Power efficiency

[Dirk Broer]

A while ago one of the postings on our forum read "hot=inefficient" and while the hot in this equation is wasted heat, one might ask oneself whether a 'hot' CPU is actually less efficient than a cool one, when efficiency is measured as to how much of the wattage spend goes into computing. I have compared all recent CPUs from both blue and green and there are some surprising results. How did I do this? I took for each CPU the number of threads that the CPU can spend on BOINC CPU projects and divided that by the TDP value for that CPU, so you can see how many Watts go into a single thread. Are AMD FX CPUs hot and therefore inefficient? Are i5s the total kings of cool? Not quite, and though Intel takes four of the best scores, we encounter the first -95 Watt- Bulldozers and Piledrivers on a shared 7th spot:

Ranking	Brand	Type	Socket	Model	Threads	TDP	Score
1	Intel	i7	1155	IvyBridge	8	77	9.63
2	Intel	i7	1150	Haswell	8	84	10.50
3	Intel	i7	2011	Ivy Bridge-E	12	130	10.83
3	Intel	i7	2011	Sandy Bridge-E	12	130	10.83
3	Intel	i7	1366	Gulftown	12	130	10.83
6	AMD	A8, A10	FM2	Richland T	4	45	11.25
7	AMD	FX	AM3+	Vishera	8	95	11.88
7	AMD	FX	AM3+	Zambezi	8	95	11.88
7	Intel	i7	1155	SandyBridge	8	95	11.88
7	Intel	i7	1156	Lynnfield	8	95	11.88
11	Intel	i7	2011	SandyBridge-EX	12	150	12.50
12	Intel	i3	1150	Haswell	4	54	13.50
13	Intel	i3	1155	IvyBridge	4	55	13.75
14	AMD	FX	AM3+	Vishera	8	125	15.63
15	AMD	FX	AM3+	Zambezi	8	125	15.63
16	AMD	FX	AM3+	Vishera	6	95	15.83
16	AMD	FX	AM3+	Zambezi	6	95	15.83
16	AMD	Phenom II X6	AM3	Thuban	6	95	15.83

The shared 19th place sees a lot of 65 Watt quads, amongst them the most efficient AMD APUs. The reknowned Intel i5 does not make the cut as it needs too much Watts to feed it's four threads, no matter what socket it is combined with. The Intel Pentium G (incarnation Clarkdale, Socket 1156) is amongst the worst performers, its mere 77 Watt is needed for just two threads. Even worse is AMDs Phenom II X2 Callisto: 80 Watt for just two threads.

[Dirk Broer]

Previous posting was soley about desktop CPUs and did not take into account the low-energy versions of the various CPUs, due to their unavailbility to the general public.
But now that we know that it is possible to build a system that actually performs and that uses only about 10 Watt per thread, let's see whether we can improve on that.

Approach 1.
We try to run as many threads as possible for a wattage as low as possible, where do we end up? When we use AMD CPUs, a mobo supporting four 16-core Opteron 6366HEs, like the Supermicro H8QGL-iF. That's 64 threads running on 4x 85= 340 Watt, that a power efficiency score of 5.31
Intel's best offering comes when you stick four Xeon E5 4640 V2's in a Supermicro X9QR-7TF+: you can run 96 threads for a tdp of 4x95 = 380Watt, that just under 4Watt per thread! But you have to bring buckets of cash in order to buy such a combination...


Approach 2.
As approach 1 is frightfully expensive, we will attack the problem from another angle: running a thread on as few Watts a possible.
The ARM-based Rapberry Pi springs into mind, using just 3.5 watts for a thread. It takes some 51 hours to crunch a e.g. Asteroids WU with it though. A i7-3770 will do that in 2 hours,
so consuming less power in that time -and it will even crunch another seven at the same time... An 8-core FX will have that same WU finished in a little under 3 hours, also doing 7 others at the same time.
We can increase the output of a Rasberry-Pi system by making a cluster of it, but at the cost of added power. A RaspiCluster doing 8 WUs at a time wil need 8x 3.5= 28 Watts, and is stil 51 hours under way producing results. AMD CPUs with a likewise power consumption are the Z-60 (4.5 Watt) and A4-1200 (3.9 Watt). Both are dual core CPUs, but I lack BOINC data bout them.
There is also a quad AMD A6-1450 that runs on 8 Watt

[Jason1478963]

There is also the older Opteron 2419 that does pretty well if you get the 40 watt version at 6 cores a piece and running 2 per board. This is a nice machine for badges at WCG and the points it was getting at WCG is similar to a x6 1090T.

[Dirk Broer]

Funny, read an article about the real difference between a normal CPU and it's so-called energy efficient version. It appears that the main difference is the lowered clock of the energy efficient version and that when you run them both at the same speed no difference is to be measured. Oke, it was about the Celeron G1610 and 1610T, but I discovered that I can kick my AMD A8-3820 up to A8-3850/70 level too -at the cost of running it at 99 Watt instead of the normal 65 Watt.

Bottom line: You do not need to pay more for an energy efficient version. Just buy the cheaper standard CPU and when you want to cut on your power bill, lower the clock and voltage to the level of it's energy-efficient brother.

[drezha]

I think you've gone about this the wrong way as one AMD core doesn't quite equal one Intel core. Be nice if it was.

Ideally, you'd need to compare them at Points Per Watt (or basically the performance of whatever you're measuring Vs power). Whilst it's nice to see here, unless you compare what you're going to use them for, it's not much help.

Regarding the clusters of Pi's, there's a nice article here that compares the efficiency of an ARM cluster against an i7 and AMD offerings.

Normally, I would suggest when calculating the power draw to look at it's minimum load and the estimated time at max - almost all CPU's now support dynamic frequncies and underclock if not being used. If you're intrested in day to day usage, look at the time to finish a task - if it takes 2 hours on an i7 (77W) but 52 on a Pi (3.5W), then the Pi will use 178.5W to do the task the i7 uses 144W to do. However, looking at crunching, if you're 24/7 at max, then you need to look at the above - performance per watt, rather than the watts per thread.

[Dirk Broer]

I had specially confined myself to watts per thread because points per watt is also dependent on OS and Project.
I thus have purely looked at how much it costs to keep a given CPU and its threads running.
You can do a lookup in the WUProp tables (select WUProp in the boinc manager under 'projects' and select 'results')
to see how much credit a given CPU thread will give you per project and OS, 32 or 64 bit.
The resulting table would be too big to be meaningful in a posting on this forum.
We can, however, have a discussion on the CPU efficiency per project
This will ensure a nice summer full of discussions about which new CPUs to buy, when money is available...

[Nflight]

Very interesting analysis keep up the good work, I like what I am seeing. I just might be in the midst of a purchase in the near future! This is valuable if your a penny pincher like i am.

[Nflight]

Here is a link That could reduce the energy usage for Chips and Chipsets.

[colbyslanbia]

well haswell is the new king of the processor it's have everything what a pure processor wants which is high processing power and lower power consumption and it hits the market pretty well and i'm sure AMD is also preparing for a new beast for coming up

[Dirk Broer]

I hope the delay of Steamroller is used to produce it in an 8-core APU version.
Such a processor I can even forgive a 100Watt tdp, as it would have the graphical power of a 100+Watt GPU included
(4 core Kaveri is already rumored to have the graphical power of a HD 7770/7790, a new 8-core FX APU would double that)