Optimal memory for P4C800-E Deluxe

[_R]

I'll be picking up a P4C800-E soon. I'd like to optimize the $ spent
on memory--I'm figuring 1GB initially, to be expanded to 2GB later.
I'll be running music software (samplers are RAM-hungry) and software
development tools.

Is dual channel worthwhile? DDR400 vs slower??? Preferred brands?

Any comments are welcome.

 

[KC Computers]

I'll be picking up a P4C800-E soon. I'd like to optimize the $ spent
on memory--I'm figuring 1GB initially, to be expanded to 2GB later.
I'll be running music software (samplers are RAM-hungry) and software
development tools.
Is dual channel worthwhile? DDR400 vs slower??? Preferred brands?
Any comments are welcome.

We are a dealer and have had good luck with Geil, Kingston and Corsair
PC3200 modules with that motherboard. Don't be fooled by memory
sold as dual channel. Any two same-sized modules (ideally the same make and
model)
will run in dual channel mode. Dual channel mode can boost performance
a couple percent.

E-mail or call us if you have any more questions.

 

[DanO]

Note: If you go Dual Channel now for 1GB (i.e. 2 x 512) when you later try
to go to 4 dimms (4 x 256) for 2 Gig, certain memory issues can arise that
will result in a chnage in performance of the memory system. There was a
lot of tests posted way back when the 875 chipset first hit the streets
showing this impact. Not sure if it would be a problem for you, but you may
want to think about it. It might be something like PAT is disabled with 4
dimms (IIRC.)

 

[Robert Hancock]

Note: If you go Dual Channel now for 1GB (i.e. 2 x 512) when you later try
to go to 4 dimms (4 x 256) for 2 Gig, certain memory issues can arise that
will result in a chnage in performance of the memory system. There was a
lot of tests posted way back when the 875 chipset first hit the streets
showing this impact. Not sure if it would be a problem for you, but you may
want to think about it. It might be something like PAT is disabled with 4
dimms (IIRC.)

Most likely what happens is that at DDR400 with 4 DIMMs it uses 2T
command rate instead of 1T, which reduces the memory bandwidth (I've
heard the figure of 20%).

 

[Paul]

Most likely what happens is that at DDR400 with 4 DIMMs it uses 2T
command rate instead of 1T, which reduces the memory bandwidth (I've
heard the figure of 20%).

The datasheet for the 875 mentions neither a register for PAT,
nor a register that controls command rate (1T/2T). So, what
I decided to do today, is run some tests on my P4C800-E, and
see what impact four sticks has on bandwidth.

The first test is two sticks in dual channel, with FSB and MEM
operating sync. I then added two more sticks (all four matching
Ballistix PC3200), running the FSB and MEM at the same standard
speeds. For the third test, I bumped up the clock by 5MHz,
but still keeping the FSB and MEM operating sync. The third test
was chosen, due to the known issue where PAT is disabled if four
sticks are used, and the clock is not exactly 200MHz. The PAT
condition was verified with both CTIAW and CPUZ.

2x512MB DS 2-2-2-6 DDR400 FSB800 P4C 2.8Ghz
Main memory speed (MB/s): Read=3609.4, Write=2410.7 (measured by cachemem)
Read=2955MB/sec (measured by memtest86+ 1.4)
Both CTIAW and CPUZ claim memory is actually set to 2-2-2-5
PAT "(1) fully enabled"

CPU-Z version 1.22 Latency Test
stride->4 8 16 32 64 128 256 512
size(Kb)
| 1 2 2 2 2 2 2 2 2
v 2 2 2 2 2 2 2 2 2
4 2 2 2 2 2 2 2 2
8 2 2 2 3 2 3 3 3
16 3 3 7 15 20 19 18 19
32 3 4 8 17 19 18 19 19
64 3 4 9 18 18 18 19 19
128 3 4 9 18 18 18 18 19
256 3 4 9 18 19 19 20 21
512 3 4 9 18 20 20 21 25
1024 4 8 15 28 53 180 172 210
2048 5 9 15 29 55 178 176 187
4096 5 8 16 30 58 215 173 190
8192 5 8 15 28 58 216 176 190
16384 5 8 15 28 58 216 170 189
32768 4 9 16 29 52 183 174 186

4x512MB DS 2-2-2-6 DDR400 FSB800 P4C 2.8Ghz
Main memory speed (MB/s): Read=3307.7, Write=1858.5 (measured by cachemem)
Read=2733MB/sec (measured by memtest86+ 1.4)
Both CTIAW and CPUZ claim memory is set to 2-2-2-6 (agrees with BIOS)
PAT "(1) fully enabled"

CPU-Z version 1.22 Latency Test
stride->4 8 16 32 64 128 256 512
size(Kb)
| 1 2 2 2 2 2 2 2 2
v 2 2 2 2 2 2 2 2 2
4 2 2 2 2 2 2 2 2
8 2 2 2 4 2 2 3 3
16 3 3 7 15 18 19 19 19
32 3 4 8 17 18 18 19 19
64 3 4 9 18 18 18 20 19
128 3 4 9 18 18 18 18 19
256 3 4 9 18 19 19 19 21
512 3 4 9 18 20 23 21 25
1024 5 9 16 30 58 209 208 239
2048 5 9 17 32 63 210 203 214
4096 5 9 17 33 64 239 201 216
8192 5 9 16 31 58 208 203 214
16384 5 9 17 33 63 210 201 215
32768 5 10 17 31 58 210 201 239

4x512MB DS 2-2-2-6 DDR410 FSB820 P4C 2.87Ghz (clk=205MHz)
Main memory speed (MB/s): Read=3230.7, Write=1890.8 (measured by cachemem)
Read=2802MB/sec (measured by memtest86+ 1.4)
Both CTIAW and CPUZ claim memory is set to 2-2-2-6 (agrees with BIOS)
PAT "(0) reserved" - means, AFAIK, that PAT is disabled.

CPU-Z version 1.22 Latency Test
stride->4 8 16 32 64 128 256 512
size(Kb)
| 1 2 2 2 2 2 2 2 2
v 2 2 2 2 2 2 2 2 2
4 2 2 2 2 2 2 2 2
8 2 2 2 3 3 2 2 3
16 3 3 7 15 18 19 19 19
32 3 4 9 17 18 19 19 19
64 3 4 9 18 18 18 19 19
128 3 4 9 18 18 18 19 18
256 3 4 9 18 19 19 19 19
512 4 4 9 18 20 23 22 26
1024 5 9 17 32 62 223 222 229
2048 5 9 17 32 62 223 217 229
4096 5 9 17 32 61 224 218 229
8192 5 9 17 32 62 222 215 229
16384 5 9 17 34 65 223 216 228
32768 5 10 17 33 61 222 245 255

What is interesting to me, is the most significant
effect seems to be the transition from 2 sticks to
4 sticks. In both cases, PAT is supposed to be enabled,
at least according to CPUZ and CTIAW. But the second case
is slower. Now, I did notice that even though the BIOS
memory timings were set the same for all tests (2-2-2-6
in the BIOS), the last param was actually 5 when using two
sticks, and was set to 6 when using four sticks. (The BIOS
must have a subtle bug in it ? Or is this evidence of
something ?)

Now, the other thing that is puzzling, is
the last test, with the slightly elevated
clock, has much "smoother" columns of numbers.
How can disabling PAT be causing such a phenomenon ?

Another puzzler, is the measurement of bandwidth. Between the
second and third test, memtest86+ finds a bamdwidth increase
which is exactly equal to the clock speedup factor. Cachemem
got a drop on read BW and an increase in write BW. And, I ran
Cachemem multiple times, and the deviation on the measured
bandwidth numbers is down to the last digit, so the drop and
the increase are reproducible.

All I can conclude, for the benefit of the OP, is that
using four sticks is costing 7.5% memory bandwidth. And
when you raise the clock, and PAT is disabled, what happens
there really depends on which tool is measuring the bandwidth.
It looks like if you raise the clock high enough, you'll get
the bandwidth back, so that is always an option.

This is one case, where a logic analyser would be called for,
rather than trying to conclude anything from software testing.

Paul

 

[_R]

All I can conclude, for the benefit of the OP, is that
using four sticks is costing 7.5% memory bandwidth. And
when you raise the clock, and PAT is disabled, what happens
there really depends on which tool is measuring the bandwidth.
It looks like if you raise the clock high enough, you'll get
the bandwidth back, so that is always an option.

This is one case, where a logic analyser would be called for,
rather than trying to conclude anything from software testing.

Paul

That's a lot of interesting info, Paul. I'm sure it will make sense
after I have a chance to ramp up on the current gen of motherboards.
I have logged it for later study.

For now, though, I'm trying to figure out the best way to spend $ on
RAM for a P4C800-E Deluxe (encouraging to see you're using this
board).

I try to stay a step back from the leading edge ($$$) so I'm looking
for whatver will yield the best performance at reasonable cost.

I do plan on either going with 2Gigs right away, or maybe 1G now
+ 1G later. (Depends again on price)

1: Is it worthwhile to buy DDR400, or is this overkill?
2: Is dual channel worthwhile? Any caveats here?
3: If I go with dual channel should I try to stay with two sticks?
4: Preferred brands?

Does the following make sense, or are there better ways to allocate
the budget?

http://www.zipzoomfly.com/jsp/ProductDetail.jsp?ProductCode=80097-41

 

[_R]

All I can conclude, for the benefit of the OP, is that
using four sticks is costing 7.5% memory bandwidth. And
when you raise the clock, and PAT is disabled, what happens
there really depends on which tool is measuring the bandwidth.
It looks like if you raise the clock high enough, you'll get
the bandwidth back, so that is always an option.

Now that I think about it, I'd much rather go with something less
expensive. Maybe:

http://www.zipzoomfly.com/jsp/ProductDetail.jsp?ProductCode=80098-21

Would using 4 sticks (2 GB) negate any advantage in dual channel?
Going to 2GB this way would be half the price of the 2GB sticks I was
looking at.

I figure it's going to be obsolete in a few months anyway, so there's
no sense in getting too caught up in getting the last nanosecond.
OTOH, I don't want to pinch pennies if a serious perfomance benefit is
possible with close to the same budget.

 

[Paul]

Now that I think about it, I'd much rather go with something less
expensive. Maybe:

http://www.zipzoomfly.com/jsp/ProductDetail.jsp?ProductCode=80098-21

Would using 4 sticks (2 GB) negate any advantage in dual channel?
Going to 2GB this way would be half the price of the 2GB sticks I was
looking at.

I figure it's going to be obsolete in a few months anyway, so there's
no sense in getting too caught up in getting the last nanosecond.
OTOH, I don't want to pinch pennies if a serious perfomance benefit is
possible with close to the same budget.

Performance is asymptotic in all aspects of computing. A 3GHz
processor is $200 and a 3.8GHz is $1000. Is the extra 800MHz
worth $800 - it must be, as somebody is paying that price as
we speak.

Memory is the same way. You can double or trip the price, to get
an extra bit of performance. Is it worth the money ? How many
people (other than the SETI or folding crowd) run their computers
flat out all the time ? Most computers are idle most of the time.
We are talking about optimizing the performance of bursts of
computing that are so short, the computing is done before you blink.

Tomshardware has an article, that compares various options for
RAM (in particular, how little difference moving from CAS3 to CAS2.5
or CAS2 makes). Maybe after reading this, you'll have second
thoughts about the utility of CAS.

http://www6.tomshardware.com/motherboard/20040119/index.html

An overclocker builds a fast system, but seldom cares about the
quantity of memory present in the system. You, on the other hand,
want more memory than can be usefully overclocked, so you aren't
in the pure overclocker camp.

With two sticks running dual channel, you can do:

1a) FSB800, DDR400, CAS3 two sticks (bog standard, cheap solution)
1b) FSB1000, DDR400, CAS3 5:4 ratio (overclock CPU 25%, same memory)
1c) FSB1000, DDR500, CAS3 1:1 ratio (overclock CPU and use $$$ memory)
1d) FSB1000, DDR500, CAS2 1:1 ratio (OCZ VX memory 3.3V overvolt $$$$)

Those options will probably change the price of the memory by a factor
of three. You'll get more memory bandwidth, but will you notice a
difference ?

Going from option 1a to 1b, you use the same cheap memory. You
overclock the processor by 25%, and on compute bound applications
get 25% more benefit. If you improve the memory by 25%, you get 8%
more performance, and it costs mondo bucks to get that extra 8%.

BTW: Your Zipzoomfly memory was $100 and the OCZ VX is $275. You
cannot actually use the OCZ VX, because I doubt the P4C800-E can
hack operating the Vdimm at 3.3V.

With four sticks running dual channel, you can do:

2a) FSB800, DDR400, CAS3 four sticks (bog standard, cheap, -7.5% BW)
2b) FSB1000, DDR400, CAS3 5:4 ratio (overclock CPU 25%, same memory)
2c) FSB920, DDR460, 1:1 ratio (four sticks might run from DDR440 to DDR460)

The last option might profitably use PC3700 memory. CAS will help
only a little bit, so whether it is PC4000 CAS2 or PC3700 CAS3 likely
won't matter, but the price will matter.

Comparing case 2a to 1a, you lose 7.5% memory bandwidth, which is
2.5% application performance. A small loss, that you can recoup by
doing 2b. In 2b, all you have to do, is set the memory ratio to
5:4, bump up the FSB, and you are done. It uses the same cheap memory
as 2a.

I think two or four matched sticks of bargain DDR400 CAS3 memory will
work just fine. If you are feeling cheated, just bump up the FSB
and feel better.

Now, you might ask me, why did I buy 4 sticks of Ballistix PC3200
CAS2 I originally started with some Azenram pulled from another
computer, which was DDR400 CAS3. The RAM died after a little more
than a year of use. My A7N8X-E is very picky about RAM, and really
needs CAS2 for stability. I decided to buy memory that could be moved
from system to system, without worry or compromise. So far the
Ballistix has been flawless. I could probably have bought CAS2 for
the A7N8X-E, and CAS3 for the P4C800-E, but figured, what the hell.
That is why I was able to do those tests for you with four sticks
of 512MB - normally I only use 2x512MB.

Paul

 

[DanO]

Paul,
Wow! Nice tests and write-up. Now we all know for sure the impact of 2
vs. 4 sticks of ram.

Thanks for all of the hard work!

DanO

 

[_R]

....
Tomshardware has an article, that compares various options for
RAM (in particular, how little difference moving from CAS3 to CAS2.5
or CAS2 makes). Maybe after reading this, you'll have second
thoughts about the utility of CAS.

http://www6.tomshardware.com/motherboard/20040119/index.html
......
I think two or four matched sticks of bargain DDR400 CAS3 memory will
work just fine. If you are feeling cheated, just bump up the FSB
and feel better.

Paul

(Lots of great info snipped above)

Hi Paul, I don't keep up with overclocking. Programming is tough
enough<g>. I've read thru your post a few times, though, and I
think I'm assimilating some stuff. Thanks much for taking your time
to do all the tests!