CCL:RIMM vs SDRAM

[Joshua Halpern]

The problem might be in the northbridge.  The fact that
the difference appears memory access driven makes this
more likely.  If you look at the results from the 
VIA KT266A revision, you can see the cited 20% 
difference in performance disappear

http://www6.tomshardware.com/mainboard/01q3/010902/kt266a-03.html

The review starts at
http://www6.tomshardware.com/mainboard/01q3/010902/


Without knowing which motherboard Scott was using, it
is hard to be definite, but the superiority of the
KT266A over the AMD 760 and the SIS 735 chip sets
(as well as blowing away the KT266) makes it almost
certain that the differential between the P4s boards
and the Althons has been recovered.

My solution would be to wait for new motherboards with the
KT266A chipset.

Of course, if your application is really memory limited, you
could always wait even longer for the nForce 220 and a 
processor that could efficiently use its memory bandwidth:)

http://www.anandtech.com/chipsets/showdoc.html?i=1535&p=6

Regards
Josh Halpern

Eugene Leitl wrote:
> 
> ---------- Forwarded message ----------
> Date: Tue, 23 Oct 2001 08:34:14 -0600
> From: Scott Anderson <anderson at chem.utah.edu>
> To: chemistry at ccl.net
> Subject: CCL:RIMM vs SDRAM
> 
> Dear Tom:
> 
> Regarding your question about RIMM (RDRAM) v.s. SDRAM (or DDR SDRAM) for
> Gaussian:
> 
> I have done a fair bit of comparing machines for Gaussian, as I am planning
> to buy a bunch in the near future.  I have been using an MP2 Opt Freq job
> that takes a few hours as my benchmark.  I have also run much longer Direct
> Dynamics trajectory jobs on several different configurations.  The MP2 job
> was deliberately chose to avoid a lot of disk i/o because I really wanted to
> focus on the computational speed.  I have compared 1 and 1.4 GHz Athlons,
> PIIIs, and P4s.  The conclusion is that for Gaussian, the P4 is 20-30%
> faster than the Athlons for the same clock speed, and since you can get P4s
> with faster clocks, the P4 ends up being substantially faster.  This speed
> difference is quite surprising because the P4 floating point unit is
> substantially inferior to that of the Athlon, and as far as I have been able
> to determine, the rate deterimining step for these fast chips is memory
> speed.  Here the RDRAM really shines.  I have also compared Athlons with
> PC133 SDRAM and PC2100 DDR ram, which is theoretically substantially faster.
> For Gaussian jobs, I see no difference.
> 
> One interesting thing to look at in this regard is the review in the Tom's
> Hardware review of the new Athlon XP.
> http://www6.tomshardware.com/cpu/01q4/011009/index.html     They compare
> several different Athlons with different memory configurations, and also
> 2GHz P4s with both RDRAM and DDR memory.  The Athlon XP looks great on most
> of the benchmarks -- equalling or exceeding the P4, even though the P4 clock
> speed is much higher.   The one area where the XP fails, however, is memory
> speed.  Note that the P4 with RDRAM does 50% better than the P4 with DDR,
> and nearly a factor of two better than the Athlons with either PC133 or DDR
> memory.
> 
> My conclusions:
> 1. Gaussian is probably memory-bandwidth limited, and you should spend the
> extra money RDRAM.
> 2. DDR memory doesn't really make a huge difference -- certainly nothing
> like the factor of two you might expect.
> 3. The ideal configuration is very application dependent.  For Gaussian, I
> would go with the P4.  If you are running a mix of programs, you really have
> a tough decision.
> 
> Have fun
> 
> Prof. Scott L. Anderson
> Department of Chemistry
> University of Utah
> 315 S. 1400 E. Rm 1216
> Salt Lake City, UT  84112
> (801)585-7289
> FAX(801)581-8433
> www.chem.utah.edu/chemistry/faculty/anderson/anderson.html
> ----- Original Message -----
> From: "Tom Kuppens" <tom at hartree4.rug.ac.be>
> To: <chemistry at ccl.net>
> Sent: Tuesday, October 23, 2001 5:07 AM
> Subject: CCL:RIMM vs SDRAM
> 
> > Hi there,
> >
> > We have several linux machines that run Gaussian. Now we want to add some
> new
> > power to our park i.e. a PIV - 1700MhZ with 512M RAM (SDRAM or RIMM).
> >
> > Is there someone who can provide me with some info on the Gaussian and/or
> > overall performance with RIMM versus SDRAM modules.
> >
> > Thanks in advance,
> >
> > Tom Kuppens
> > Ghent University
> >
> > --
> > tom at hartree4.ac.be
> > http://studwww.rug.ac.be/~tkuppens/quantum/
> > ---
> > It seems to me, Golan, that the advance of civilization is nothing
> > but an exercise in the limiting of privacy.
> >      -- Janov Pelorat in Asimov's Foundation's Edge
> >
> >
> >
> > CHEMISTRY at ccl.net -- To Everybody  | CHEMISTRY-REQUEST at ccl.net -- To
> Admins
> > CHEMISTRY-SEARCH at ccl.net -- archive search    |    Gopher: gopher.ccl.net
> 70
> > Ftp: ftp.ccl.net  |  WWW: http://www.ccl.net/chemistry/   | Jan:
> jkl at osc.edu
> >
> >
> >
> >
> >
> >
> >
> 
> -= This is automatically added to each message by mailing script =-
> CHEMISTRY at ccl.net -- To Everybody  | CHEMISTRY-REQUEST at ccl.net -- To Admins
> MAILSERV at ccl.net -- HELP CHEMISTRY or HELP SEARCH
> CHEMISTRY-SEARCH at ccl.net -- archive search    |    Gopher: gopher.ccl.net 70
> Ftp: ftp.ccl.net  |  WWW: http://www.ccl.net/chemistry/   | Jan: jkl at osc.edu
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf