Index  | Recent Threads  | Unanswered Threads  | Who's Active  | Guidelines  | Search
 

Quick Go »
No member browsing this thread
Thread Status: Active
Total posts in this thread: 4
[ Jump to Last Post ]
Post new Thread
Author
Previous Thread This topic has been viewed 41984 times and has 3 replies Next Thread
Hypernova
Master Cruncher
Audaces Fortuna Juvat ! Vaud - Switzerland
Joined: Dec 16, 2008
Post Count: 1908
Status: Offline
Project Badges:
Reply to this Post  Reply with Quote 
Comparative study 100 Mac Minis VS Custom Built

On another thread a member has proposed to build a WCG crunching system with 100 Mac mini's. Why not. The idea may seem interesting.
I have tried here to compare such a system versus mine with 10 custom built desktop machines.

1) Energy consumption
A) The Mac Mini System (MMS)
Physical cores of 100 minis = 200.
Consumption is 11 kW in total according to Apple specs. Measured at full crunching power this will be more as companies tend to be very conservative in their spec. Here I will be a real gentlemen and use the spec value. But I will add another 1kW cooling which is the minimum that must be considered. The total is 12kW
We have a theoretical ratio of 60 watts/core (2.53 Ghz/core Ram at 1066 Mhz)
And a theoretical ratio of 23.7 Watt/Ghz

B) The Solar System (SSI)
Physical cores of my 10 machines = 40 with HT 80. In terms of equivalent physical core it is 50.
I use 220 Watts per machine (measured at the wall socket at full WCG crunching) that is 2.2 kW in total. This power is also measured at 3.6 Ghz running speed for the cores and 1600 (8 machines) or 2000 Mhz (2 machines) running speed for the memory. For the sake of simplicity I will take 1600 Mhz value across all the machines.
That gives us a proven ratio of 44 watts/core (3.6 Ghz/core and Ram at 1600 Mhz).
It is also a proven ratio of 12 Watt/Ghz


Conclusion on Energy:
MMS consumption is 36% more per core than SSI.
MMS consumption is 97.5% more per Ghz, than SSI

2) Processing cost
A) The Mac Mini System
The fastest mini CPU is a 2.53 Ghz Intel Core 2 with memory running at 1066 Mhz. The cost of this mini is 800 US$ in the US. Let's suppose that we get a 10% rebate for 100 that would be 720 US$ each.

Total cost 72'000 US$.
Cost is 360 US$/core.
Cost is 142 US$/Ghz.

B) The Solar System
The CPU's are 8 with I7 950, one with I7 975X and one with I7 920. The cost of the I7 920 and 975X do average to a cost of two 950. All my machines are completely custom built by myself. The cost of the complete standard rig is 1'904 US$ but at European Prices. To be fair and compare with the minis I should use US prices which are allways lower than US prices. As the price differential seems low nowadays I will keep as is.

Total cost 19'040 US$
Cost is 380 US$/core
Cost is 105 US$/Ghz

Conclusion on cost
MMS cost is 5.2% less per core than SSI.
MMS cost is 35.2% more per Ghz than SSI

These two systems are to be used 24/7 so we must add a cost coefficient to take into account energy consumption costs.
The ratio MMS/SSI is 12/2.2 = 545%. To run the MMS the cost will be more than 5 times the cost for the SSI.

Final Conclusion
We must now add an additional factor and that is the Memory Bandwith difference between the QPI memory management design of the i7 950 against the Intel Core Duo 2 FSB older design. This is very important for crunching tasks where memory access speed is a limiting factor.
I estimate that the design difference should give on a par frequency in crunching tasks at least 30% better throughput for the i7. If some of you have a better estimation of this factor I am interested to know.
On that base we get the following final ratios:

MMS energy consumption is 126.7% more per same volume of work done than SSI
MMS cost is 45.7% more per same volume of work done than SSI

These two systems are to be used 24/7 so we must add a cost coefficient to take into account energy consumption costs.
The ratio MMS/SSI is 12/2.2 = 545%. To run the MMS the cost will be more than 5 times the cost for the SSI.
If corrected for the difference in performance design then the running costs per same volume of work are 7 times more for the MMS system.

It is clear that the Mac Mini is not the right solution. This conclusion is logical. I built my rigs to crunch. The Mac mini was never designed for that.

Now let's suppose that we buy a future Mac Mini with 4 cores at the same price and same speed and same consumption. We will end up with 400 cores.
I will suppose that on my side that I change all the CPU's with the new i7 hexacores with a version with the speed, consumption and price of the 950. I will end up with equivalent to 75 physical cores.

MMS will have a cost/core of 180 US$
SSI will have a cost/core of 253 US$ (40.5% more expensive than MMS)

MMS will have a cost/Ghz of 71 US$
SSI will have a cost/Ghz of 70.5 US$ (0.7% less expensive than MMS)

These values here over do not consider a coefficient for design performance as I have no idea what Apple will put as a quad processor. Will it be an FSB solution or QPI based.

Anyway three last things to consider:
The base investment for the MMS solution is 378% more expensive than SSI and will be at par with SSI in terms of performance but only with a future quad core solution.
The energy consumption for the MMS solution is at least 500% more expensive than the SSI.
And last but not least, managing 10 devices is a breeze compared to managing 100 devices.

Definitely powerful Multicore Desktops are for us home based crunchers the best investment and solution. smile
----------------------------------------

----------------------------------------
[Edit 2 times, last edit by Hypernova at Mar 14, 2010 7:40:59 PM]
[Mar 14, 2010 7:37:28 PM]   Link   Report threatening or abusive post: please login first  Go to top 
rilian
Veteran Cruncher
Ukraine - we rule!
Joined: Jun 17, 2007
Post Count: 1453
Status: Offline
Project Badges:
Reply to this Post  Reply with Quote 
Re: Comparative study 100 Mac Minis VS Custom Built

dont forget that buying macmini you pay xx% for brand name
----------------------------------------
[Mar 14, 2010 11:27:21 PM]   Link   Report threatening or abusive post: please login first  Go to top 
Former Member
Cruncher
Joined: May 22, 2018
Post Count: 0
Status: Offline
Reply to this Post  Reply with Quote 
Re: Comparative study 100 Mac Minis VS Custom Built

A small but crucial error in the calculation. The 110 Watt you mentioned is the maximum the power supply of the minis can deliver. They take much much less than that, under 50 Watt (full load, 2,53 GHz Version) as I read here (http://www.digital-world.de/artikel/_Tests/1260266/apple_mac_mini/3).
[Mar 22, 2010 9:43:30 PM]   Link   Report threatening or abusive post: please login first  Go to top 
Hypernova
Master Cruncher
Audaces Fortuna Juvat ! Vaud - Switzerland
Joined: Dec 16, 2008
Post Count: 1908
Status: Offline
Project Badges:
Reply to this Post  Reply with Quote 
Re: Comparative study 100 Mac Minis VS Custom Built

A small but crucial error in the calculation. The 110 Watt you mentioned is the maximum the power supply of the minis can deliver. They take much much less than that, under 50 Watt (full load, 2,53 GHz Version) as I read here (http://www.digital-world.de/artikel/_Tests/1260266/apple_mac_mini/3).


Thank you for this article darkeins, it is interesting. However the value of 50 Watts is to be taken with some care.
1) None of the tests to my knowledge load the multicore CPU full and for a sustained period like Prime95 as an example. Ideally we should have Boinc running on it and do the measurement.
2) We do not know if the 50 watts are an average between tests, or on a given test an average over time.

However if we consider the data from the article and that real consumption running Boinc WCG is half the specs then we improve a little the result for MMS, but it's consumption will still be 3.5 times more than the SSI solution.

If you look at the following post it seems that I have been on the contrary overly optimistic and that cooling power should be counted as much as processing power. We are back again to the ratio of 7.

https://secure.worldcommunitygrid.org/forums/...ad,28602_offset,20#271876

As we chat a new player has entered the arena and this is the i7 980X which improves the crunching capability of SSI by 60% under the existing power envelope. You can see my first test results on the following post

https://secure.worldcommunitygrid.org/forums/...d,28326_offset,120#272515

This factor of 7 is now even higher, we should be near 10 now. I know it is a moving target and the mac mini is definitely the wrong way to go.
----------------------------------------

[Mar 23, 2010 12:07:32 PM]   Link   Report threatening or abusive post: please login first  Go to top 
[ Jump to Last Post ]
Post new Thread