From: Eugene Leitl (eugene.leitl@lrz.uni-muenchen.de)
Date: Thu Jul 15 1999 - 17:26:15 MDT
"Phillip D. Matz" <matz@wsunix.wsu.edu>
I have two C300As overclocked to 450Mhz which have been running Win95 24/7
since December 1998. They have been running Gaussian 98W, a molecular
modeling program which is fpu intensive, nonstop 24/7 as well. The computers
have no overclocking stability issues with G98W, btw it has been my experience
that if a pc is stable enough to run a gaussian calculation without crashing
then nothing will crash it.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From: "Donald B. Kinghorn" <kinghorn@U.Arizona.EDU>
I'd like to make two comments;
1) I do (very) high precision QM energy calculations with wave function
optimizations that may run for months --- I would be very uneasy
about publishing new energy upper bounds that were computed with
hardware that was pushed beyond specifications. Maybe it's safe but
I wouldn't want to stake my reputation on it. [Remember the FOOF bug?
For a short time results computed on Intel hardware were unpublishable.]
2) A killer machine for one problem may be a piece of junk for another problem.
[I think this has been stated several times] The memory subsystem can have
a large impact on performance for some calculations.
Example:
I put together a PII400 system with 128M PC100 cas 2 SDRAM and was
casually testing it against a PPro 180 system with 128MB 60ns DIMM (at 60MHz)
The PPro has 256k full speed cache and the PII400 has 512k 1/2 speed cache.
(some other difference too I assume)
For "small" jobs the PII400 was basically twice as fast as the PPro180.
For "large" jobs (80MB RSS) the PII400 was about 3.5 times faster!!
[there was no disk or network IO on either system during the calc's]
I was very surprised by this. I can only assume that it was due to the
(overall) faster memory subsystem on the PII400 system.
[I'm eagerly awaiting a K7 system with a 200MHz memory subsystem
that might be real interesting if the price is right]
Bottom line: A system is only killer if it works well on your problems
and gives you results that you can use.
but please don't stop experimenting with the overclocking --- this is
very interesting stuff. [I'm going to try it too] :-)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From: Glenn Plas <glenn.plas@mchugh.com>
Duh, the right link to the SMP smackdown is:
http://www.ars-technica.com/cpu/3q99/smp/smp-1.html I switched the 'cpu'
part to 'spu'. Copy/pasting the link was too much trouble :P huh ...
Sorry for the extra mail.
I was not going to post the vendors at first, I was a little worried some
of you would think I was making free publicity. But the place I bought it
supports 100% 'the cause', meaning 'the overclockers cause'. I got my
CPU's over international mail from the US to Belgium (yes, the only
chicken-less, coke-less lil' country in Europe for a while if you followed
the news lately). The place is called http://www.outsideloop.com . Take a
look at their site, they are cool. Their prices might be a little higher
than others but they hand-pick Processor batches and for a buck or 10 more
you have a guaranteed OC'er. Lemme tell you that what they charge over
there for their stuff + VAT + shipping oversea to my front door in Europe
costs less than buying the stuff over here. And a lot of the stores DO not
ship outside the US. Heads up for them.
There are other vendors that I know a lot of my fellow clockers use (almost
all in the US unfortunately for us Euro's). One of my favorite sites
among ARS is HardOCP ( http://www.hardocp.com ). There are links to Azzo
Computing and Dallasmemory and other shops known in the field.
The 300a thing is indeed a good question. I would still go for them I
think considering their price. The main reason is because they easily run
at 100Mhz FSB. Since Intel knows people are seriously clocking them, all
the new PPGA celeron's available today have huge multipliers (6, 7, etc ..)
now since they are supposedly only good according to Intel for the 66 Mhz
Bus. When you set the FSB to 100Mhz, a multiplier of 7 is just too good to
be true. (700Mhz anyone ?? :P). As a side effect when clocking in the
case of high multiplier locks, you'll have to use a different FSB (83, 75,
etc ) to achieve a stable system. This has implications for the PCI bus
running seriously out of spec at 40 +mhz in some cases because the PCI bus
divider on most motherboards instead of the default speed of 33 Mhz . At
66 Mhz and 100Mhz the divider becomes respectively 1/2 and 1/3 .
In short, in that case you're overclocking every CARD in your PC , more
correct your whole system. Most lower end cards blow chow on that.
Sometimes this can be safely done but the successrate is less. Your
mileage WILL vary. The latest news is that the 366 PPGA celeron is a
good candidate. But then the Epox KP6-BS is not a good choice since it
does not have a lot of different FSB's possibities. I would defo go for
the Abit Dual BP6 Mobo in that case. In fact, that motherboard wasn't out
yet when I designed my system, otherwise I would've taken that one. No
SLOT1/PPGA convertors to mess with.
Pfff, my hands hurt.
Here are my specs:
2 * cele 300a @ 504
2 * MSI 6905 Ver. 1.1 S370
Epox KP6-BS dual board
Adaptec U2W 2940 SCSI
IBM U2W LVD 4.5 GIG System HD (don't know the type# by heart)
Quantum Fireball 6 Gigger file-storage HD
Intel Etherexpress Pro 100 fast ethernet NIC
2* Creative Labs VoodooII (SLI)
Creative Labs RIVA TNT (AGP)
Somewhat older ISA Creative AWE64 Sound blaster
198 Meg SDRam (6ns) 1 * 128 from Siemens and 1 * 64 from Samsung (Micron is
also a very good brand, go for atleast CAS2 ram)
2 * GlobalWin CPM25603-16 CPU cooler/heatsink combo
2 Case 25 cfm 12V blowers Nidec Gamma 30 (extra ones on top of the standard
case fans) http://www.nidec.com/gamma30.html
Case Supermicro SC 750A http://www.supermicro.com/PRODUCT/Chassis/sc750.htm
MGE UPS Pulsar 750VA 8+
And my shoesize is a 9 and a half :-] thus a 43 in Europe
Good luck,
/// Glenn
Linux -is- user friendly. It's just very selective about who its |
friends are. And sometimes even best friends have fights
-----Original Message-----
From: Fulton CR (Charles) [SMTP:fultoncr@ucarb.com]
Sent: Wednesday, July 14, 1999 4:12 PM
To: 'glenn.plas@mchugh.com'
Subject: a very nice system (from beowulf list)
Glenn,
Any chance you could send me a full listing of all the parts in your system
w/ approx prices. I've started doing some research on the feasability of
building a dual system at home. Your system sounds quite nice.
I tried the link at arstechnica.com but it didnt work. but that is a very
nice site thanks.
If you were building it today would you still use the cel. 300a's ?
Also any hints for online purchasing would be great. I always like to see
other ppl's resources. well this is just the thing that is going to make my
new bride very upset hehehe. "Another computer!!!" :)
thanks in advance,
-charlie.
p.s.
I just wanted to mail direct since this is kinda off topic.
<your sys>
motherboard - Epox KP6-BS motherboard (200$?) (slot1)
processors - 2 Celeron 300a overclocked to ~504Mhz
case - ?
extras - 2 Nidec Gamma 30 blowers
hd (s) (scsi/ide?) -
ram -
video card(s) - 3dfx vodoo2
sound card -
</your sys>
> The most important thing perhaps is cooling. The CPU coolers I have are
> the best of breed(globalwin). Also I've spent a lot of money on a very
> good case (one that has very good airflow) with 2 extra Nidec Gamma 30
> blowers. The second most important is the PSU. Go for 300 WATTS or
> more.
> I've seen huge powerdrops on the -12v and the +12 Volt line when
> starting
> Quake3 when this board was still in an old case. That does not help
> stability.
>
> On Tue, 13 Jul 1999, Glenn Plas wrote:
>
> > I have made a (non beowulf) dual Celeron machine based on the Epox
> > KP6-BS motherboard. It is a SLOT 1 motherboard . I used 2 MSI 6905
rev
> > 1.1 Slotkets to SMP-enable the PPGA celerons'. It also allows you to
> > overclock them. I'm currently running Suse 6.1 on it. The Celeron's
> > are 300a AkA hard-to-get these days. They are clocked stable at 504
Mhz.
> > yes, you read that, I clocked them 204 Mhz higher than their specs.
> > It's outside the scope maybe but I'll continue anyway. It's the
> > cheapest wy to build a rocking fast dual box. 504 is not a guaranteed
> > FSB though. 450Mhz is reached with almost all 300a's (FSB 66 -> FSB
> > 100_), 464 with a FSB of 103 is being reached by a lot of people too.
> > 504 is more an exception.
>
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Glenn Plas <glenn.plas@mchugh.com>
nlike your statement, the Celeron is not from the same core as the PII.
As a mather of fact it's design looks more like the Intel Xeon PII than
the regualar PII.
The Celeron's Core is Mendocino. The PII's core are either a
Klamath/Deschutes core . the big difference is the on die L2 cache,
which runs at core speed. The PII's cache runs at core speed / 2.
Celeron's really rock. Intel's money cow, the PII is just a big joke.
The celeron's that run at the same core speed as a PII kicks serious
butt. Including PIII's.
Because of the bad rep the first cacheless Celeron's had, the CPU was
destined to be sold to the lower segment of the market. And that's great,
because 'we' can buy dirt cheap CPU's that rock.
Also, PC100 RAM has totally no advantage if you're Front bus speed is set
too 66Mhz. Like someone stated before, celeron's overclock easily. Abit
motherboards are your friends.
Take a look at the BP6 Motherboard from abit. A dual PPGA motherboard and
probably one of the most stable boards in the world. (and cheap ..)
As long as no program uses the new KNI instruction set for the PIII, it's
really a waste of money to pay 3 times as much for it than the cost of King
Celeron.
/// Glenn
Unix -is- user friendly. It's just very selective about who its |
friends are. And sometimes even best friends have fights
-----Original Message-----
From: Robert G. Brown [SMTP:rgb@phy.duke.edu]
Sent: Tuesday, July 13, 1999 11:35 PM
To: Christopher J. Morrone
Cc: Conor Masterson; beowulf@beowulf.gsfc.nasa.gov
Subject: Re: Fast Ethernet - worth it?
On Tue, 13 Jul 1999, Christopher J. Morrone wrote:
> On Tue, 13 Jul 1999, Conor Masterson wrote:
>
> > Hi,
> >
> > > > BTW: AMD K-6 CPUs - good, bad or indifferent?
> > >
> > > For FPU intensive tasks K6 is in general not a good choice, for
integer operations K6 is a good choice. But maybe AMD-K7 will change the
FPU dominance of Intel stuff.
> >
> > We're also looking at buying a Beowulf cluster for a CPU/floating point
> > intensive application, and Intel Celeron 466 CPUs seem to be the best
value,
> > as they have the same performance on that type of application as PII's,
the
> > only difference being a smaller L2 cache. You can also get a dual
Celeron box
> > for roughly the price of a single PII of the same speed. Has anybody
tried
> > this configuration?
>
> Doesn't the Celeron limit you to 66MHz memory access vs the PII's 100MHz
> memory access?
Yes, but for many things this doesn't seem to matter much. For others,
of course, it will. I've a bunch of applications (including my
numerical Monte Carlo code, but other stuff too) that runs within 5% of
the same speed on PII's and Celerons, with or without PC100 SDRAM. This
isn't unreasonable -- the PII and Celeron share the same core.
For many things, clock speed rules. For others, cache size and memory
speed are important. In my mind (because my applications fit in the
former category) the Celeron is the best deal on the market -- a beowulf
node can be built for as little as $350 with careful shopping, and it is
pretty easy to build nodes for $500-750 (depending on how much memory
and disk and what kind of network). This is pretty close to what PIII
processors cost all by themselves.
Between the Celeron at one end (most TOPS/FLOPS/IPS per dollar) and the
21264 Alpha at the other (most TOPS/FLOPS per node at a still-sane cost
per dollar) is a vast wasteland of less-than-optimal cost benefit where
only a few people will be wise to buy. To find out if you're one of
them, you pretty much have to benchmark and test, ideally with the
actual application you plan to run. Realize that you can get two, maybe
three 466 MHz Celeron systems for what a single 500 MHz PIII system
costs -- there has to be a pretty significant advantage to the PIII to
make it worth it.
rgb
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
From: "Phillip D. Matz" <matz@wsunix.wsu.edu>
Allow me to specify in greater detail the comments I made in the previous post.
I have a genuine 450MHz PII system (no overclock) which I run gaussian98
calculations on. I have the two previously mentioned C300As overclocked to 450MHz,
and I have a K6-2 333MHz system (so a total of 4 different computers).
Now, it does NOT matter which computer I run a gaussian98 on I ALWAYS get the same
answers (energies, MO coefficients, etc.). I have checked and rechecked my results
between all four computers as well as published results in the literature (JACS, J
Chem Phys, Etc.). If these calculations are subject to ANY hardware derived errors
then all four of the above machines must share that same flaw.
Also, my reference to the stability of an overclocked computer with respect to
gaussian98 is derived from the following. First, all four cpus listed above can
still be further overclocked without any degradation in system stability during
normal usage (i.e. running netscape to check my email). In fact they can all be
overclocked another 12% (100MHz FSB -> 112MHz FSB) and run just fine (i.e. no
crashes or errors when running any/all my current software programs) until I attempt
to run a gaussian98 calculation. For some reason the gaussian98 code seems
particularly sensitive to cpus operating in their marginally unstable regions. I
have verified that this phenomena is not an artifact of the 112MHz bus since the K6
is not clock-multiplier locked and the motherboard can go up to 124MHz FSB with no
g98 issues unless I also try and clock the cpu above 333MHz. Bottom line is that if
your computer is overclocked and does not crash when running G98W then I would say
it is as stable as it will ever need to be.
Phillip Matz
This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 15:04:29 MST