running 5:4 at 275x12=3300, my PC3700 is running at a slower
 > 220
  > > mhz with tighter timings of 2,3,2,6. My Sandra memory scores are 5800+
  > >
 >
 > Interesting. I'm actually running at 2,2,3,5 at 230 FSB and still getting
 > less MBW than you are at 220. Both of us are using PC3700 memory, so
there
 > must be some other differences in our setup. I'm running Win2k without
 > hyperthreading and two 256 Mb sticks (512 Mb) of memory. I'm guessing
 > you're running 1 Gb of memory and/or WinXP. Either that or we may be
 > running different versions of Sandra.

Like I said, there are too many variables to declare any particular approach
the best.

But for the record, I am running Win XP Home, P4C with hyperthreading
enabled, Abit IC7 with PAT enabled, latest freeware version of Sandra Max,
with 2x256 Buffalo Tech PC3700 (Winbond BH-5 modules, same as Mushkin PC3500
but lots cheaper.) Sandra scores fluctuate a little, as you know, but from a
fresh boot with my usual NAV, ZoneAlarm and KillPopup running at startup,
mine are usually 5840-50.<!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: Memory performance with high FSBs (for P4) 

"Andy Cunningham" <asc99c DeleteThis @hotmail.com> wrote in message
news:blndaa$69q$1$830fa17d@news.demon.co.uk...
 > Well so far there doesn't seem to be any strong indication that the 1:1
 > ratio is a particularly great performance boost. How good is Geil memory?
 > I've got my eye on some PC3200 stuff which costs little more than generic.
 > What sort of speed is likely possible; PC100 RAM very often managed 120
MHz
 > or even 133 MHz, but is more recent memory built with less tolerance?

DDR in dual channel mode should be harder to overclock, but some of it does
really well. There is no recognized certification standard above PC3200, so
most PC3500-4000 is actually overclocked PC3200 with relaxed timings to get
it stable at higher speeds.

I have no personal experience with Geil, but it is one of those love it or
hate it brands with as many supporters as detractors. They do roll their own
modules, like Crucial/Micron, where most other brands buy chips from various
makers and slap them on a pcb with a heatspreader to hide their rating and
origin.

OCZ PC3700 Gold, for instance, uses 5ns Samsung actually rated at PC3200. My
Buffalo Tech PC3700 uses the same Winbond BH-5 modules as Mushkin PC3500
Black Label Level II, but costs a lot less.

Memory is the key to a good overclock on an Intel Springdale or Canterwood
board, and buying it can be very confusing. Speed labels and price tags are
not the best indicators of success.

There are some good articles and comparisons in the memory section at
<a style='text-decoration: underline;' href="http://www.anandtech.com," target="_blank">www.anandtech.com,</a> and a link to lots of memory reviews on the right side of
the page at <a style='text-decoration: underline;' href="http://www.amdboard.com." target="_blank">www.amdboard.com.</a><!-- ~MESSAGE_AFTER~ -->
 >> Stay informed about: Memory performance with high FSBs (for P4)