P4: 865xx vx 875 for memory speed

  • Thread starter Thread starter H.W. Stockman
  • Start date Start date
H

H.W. Stockman

I've seen arguments that the 875 chipset should produce faster memory access,
but nothing quantitative (other than and "overall" performance boost of 3-5% --
I don't know what that includes as a mix of CPU/memory stress).

I've looking at performance for a very memory-intensive application, without
threading, similar to the stream benchmark (the one in SiSoft Sandra is an OK
analogy).

Anyone?
 
H.W. Stockman said:
I've seen arguments that the 875 chipset should produce faster memory access,
but nothing quantitative (other than and "overall" performance boost of 3-5% --
I don't know what that includes as a mix of CPU/memory stress).

I've looking at performance for a very memory-intensive application, without
threading, similar to the stream benchmark (the one in SiSoft Sandra is an OK
analogy).
Click to expand...
There are plenty of comparisons, for instance here,
http://www.tech-report.com/reviews/2003q2/intel-865/index.x?pg=5 (both
sandra memory bench and linpack numbers).
(some i865 boards also allow to enable PAT (except it has to be named
differently in the bios), those will be just as fast as the i875 boards
(chipsets are 100% identical anyway, intel just disables PAT so it can
sell the i875 for higher price to the enthusiast crowd). I don't know if
this is still possible though or if intel prevented that in later i865
steppings.)

Roland
 
"H.W. said:
I've seen arguments that the 875 chipset should produce faster memory access,
but nothing quantitative (other than and "overall" performance boost of 3-5% --
I don't know what that includes as a mix of CPU/memory stress).

I've looking at performance for a very memory-intensive application, without
threading, similar to the stream benchmark (the one in SiSoft Sandra is an OK
analogy).

Anyone?
Click to expand...

The 875 is the king of the heap on the Intel side.
(Based on reading this one article :-)))
http://www.anandtech.com/mb/showdoc.html?i=1930&p=11

But the FX-51 blows it away, due to the memory controller
being on the processor and not the Northbridge. Then, the
question is, how does the rest of the processor compare
to the Intel side ?
http://www.anandtech.com/mb/showdoc.html?i=1917&p=9

Memory performance also has a dependency on stride.
The memory patterns in your real world application might
not match how Sandra does things. You may want to look
at the Intel memory guide for the 875/865, to see what
some of the issues might be.

ftp://download.intel.com/design/chipsets/applnots/25273001.pdf

You might want to visit the forums at Abxzone, and
see what memory configurations give the best performance
with four sticks of memory. One guy did DDR440 with four
sticks of memory on an 875/865 board, and by now, someone
will have surpassed that (that was months ago).

At my previous employer, local vendors would provide a
loaner, so that performance could be evaluated with the
exact application the customer has in mind. You cannot
expect to make fine distinctions between designs based
on a few web reviews. It is either look for loaner
machines, or offer to distribute a benchmark application
from your web page, and have hobbyists run a representative
part of your benchmark on their machines.

We could rate motherboards in terms of how many "Stockmans"
it can do per second :-)

HTH,
Paul
 
Paul said:
on a few web reviews. It is either look for loaner
machines, or offer to distribute a benchmark application
from your web page, and have hobbyists run a representative
part of your benchmark on their machines.

We could rate motherboards in terms of how many "Stockmans"
it can do per second :-)
Click to expand...


I used to distribute a benchmark (an abbreviated version of the actual program I
run). However, I found the process of collecting results from people
surprisingly tedious and never very efficient. I found that performance on
McCalpin's stream correlated well with the performance on my specific benchmark.

I cringe when I see anandtech using "ScienceMark." Enough said.
ftp://users.viawest.net/public_html/bench/daxpy_p4&Ath.gif
 
"H.W. said:
I used to distribute a benchmark (an abbreviated version of the
actual program I run). However, I found the process of collecting
results from people surprisingly tedious and never very efficient.
I found that performance on McCalpin's stream correlated well with
the performance on my specific benchmark.

I cringe when I see anandtech using "ScienceMark." Enough said.
ftp://users.viawest.net/public_html/bench/daxpy_p4&Ath.gif
Click to expand...

http://www.cs.virginia.edu/stream/standard/Bandwidth.html
ncpus COPY SCALE ADD TRIAD
Asus_P4T533-P4-2800 1 1832.6 1736.2 2249.2 2249.2
Intel_875_P4_2.80 1 2391.0 2376.0 2742.0 2715.0
AMD_Opteron_248 (2.2GHz/PC3200) 1 2819.8 2815.4 3146.0 3144.0

I don't know if you are looking for a solution that will scale
in any sense, but maybe a P4C800-E Deluxe would be a better
platform for your experiments, because there are more adjustments
and more opportunities for overclocking. (You will need a cheap
video card.) You could take a 2.8GHz P4 and overclock it to 3.6Ghz
on air, and that would allow a higher memory bus speed at the same
time. You could try out some PC4000 memory (2x512MB in dual channel
config). If your program needs more memory than this, then the 5:4
CPU:FSB ratio and some 4x512MB Winbond BH5 based memory from Ebay
is another route to take (you cannot buy BH5 low latency memory
at retail any more - the BH6 doesn't overclock quite as nicely, but
can be purchased from Mushkin).

http://www.cpudatabase.com/CPUdb (Overclocking database)

I still think if you went over to Abxzone.com and posted a challenge
based on some copies of "stream", you could get the people over
there involved. I think they are getting bored these days, waiting
for the BTX motherboards to come in, and many of them are torturing
mobile Athlon processors and could use a break :-) They also
get their thrills running OCZ 2.5V memory at 3.3V (not recommended!).

Paul
 
I might reconsider the Opteron. Early benchmarks didn't show the greatest
memory performance...
 
Paul said:
I still think if you went over to Abxzone.com and posted a challenge
based on some copies of "stream", you could get the people over
there involved. I think they are getting bored these days, waiting
for the BTX motherboards to come in, and many of them are torturing
mobile Athlon processors and could use a break :-) They also
get their thrills running OCZ 2.5V memory at 3.3V (not recommended!).
Click to expand...


Well, I have done just that, and also posted requests on this newsgroup and
comp.sys.intel. So far, the response has been: zip. Over at abXzone, they
seem more interested in running benchmarks that confirm their choices of
architechture.
 
"H.W. said:
Well, I have done just that, and also posted requests on this newsgroup and
comp.sys.intel. So far, the response has been: zip. Over at abXzone, they
seem more interested in running benchmarks that confirm their choices of
architechture.
Click to expand...

Well, I ran your benchmark, more out of curiosity than anything
else. I got 2.73 MUPs on my P4 1.8GHz (SDRAM based), which is
hardly the class of machines you are interested in.

I also noticed the two threads you posted in, over on Abxzone.
Not a very good response on their part. And it will probably
take too long, to figure out how to motivate any of them
to help you. (I don't have an account over there, but I
do use there search engine a lot.)

One of my problems with that forum, is they don't know what a
thread is for. This makes their search engine rather useless
at times, because if you get a hit in a thread with 3500 posts
in it, you are unlikely to read all the posts, to see if one
of the posts is relevant or not. (Apparently thread hijacking
is accepted, if it interests them.)

About the only other thing you could look for, is to find some
Sandra unbuffered memory bandwidth pictures, posted for the
various machine configurations you are interested in. I don't
know how you feel about Sandra, as some of Sandra's other
benchmarks aren't that well constructed (give weird results).

Maybe you'll have to tap into a mailing list somewhere, where
people are involved in benchmarking. I think you've tried some
other USENET news groups in the past, so that is an option
as well.

There are some adults having fun out there:
http://www.csm.ornl.gov/~dunigan/opteron/#mem

HTH,
Paul
 
Back
Top