VIA mini-itx DF performance

[erk]

Has anyone tried DF with the VIA mini-itx boards?

I tried a VIA EPIA-CL10000 board (C3 1Ghz) with FreeBSD 5.2 and it did not perform very well, much slower than a 1GHz PIII, however on most normal tasks like compiling a kernel etc. it was quite quick. It was not my system so I could not test other OS's and DF versions like the gcc vs. the Intel compiler versions under Linux etc. I like the idea of the mini-itx as a diskless folding farm due to the very low power consumption.

[allenfinch]

I don't have one, but low power computing is of some interest to me so I try to keep an eye on developments here.

My understanding of this project is that it is much less floating point intensive than many other distributed projects. The reason I point this out is that I would expect this project to be the one where the c3 cores would have the best hope of performing decently, relative to other chips. FP performance has always been the low point of these chips. If you are looking to build a farm for application to other distributed projects as well, the via becomes even less appealing.

The low power consumption is certainly enticing but, when you consider that distributed computing puts a heavy load on the processor, performance per watt for the via chips isn't that impressive. In other words, you are generally better off getting fewer amd xp, even if power demand is the key criterion you are looking at for a farm. In general these chips are still of appeal primarily where size and noise are the key concerns and where performance can be sacrificed.

The mobile amd xp chips would probably be the best choice for such a thing, since you have more flexibility in being able to put these in desktop motherboards(though this isn't always a sure thing). The pentium M would be impressive for such a thing(less so if cost is an issue), but you can't readily buy a motherboard and plunk an M in it.

Don't know much yet about the performance of the via nehemia core. Hopefully it will be more readily available by mid year. The preliminary numbers I've seen, though, suggest it isn't likely to alter the balance of things here.

[erk]

Originally posted by allenfinch
I don't have one, but low power computing is of some interest to me so I try to keep an eye on developments here.

My understanding of this project is that it is much less floating point intensive than many other distributed projects. The reason I point this out is that I would expect this project to be the one where the c3 cores would have the best hope of performing decently, relative to other chips. FP performance has always been the low point of these chips. If you are looking to build a farm for application to other distributed projects as well, the via becomes even less appealing.

The low power consumption is certainly enticing but, when you consider that distributed computing puts a heavy load on the processor, performance per watt for the via chips isn't that impressive. In other words, you are generally better off getting fewer amd xp, even if power demand is the key criterion you are looking at for a farm. In general these chips are still of appeal primarily where size and noise are the key concerns and where performance can be sacrificed.

The mobile amd xp chips would probably be the best choice for such a thing, since you have more flexibility in being able to put these in desktop motherboards(though this isn't always a sure thing). The pentium M would be impressive for such a thing(less so if cost is an issue), but you can't readily buy a motherboard and plunk an M in it.

Don't know much yet about the performance of the via nehemia core. Hopefully it will be more readily available by mid year. The preliminary numbers I've seen, though, suggest it isn't likely to alter the balance of things here.

I understand what you are saying, and it holds true for small folding farms of a few machines. Currently I have 20 machines doing Distributed Folding, a mixture of P4's and Athlons with a few PowerPC's and a Celeron. I have shelled out about $1,800 US in electricity for my folding farm in the last 12 months, however the heat loading is so high that my airconditioning bill the farm generates is almost twice that, the compressor is running at 80% duty cycle up from 30% 12 months ago, and it runs at 100% for several weeks over summer. It's so high that I have had to stop adding to my folding farm for the last few months.

I have calculated that I could cut my DF power/aircon bills by 70% by using the same GHz worth of mini-itx boards, they pull less then 10watts each as reported in this article http://www.mini-itx.com/projects/cluster/ vs. over 100watts for a basic P4 system.

As it's all coming out of my own pocket I am quite keen on investigating any way of reducing heat and power.

[rsbriggs]

Via 800 Mhz board won't run the DF code. You just get an instruction exception. I bought a batch of them, hoping to make a DF mini-super-comp out of them....

[erk]

Originally posted by rsbriggs
Via 800 Mhz board won't run the DF code. You just get an instruction exception. I bought a batch of them, hoping to make a DF mini-super-comp out of them....

I have tried the M9000 and CL10000 933Mhz and 1GHz boards both under FreeBSD with no problems. As there are different clients for different OS's you might have some issues on one OS and not the other.

[allenfinch]

I think you are making two important errors.
"I have calculated that I could cut my DF power/aircon bills by 70% by using the same GHz worth of mini-itx boards..."

I don't question your air conditioning calculations but you are making a serious error if you are assuming that the same ghz worth of mini-itx is going to perform anywhere near a comparable ghz in intel or amd.

"they pull less then 10watts each as reported in this article http://www.mini-itx.com/projects/cluster/ vs. over 100watts for a basic P4 system."
These numbers suggest you are comparing an idle via board and chip to an intel under load. Amd chips are also not quite the power gluttons that the p4s are. A 2600 barton, for example, has a max a bit under 70 watts. The c3s you are looking at will max around 15-24watts depending on chip generation. The draw under load is what you need to be looking at, since these chips will be crunching 24/7. I would bet my left arm that the 2600 barton will be more than 3-5 times the performance of the 1ghz c3 you are looking at, particularly with the majority of distributed projects and their floating point demands. As I said before this project is a bit unusual in that regard.

Take a look at this graphic. This is a synthetic bench and I don't claim it is a definitive answer to anything but it doesn't appear to be too out of whack with what I have seen elsewhere.
http://www.overclockers.com.au/image...33/figure6.jpg
Note especially the fpu performance of a 1ghz down BELOW a p2 333. What advantage are you getting in pulling perhaps 1/5 the power of a faster xp if you are getting 1/8 the performance?


I read that article you linked to. I don't have other performance numbers for the 1ghz c3 you are looking at, but I ran some numbers quite some time ago around the older ezras and folding at home. I really don't believe for an instant that his 12 800 mhz node system performs at 4 2.4ghz p4s. That would make it clock for clock performance equivalent. The vias do NOT perform like that. They may very well appear to give equally satisfying performance for certain tasks, but I think it would be a serious miscalculation to believe that they are going to crunch for distributed projects on an equal performance per ghz basis. The newer chips, of course, perform better but then again newer amds perform better for the same power budget(of previous amds) as well. I haven't seen anything yet which suggests that the previous balance has been significantly upset.


http://users.erols.com/chare/elec.htm gives some processor electrical specs. Though it doesn't specifically give maximum dissipation for the chip you are talking about it does give typical figures, which will be lower.

Creating the same ghz in via systems would lead to significantly less crunching power.
It wouldn't really make sense to spend a lot of money to create a folding farm that consumes less power than what you have now, if you also take a major performance hit comparable to the power you save. You would be much better off reducing costs by simply shutting down your lower performance rigs.

The key issue is performance per watt. Although performance per watt may, in some instances, be in the same ballpark as what you get with an amd or intel, it is not superior. If you wanted to recreate a farm of equal performance to what you have now at reduced power consumption you would be more likely to accomplish that at lower initial cost and lower total power budget by buying a small number of mobile xps rather than a greater number of vias.

In short I think your reasonable options come down to the following.

A. retire your lowest performance systems(under 1-1.5 ghz).
B. Retire groupings of your lowest remaining performers with a smaller number of amd mobile (or even desktop) xps in the 2400-2700xp range(in general I think the higher clock of the thoroughbred would be better than the lower clock but larger cache of the comparably named bartons). If up front cost is no object you could scale the processor up to the high end, but as you move up you will quickly pass the point where additional up front cost will outweigh operational savings on electrical cost. Any attempt to replace your farm with vias will either force a major performance hit or blow up your up front costs by demanding a large number of via chips(which when all is totaled up will not likely significantly reduce your total power demand).

I want to make it clear I have no connection or emotional attachment to amd or intel. I'm not a fanboy of any company or chip. I would love to see the vias outperform the big names on a performance for given power budget basis. My dream has been a solar powered farm of significant crunching power. We just aren't there yet. I simply feel you are setting yourself up for disappointment and significant expense which isn't likely to accomplish what you are hoping to accomplish.
The big name chips do consume more power, but they also perform at a level commensurate with the increased power required. While the via is impressive from an absolute power draw standpoint, it isn't impressive when you are trying to achieve low power draw while meeting high performance computational requirements.

I should have added for anyone who may be interested in mobile athlon xp that it is feasible to use them in readily available boards(unlike many other mobiles), but it isn't a sure thing that you can always just plug one into a mb that can hold the normal desktop chip. Search for some of the latest research as to which boards are best for this.

[MerePeer]

Nice reply in post above. I especially liked (B) suggestion of using a Mobile version.

[erk]

Yes, lots of good information there, I have investigated the Athlon-M situation, and the Australian distributor does not import them, unbelievable! They seem to be quite popular with overclockers, but my interest is in their power savings ability compared to normal Athlon XP's when folding all day long.

[HaloJones]

The Mobile Xps will use less pwer than a desktop equivalent but many of the advantages of the mobiles (such as slowing themselves down when not in use etc) won't apply to a 24/7 farm.

[jand]

if the beta client's the new client update, we should see more power savings whilst using mobile amd processors.
the power savings come in during uploads as you'll need to verify each data sent. so while it's doing that, you're powering down the cpu.
every little bit adds up

[erk]

This thread got me digging around looking from some other low power chips, there is a thing called an Opteron EE that does only 30watts.

http://www.amdboard.com/opteron_low_power.html

Very interesting.

[tpdooley]

Take a look at the prices of an Opteron HE or EE based system; complete with the registered ram.

It'd probably be cheaper to continue paying the electricity costs for a year and let their pricing drop before replacing your farm with them...

I was looking for sites that would list the actual heat production of the A64 line vs Bartons vs Opterons and couldn't find one. (They usually just list the quoted Max Thermal value). Perhaps if you could find a site that did power draw comparisons on the various systems with similar setups, you could use that to make replacement decisions.

[Grumpy]

There are 3 Opterons..Low Medium and Normal Power Models. An Opteron 240 Dual System with Integrated Rage will probably use less power than a Dual 2600MP and give you a lot more output. Me, I am looking at the 2500XPM and the soon to be released Pentium M 2 Ghz..these 2 would be a good pick, thought the 2 Ghz Dothan may hurt the pocket like the Opteron

[erk]

Originally posted by tpdooley
Take a look at the prices of an Opteron HE or EE based system; complete with the registered ram.

It'd probably be cheaper to continue paying the electricity costs for a year and let their pricing drop before replacing your farm with them...

I was looking for sites that would list the actual heat production of the A64 line vs Bartons vs Opterons and couldn't find one. (They usually just list the quoted Max Thermal value). Perhaps if you could find a site that did power draw comparisons on the various systems with similar setups, you could use that to make replacement decisions.

It is very hard to find accurate power measurements of CPU's. I am going to have to get a current meter and start measuring some of these chips when folding myself. What does "Max Thermal value" mean, is it just a point where the device starts to fail or a typical expected power usage?

[allenfinch]

The following terms are not all defined in exactly the same way, nor will they necessarily be precisely the same across manufacturers.
These are safe working definitions, not precisely accurate ones. In some cases you might be lucky enough to see explicit definitions included with numbers.

For our purposes however I think you can view max power levels such as maximum dissipation and thermal design power as worst case instruction mix under normal operation(voltages).

Max temp figures such as maximum case temp(referring to chip packaging, not ambient case temp), as do not exceed-red light-possible damage zone figures, NOT expected operating condition figures.

Although it would be interesting to do some tests of your own, I'm not sure that it would actually clarify things(unless you had a background in designing and conducting scientific tests), nor do I think trying to come up with exact real world numbers is worth the trouble(power dissipation varies with instruction mix, for example). Given the behavior of most systems running distributed computing projects relative to day to day use(temp behavior and so on), I think assuming these projects more closely approximate the worst case instruction mix(whatever actually makes up that mix) is a safe assumption and so I feel comparing max power dissipation levels is close enough.
I wouldn't fuss beyond this, unless you just enjoy such a thing and want to do it for pleasure. You are trying to do relative comparisons of chips, after all, not come up with design numbers. I don't think any definitional or testing variation across manufacturers invalidates using the provided numbers for relative comparison purposes.