Most processing power/Watt

[~misfit~]

Something I've been thinking about recently, of the CPUs in the 700Mhz and
upwards range, which CPU, or family of CPUs gives the most for the least,
electrical-consumption-wise?

I'm running a few machines for SETI and am on a budget. I have a bit of
hardware and am able to decide which machines to run. I'd just like to know
which will give the best return for money spent on power. Hell, I just run
the fastest of what I have at the moment and damn the cost. It'd be nice to
know though, I'm on a budget.

Any ideas folks?

Cheers,

 

[Michael Brown]

Something I've been thinking about recently, of the CPUs in the
700Mhz and upwards range, which CPU, or family of CPUs gives the most
for the least, electrical-consumption-wise?

Of commonly-available, easy to buy, CPUs, probably the Via/Cyrix Nehemiah
ones. But more because they have incrediably low power consumption (sub 10W
IIRC), rather than they have particularily high computing power. The P4E's
are definately off the menu (too much heat), as are the Celerons (as much
power as the P4's, a fraction of the computing grunt), and most of the A64's
(high power usages as well). For off-the-shelf CPUs, the 35W Athlons would
probably take second place. Mobile P4's are generally hamstrung by slow
busses and slow speeds, and don't really perform to well for the power they
comsume AFAIK (though I haven't really been out shopping for them ).

If you're willing to go exotic, you can get a lot better. You can get
XP1800's that use up just 16W (if you can find a source, and if you can get
your hands on a uPGA motherboard). I'm sure Intel has Pentium-M's that can
get pretty close as well. These might even threaten the Via chips because of
their higher computing power.

Of course, it all comes down to initial cost as well. With the exception of
the Via chips (which are relatively cheap to start off with), the cost is
inversely proportional to the computing-grunt-per-watt. If you want a 16W
Athlon in a uPGA board, you're going to have to shell out some big $$$ to
get it.

[...]

 

[Cuzman]

" Something I've been thinking about recently, of the CPUs in the 700Mhz
and upwards range, which CPU, or family of CPUs gives the most for the
least, electrical-consumption-wise? "


You'll be hard-pressed to find something that outperforms an Athlon XP
Mobile in that respect.

 

[~misfit~]

Something I've been thinking about recently, of the CPUs in the
700Mhz and upwards range, which CPU, or family of CPUs gives the most
for the least, electrical-consumption-wise?

Of commonly-available, easy to buy, CPUs, probably the Via/Cyrix
Nehemiah ones. But more because they have incrediably low power
consumption (sub 10W IIRC), rather than they have particularily high
computing power. The P4E's are definately off the menu (too much
heat), as are the Celerons (as much power as the P4's, a fraction of
the computing grunt), and most of the A64's (high power usages as
well). For off-the-shelf CPUs, the 35W Athlons would probably take
second place. Mobile P4's are generally hamstrung by slow busses and
slow speeds, and don't really perform to well for the power they
comsume AFAIK (though I haven't really been out shopping for them
).

If you're willing to go exotic, you can get a lot better. You can get
XP1800's that use up just 16W (if you can find a source, and if you
can get your hands on a uPGA motherboard). I'm sure Intel has
Pentium-M's that can get pretty close as well. These might even
threaten the Via chips because of their higher computing power.

Of course, it all comes down to initial cost as well. With the
exception of the Via chips (which are relatively cheap to start off
with), the cost is inversely proportional to the
computing-grunt-per-watt. If you want a 16W Athlon in a uPGA board,
you're going to have to shell out some big $$$ to get it.

[...]

Thanks Michael.

 

[~misfit~]

" Something I've been thinking about recently, of the CPUs in the
700Mhz and upwards range, which CPU, or family of CPUs gives the most
for the least, electrical-consumption-wise? "


You'll be hard-pressed to find something that outperforms an Athlon XP
Mobile in that respect.

I think you may be right there, especially at stock speed. However I'm
financially limited to older machines at present. I have a Tualatin Celeron
and another one or two on the way, I think they should be fairly good in
that regard. I'm just trying to work out which machines (I have an XP1800+
OCed to 2.1Ghz, 10.5 x 200, 1.8vcore, but that sucks the power) to leave
running 24/7 to maximise SETI contributions per dollar spent on power. I'm
on an invalid's benefit and certainly not rich.

Cheers,

 

[Phil Weldon]

The SETI@home application is very sensitive to L2 cache size, and needs
neither a display device, a hard drive, nor very much random access memory.
A loosely coupled cluster would likely provide the most bang per buck. With
a large L2 cache, main memory speed is not much of a factor.

You might also consider one, very efficient, switching power supply for the
entire cluster (perhaps a Lamda from ebay?)

Eliminating as many support elements as possible may lower power consumption
per work unit more than going with lower power CPU's.

--
Phil Weldon, pweldonatmindjumpdotcom
For communication,
replace "at" with the 'at sign'
replace "mindjump" with "mindspring."
replace "dot" with "."

 

[~misfit~]

The SETI@home application is very sensitive to L2 cache size, and
needs neither a display device, a hard drive, nor very much random
access memory. A loosely coupled cluster would likely provide the
most bang per buck. With a large L2 cache, main memory speed is not
much of a factor.

You might also consider one, very efficient, switching power supply
for the entire cluster (perhaps a Lamda from ebay?)

Eliminating as many support elements as possible may lower power
consumption per work unit more than going with lower power CPU's.

All good points Phil. However, I'm running the new BOINC client now and I'm
fairly sure it needs a HDD. At least, I don't have the know-how to do it any
other way.

I have heard others mention that Seti is L2-dependant as well. However, from
personal experience I have noticed little, if any, discernable difference
between similarly-clocked Coppermine and Tualatin Celerons (128 and 256KB
respectively). Maybe it's only over 256KB that the benefits kick in.

Thanks for your input.

 

[Phil Weldon]

Try this URL
http://arstechnica.com/wankerdesk/3q99/setiopt-1.html

and this URL
http://www.cox-internet.com/setispy/

and this URL

The L2 benefit really kicks in at 1 MByte, unless new versions of the
SETI@home client have been written to make more effecient use of 256 KByte
and 512 KByte L2 caches.

--
Phil Weldon, pweldonatmindjumpdotcom
For communication,
replace "at" with the 'at sign'
replace "mindjump" with "mindspring."
replace "dot" with "."

 

[kony]

Something I've been thinking about recently, of the CPUs in the 700Mhz and
upwards range, which CPU, or family of CPUs gives the most for the least,
electrical-consumption-wise?

I'm running a few machines for SETI and am on a budget. I have a bit of
hardware and am able to decide which machines to run. I'd just like to know
which will give the best return for money spent on power. Hell, I just run
the fastest of what I have at the moment and damn the cost. It'd be nice to
know though, I'm on a budget.

Any ideas folks?

Cheers,

SInce most people are not using their systems to 100% processing
capacity most of the time, it could be most power efficient to
just run SETI on your main system as low-priority, and have HDD
spindown set low enough that they're usually asleep when you're
not using system otherwise.

As for a 2nd system, I agree with other posters that large L2 is
most beneficial. Coppermine CPU topped out at 256K L2, and old
P3 ~ 600MHz wasn't as good of a heat vs performance ratio so best
bet is probably a Tualatin P3 w/512K L2 (except too expensive) or
a Barton Mobile underclocked/undervolted.

 

[~misfit~]

SInce most people are not using their systems to 100% processing
capacity most of the time, it could be most power efficient to
just run SETI on your main system as low-priority, and have HDD
spindown set low enough that they're usually asleep when you're
not using system otherwise.

As for a 2nd system, I agree with other posters that large L2 is
most beneficial. Coppermine CPU topped out at 256K L2, and old
P3 ~ 600MHz wasn't as good of a heat vs performance ratio so best
bet is probably a Tualatin P3 w/512K L2 (except too expensive) or
a Barton Mobile underclocked/undervolted.

Thanks Dave,

 

[~misfit~]

Try this URL
http://arstechnica.com/wankerdesk/3q99/setiopt-1.html

Ummm, when was that was written? The guy mentions his machine, a Celeron
300a @ 450Mhz and version 1.06 was the latest version of SETI GUI (Now
3.08). I read all three pages and saw no mention of L2 cache.

and this URL
http://www.cox-internet.com/setispy/

Having used SetiSpy myself I'm familiar with this page and have exchanged
emails with Roelof on occasion. However, once again no mention of L2 cache
and it's effect on WU throughput times at this page.

and this URL

The L2 benefit really kicks in at 1 MByte, unless new versions of the
SETI@home client have been written to make more effecient use of 256
KByte and 512 KByte L2 caches.

Ok, that could be true. I have only set SETI up on machines having L2 caches
up to 512KB. My girlfriends Barton at 2.2Ghz and my Tbred at 2.1Ghz are very
similar in WU times, hers is slightly faster as you would expect from the
clock-speeds, so I doubt there is much benefit going from 256KB to 512KB. I
would think that, until recently, hardly anyone (considering the user-base
of Seti) had machines with greater than 512KB of L2 cache so it seems a bit
strange to have software that seems optimised for machines other than the
ones it's going to run on.

Cheers,

 

[N. Thornton]

I've no idea what your knowledge is misfit, so this might be,
something. Shutting down as many software processes as possible frees
cpu and ram up for the seti task.

The other question is why youre doing seti. If you look in
rec.org.mensa for Penny's explanation of it you'll see that it doesnt
look too likely that its doing what they claim it is. Penny reckons
its military, and nothing at all to do with what it claims. Personally
I dont know about it, but its well worth reading what she says.


Regards, NT

 

[~misfit~]

I've no idea what your knowledge is misfit, so this might be,
something. Shutting down as many software processes as possible frees
cpu and ram up for the seti task.

Thanks. I have it running as lean as I can.

The other question is why youre doing seti. If you look in
rec.org.mensa for Penny's explanation of it you'll see that it doesnt
look too likely that its doing what they claim it is. Penny reckons
its military, and nothing at all to do with what it claims. Personally
I dont know about it, but its well worth reading what she says.

Just Googled it and read it. Interesting theory. I agree that seti isn't
ideally set-up, however I've been crunching for over five years now and have
sort of developed an inertia. <g>. Also I'm the founder of a team. To me
it's partly about my passion for building/overclocking machines and getting
the best I can out of them. Seti is a way to quantify the success or
otherwise of my hobby. I'm an invalid and it's what I do to keep myself
busy.

There are other distributed computing projects I've looked at. However, in
*all* cases it appears that commercial entities stand to make money out of
my CPU cycles. I don't want my hobby to end up making the rich richer.

Penny could be right about SETI, however, over the years I've read a lot of
criticism of it and conspiracy theories about it. Who do you believe?
Someone who may be taking lithium (Mensa? After all, they say genius is
close to madness. I should know, I once took the test to join Mensa, way
back in the pre-internet days, passed but then decided not to join) or do
you go with your gut feeling?

Thanks for the input NT.

 

[N. Thornton]

N. Thornton wrote:

Thanks. I have it running as lean as I can.

I realised once I pressed post that that was unlikely to be
appropriate advice.

There are other distributed computing projects I've looked at. However, in
*all* cases it appears that commercial entities stand to make money out of
my CPU cycles. I don't want my hobby to end up making the rich richer.

I guess any distributed computing project costs to set up, so people
will only do it when theres the possibility of profit. Only government
funds would setup something that will never return anything.

Penny could be right about SETI, however, over the years I've read a lot of
criticism of it and conspiracy theories about it. Who do you believe?

no-one. I just hope you consider your input to seti to be a fulfilling
way to spend your life. I guess its different to my approach to
things. Good luck.


Regards, NT

 

[~misfit~]

I realised once I pressed post that that was unlikely to be
appropriate advice.



I guess any distributed computing project costs to set up, so people
will only do it when theres the possibility of profit. Only government
funds would setup something that will never return anything.

I remember seeing a DC project that paid you for your CPU cycles a while
back but it wasn't feasible for anyone living outside of the US. The pay-out
looked like it was reasonable too. As in, it would have actually covered
electricity costs. I don't think you could have made a living out of it
though.

no-one. I just hope you consider your input to seti to be a fulfilling
way to spend your life.

Let's not get carried away, it's not the way I spend my life, it's a hobby.
I do other things as well, like, umm, post on usenet, and, errr, play
computer games. ;-)

I guess its different to my approach to
things. Good luck.

That's what makes the world so interesting.

 

[Phil Weldon]

"it seems a bit strange to have software that seems optimised for machines
other than the ones it's going to run on." That fact may be strange to
you but that's the way it is; and also one of the reasons that at least one
rogue version of the application was written (which couldn't be accepted
because of the need to protect the integrity of the signal processing.)

As for the effect of L2 cache size on the SETI@home application, try
http://www.cox-internet.com/setispy/efficiency.htm
which is just a little further into
http://www.cox-internet.com/setispy/ .




--
Phil Weldon, pweldonatmindjumpdotcom
For communication,
replace "at" with the 'at sign'
replace "mindjump" with "mindspring."
replace "dot" with "."

 

[Thomas]

Just Googled it and read it. Interesting theory. I agree that seti
isn't ideally set-up, however I've been crunching for over five years
now and have sort of developed an inertia. <g>. Also I'm the founder
of a team.

I am crunching SETI blocks too, but didnt bind myself to a team yet....
Wisheth thou mine allegiance, my Lord?

Hahaha.. sorry, been reading too much of 'the worm ouroborous' ;-)

Thomas

 

[~misfit~]

"it seems a bit strange to have software that seems optimised for
machines other than the ones it's going to run on." That fact may
be strange to you but that's the way it is; and also one of the
reasons that at least one rogue version of the application was
written (which couldn't be accepted because of the need to protect
the integrity of the signal processing.)

It seems even stranger then that they've continued through several revisions
without optimising it to run on what must be 98% of their users machines.

As for the effect of L2 cache size on the SETI@home application, try
http://www.cox-internet.com/setispy/efficiency.htm
which is just a little further into
http://www.cox-internet.com/setispy/ .

That's interesting, the Celeron Tualatin with it's 256KB cache is a lot
lower down the ladder than the Celeron Coppermine with it's 128KB cache.
That seems to go against the trend. I wonder why? Cache latency maybe
(Something I know v.little about).

Using the calculator on that page, for a 1Ghz Celeron, all other things
being equal, the Tui comes in at 8.8hrs and the CuM at 8.36hrs. Hmmmm......

It'll be interesting to see that list updated (if it ever is now) to include
the Pentium M Dothan with it's 2MB L2 cache. The Pentium M Banias (1MB L2)
is the fastest CPU without L3 cache, in position 3, only beaten by
v.expensive server CPUs. Also the Pentium 3 Tualatin is higher than the P4
EE with it's L3 cache. That confirms nicely what I said in my little rant in
the "Looking for someone to clear up this Prescott confusion!" thread.

Thanks for the link to that, that's pretty much what I was looking for with
this whole thread, although it doesn't include power consumption I can get
those figures elsewhere.

Maybe I'm not doing the right thing. At the moment I'm in the process of
replacing my Coppermine Celerons with Tualatin Celerons and mobos, thanks to
the generosity of a good friend who is re-organising his SETI farm. I've
always sung the praises of the Tualatin Celeron, knowing how good the P3
Tualatin was and bearing in mind that it's the only Celeron ever made that
has 256KB of L2. It's only recently I have first-hand experience of them. I
didn't think they seemed any faster clock-for-clock than the Coppermines and
put it down to the fact that the Coppermines I have are running in the
wonderful BX chipset boards whereas the Tui's are running in Via boards. It
would seem that in fact they are inferior CPUs. Cache latency..... I seem
to remember reading somewhere that Intel crippled them by increasing cache
latency. After all, the P3 Tualatin does *so* well.

Thanks Phil.

 

[~misfit~]

I am crunching SETI blocks too, but didnt bind myself to a team
yet.... Wisheth thou mine allegiance, my Lord?

That is a very generous offer, and I thank you for it. However, I see you
are from The Netherlands and the team I founded is a team for the readers
and posters in the nz.comp newsgroup. I am in New Zealand.

That being said, there is no reason why you can't be out first member from
outside NZ if you so wish. Are you still crunching Seti classic or have you
gone over to the new BOINC client yet?

Classic team: http://setiathome.ssl.berkeley.edu/stats/team/team_155750.html

BOINC team: http://setiweb.ssl.berkeley.edu/team_display.php?teamid=30823

Our members are slowly switching over to BOINC as classic will be terminated
at an undisclosed date not too far into the future. There are currently
problems with the SETI servers (when isn't there? ) so any changes
could be difficult currently.

Feel free to join if you so desire. I would welcome your contribution.

Hahaha.. sorry, been reading too much of 'the worm ouroborous' ;-)

Do you recommend it? I enjoy fantasy but have yet to read it.

Regards,

 

[kony]

That's interesting, the Celeron Tualatin with it's 256KB cache is a lot
lower down the ladder than the Celeron Coppermine with it's 128KB cache.
That seems to go against the trend. I wonder why? Cache latency maybe
(Something I know v.little about).

Note that the Tualatin P3 (version with only 256K L2) does
better. Difference is FSB & memory speed... Multiplier as
mentioned on that page.

Using the calculator on that page, for a 1Ghz Celeron, all other things
being equal, the Tui comes in at 8.8hrs and the CuM at 8.36hrs. Hmmmm......

But, CuM uses more power, and Tui can clock higher. Tui on
133MHz FSB/Mem would be good if particular specimen doesn't
require extreme Vcore increase to get there, which it probably
wouldn't since resulting speed is only 1.33GHz.

Maybe I'm not doing the right thing. At the moment I'm in the process of
replacing my Coppermine Celerons with Tualatin Celerons and mobos, thanks to
the generosity of a good friend who is re-organising his SETI farm. I've
always sung the praises of the Tualatin Celeron, knowing how good the P3
Tualatin was and bearing in mind that it's the only Celeron ever made that
has 256KB of L2. It's only recently I have first-hand experience of them. I
didn't think they seemed any faster clock-for-clock than the Coppermines and
put it down to the fact that the Coppermines I have are running in the
wonderful BX chipset boards whereas the Tui's are running in Via boards.

Via 693 was slow, but Via 694 with good bios may outperform a
BX... they're close though.

It
would seem that in fact they are inferior CPUs. Cache latency..... I seem
to remember reading somewhere that Intel crippled them by increasing cache
latency. After all, the P3 Tualatin does *so* well.

Keep in mind that the motherboard, memory, drive, are all
overheads that may not remain at a fixed constant but still are a
large overhead. If you have a CPU that's 30% faster but uses 40%
more power, you're probably still coming out ahead by using it.