What is the most power efficient CPU out there? I'm looking over my system and I'm now just about convinced that my system is too powerful for my day-to-day needs (though it's great for any heavy duty encoding jobs).

When I mean power efficient, I mean, Which CPU makes the best use of the power it draws? The Via CPUs draw about 10-20W, but their use of that power is weak at best compared to the P4-M or even Pentium M, which draws as much power, but makes far better use of it.

Right now, the Mobile A64s are looking to be the most efficient, but are there others that are just as efficient? Can most A64 motherboards handle the 35W Mobile A64s?

OT: Perhaps SPCR should come up with the Power Efficiency Unit (PEU) as a function of a benchmark score divided by Power consumption.

Any thoughts?

That would be great, although the toughest part is determining the real power draw. AMD seems to be very conservative about the true power draw of their chips, which actually is one of their greatest assets. I would assume a mobile Barton or Mobile A64 would be the best candidate though at factory specified clock speeds and voltages.

Thing is.. different chips will be able to sustain different levels of undervolting while still holding a high clockspeed, and furthermore, different motherboards will provide different levels of undervoltability. That makes it really hard to just put out a chart and say that XYZ chip will score 123 units of measure per watt on the so-and-so benchmark.

Well, for my given problem, I'm willing to undervolt to achieve a better efficiency, so I wouldn't mind knowing the most efficient processor in terms of that. But undervoltability is never guaranteed.

Real world power shouldn't be terribly difficult to calculate given a Kill-A-Watt, which is our best method for calculating power consumption. If we choose a basic reference system with a set of reference motherboards, it shouldn't be terribly difficult to come up with some numbers, or at least a scale of which processor is more efficient than another.

Well lemme go into a little more detail then. First, I'll start with the data I've got:



Now since the voltage is a constant 1.55v across all clockspeeds (I can't undervolt below 1.55v) you can see that power draw goes up in linear proportion to clockspeed. Since 1000mhz is 65 watts in prime 95 and 500mhz is 53 watts in prime95 (12 watt drop) then dropping to 0mhz should bring it down to 40 or 41, which is coincidentally the measured draw at 500mhz when running cpuidle to halt the CPU. Thus, we start out knowing that the baseline is that the system draws 40 watts even when the CPU is pulling none.

Based on that, we can also then see that at 2037mhz it draws 86 watts in prime95, which is about as intensive as most folks' apps will be. I get the same draw from prime95 as from folding and sisandra CPU benchmarks, so it seems like a good figure. That works out to 46 watts drawn associated with the CPU. This is where it gets funky though; even if the PSU is a 70% efficient (which is being VERY generous) that would mean the CPU is only pulling about 32 watts DC. That's way lower than any of the CPU power calculators estimate based on AMD's TDP or whatever. In fact, the advertized draw even for the mobile 2600+ is 45 watts, and that's at 1.45v, not 1.55v.

So... what does it mean? Is AMD really, REALLY conservative in their power draw figures? Is there actually CPU power draw as part of that 40 watts even that I don't know about, or that can never be fully halted even with cpuidle? The bottom line is.. I just don't know. Even the full 86 watts though at 70% efficiency only works out to 60 watts DC drawn, which is actually what the CPU calculators say the CPU alone should be pulling at that voltage and clockspeed.

So.. if we at least just look at whole system performance and the Sandra CPU Arithmetic benchmark, it draws 86 watts and scored 7701 at 2037mhz. So that's just shy of 90 Sandra points per watt.

This is exactly the sort of information that I'd like to know or keep track of. Personally, given the samples to review, I would post and keep track of data like this for all processors. Unfortunately, I have no such samples. Perhaps I should request some for testing purposes, but from whom?

How much the CPU itself actually draws is not as important as keeping consistant or relatively consistant numbers, i.e. by using as much of the same hardware in every test. Everyone has to use a chipset + hard drive + video card (or onboard video), so small hardware discrepancies can be tolerated for the sake of generating interesting and useable numbers for everyone.

I've always wanted to see this exact thing. It's all fine and dandy that Via puts out less than ten watts, but how much processing power are you actually getting out of the chip?

I would think you could make a test setup using very similar components for AMD, Intel, and Via chips--the only thing that would have to be different is memory and motherboard, but perhaps a bit of testing could figure out just how much variance there is in memory and motherboards.

Then you could measure the AC power draw coming out of the wall, and compare that with some benchmark measurements.


There's a number of flaws I can think of this sytem already, but there has to be some way to get a comprehensive idea of what a given chip consumes verses its actual performance in various applications. The big question in my mind would be what benchmarks to use. P4s and Athlons both have their respective strengths and weaknesses--do you choose gaming? Encoding? Folding? Compiling? Some artificial benchmark? The only thing I can think of is you're going to have to do a number of tests for multiple benchmarks, reconcile that with the power draw during said activity, and then go from there. Needless to say, it'd be difficult to draw good conclusions if this wasn't performed with the most anal retentive methods possible.




I would think you'd have to leave the underclocking/undervolting for a separate test methodology, or create numbers off to the side for each system denoting how low of an undervolt/clock was achieved, since that's yet another unknown element in the whole process.

I, for one, would love to see what processor produces the most processing power for the smallest amount of heat. It'd be a massive project, but.....damn, would that be cool.

What if we had some repository of the average or typical consumption of various components, like hard drives, video cards, and other peripherals, and then we had some sort of database that took that into account to deduct those watts from the score, leaving just a CPU or something close to it. Then you find benchmarks that have a strong corrolation with CPU performance (weak corrolation with video or hard drive performance), and base scores on that? It wouldn't be as accurate as I'd like, but at least it'd give you a ball park figure...

I have 2 two shuttle zens running P4 3.06's underclocked to 2.56 ghz and undervolted. Both boxes together consume 106 watts while each are running two instances of folding@home. That's under 50 watts for each system when you discount the Reserator that is cooling them and the monitor that still draws current even when it's "off".

A P4 becomes very efficient when slightly underclocked and then undervolted as low as it can and still be stable. May or may not be as good as the AMD mobiles, but still very good. Even the Prescott can be tamed in this way, though not as good as a northwood.

What power supply do those use?

As a review site, SPCR should be obligated to at least think up of some benchmarks that would be comprehensive enough. Certainly if any hardware were sent MY way (*hint hint generous donors*) I would run an encoding test, a gaming test, and an artificial benchmark.

I think a bigger issue would be, "How do we get this value to have any meaning?" The greatest potential for silencing without sacrificing performance? What if it turns out that the PIII series was the most Power Efficient processor of all time? The point sorta gets lost.


Err... Underclocked and undervolted, aren't you better off attempting to undervolt a slower proc?

Undervolting will guarantee an increase in Power Efficiency. Underclocking doesn't make that same guarantee.

It's an external brick that will work only with the motherboard in the zen. It must be pretty efficient given that it is encased in plastic without any vent holes and is rated for 180 watts.

I'll give you a maybe on the first question. I haven't tried to undervolt a lower rated P4 to see how it compares to the underclock. It's definetly a good question.

Your last statement, however, is false. The underclock allows for a decrease in power consumption that is greater than the decrease in cpu output. For example, I have a 3.0C that at stock speed will undervolt down to 1.4 Vcore. The system will consume ~90 watts with a full cpu load. If, however, I underclock the cpu to 2.49 ghz it will undervolt to 1.15 Vcore. Underclocked the system consumes just slightly under 50 watts. Here are the percentages:

3.0 ghz to 2.49 ghz is a 17% drop.
90 watts to 50 watts is a 44.44% drop.

In a nut shell, a small underclock allows for a much greater undervolt, resulting in greater efficiency than undervolting alone.

Returning to your first question, "aren't you better off attempting to undervolt a slower proc?" To answer that question, I've also had the very same 3.0C running at 1.5 ghz and undervolted to .8375 Vcore. In this configuration the cpu pulls about 11 watts and I was able to run it for weeks, even folding@home, completely passive. Nothing but the stock ICE heat sink, without so much as a case fan or PSU fan. No suedo passive, really passive.

You ask yourself a question. Do you think you could get the Vcore low enough on a P4 rated for 1.5 ghz so that it would consume only 11 watts under full load @ 1.5 ghz and run completely passive? I think I know the answer. And I think I know the answer to your first question.

I have my p4 2.6C underclocked slightly to 2.15GHz and undervolted to 1.175 vcore. CPU Power says this is 33.8W or a 50% reduction in power. This way my computer runs folding at home and doesnt seem to use much more power than when its idle at stock speed (the psu fan in my antec 380s stays about 1350rpm load/underclocked/volted compared with 1300rpm idle/stock)

Pragmatically, going back to your original question, I suspect a multiplier unlocked XP (e.g. Mobile XP) might be the best bet. See http://www.silentpcreview.com/article164-page1.html for what I have done. This gives a range from 300MHz to probably over 2000MHz. So very low energy when word processing but fast enough for the latest games, etc when necessary.

It seems A64s will only go down to 800MHz http://forums.silentpcreview.com/viewtopic.php?t=13058&highlight=
at which the best is perhaps using 13W at 0.95V http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf. It may be that different BIOS might meant that the FSB could be dropped below 200MHz

P4s might be better. Copper how low can you get your P4? According to CPU Power if you can get it down to 0.8375V 600Mz it would be about the same power consumption as an XP at 1.1V 300MHz. How useable is it at that?

Mark

It worked for you and that's great and wonderful. But will underclocking guarantee greater undervolting in most cases (at least sufficient enough undervolting to compensate for the loss of performance)? That's what I mean by not guaranteed.


Probably not with a 1.5 GHz, but what about a 2.4 GHz Northwood underclocked/undervolted to the same degree?

I wouldn't imagine that a 600 mhz P4 would be useful at all. I can't get it that low on my machine to try. It was fine at 1.5 ghz though, at least for internet, light gaming, and office type stuff. A 100 mhz bus and .8250 Vcore are the lower side limits in the bios.

It has worked consistantly on 2 3.06's, a 3.0c, and a 3.0e. Not to mention we have tons of examples that demonstrate the same process, but in reverse - overclocking. Up the mhz by a few percentage points and the necessary Vcore and power rises exponentially. It's not an isolated hit and miss experience. It's the way cpu's work.



Your question was: "Err... Underclocked and undervolted, aren't you better off attempting to undervolt a slower proc?"

But since you now want to know about undervolting and underclocking a slower processor, I'd have to guess that if you underclocked both a 3.0c and a 2.4c to 1.5 ghz that the 3.0c would come out on top due to the exponential nature of power reduction with clock reduction. I'd guess that the 3.0c would be stable at a much lower Vcore than the 2.4c.

Alternatively, if you halved each of the processors, one at 1.5 and one at 1.2, I'd suspect that the 2.4c would come out on top, but at a lower clock speed.

I don't know if SPCR is "obligated," since the real focus of this site is silencing one's computer. I know that thermal output and ease of silencing are inexplicably entwined, but that isn't to say that a cpu spitting out fifty to eighty watts can't be cooled with some degree of silence.

As for the bigger issue, I agree--it is dubious to find meaning in those numbers. It may be that the most efficent cpu is an old 486 and all this is a waste of time. It may also be, however, that you can see the top ten, and the difference between more modern cpus becomes evident. Perhaps the test doesn't have to be all encompassing either--perhaps we take CPU's made in the last three years and limit the options.

It may be that the most efficient cpu is an older p3, but it may also be that the second most efficient cpu is a mobile athlon.

Eeek. Rereading what I wrote, I see that I was being a bit ambiguous. Would underclocking yield better efficiency with the lowered voltage? In otherwords, you were able to lower the voltage further by dropping the clock, but is the voltage/power reduction always greater than the performance reduction? And at what level does it become pointless? Where you have to drop the clock a lot in order to reduce the power a little?


That was kinda what I was getting at actually. i.e. wouldn't you be better off attempting an undervolt/underclock on a slightly slower (and considerably cheaper) processor than on a faster processor. More specifically, I was thinking, "Aren't you better off attempting to undervolt a 2.4GHz processor rather than underclock a 3GHz to a 2.4GHz and subsequently undervolt." I guess the answer is, "Not necessarily." As you've demonstrated.

Well, what inspired this thread was taking a look at the recommended section's CPU's ranked by noise/heat. For someone wanting to purchase a reasonably fast system, the first 5 entries are meaningless. In fact, a good deal of entries are meaningless since most of them are in the sub-GHz range. It would be nice for us to provide some numbers that are a little more useful to people who want to buy a new computer/CPU.

Also, by having a system of tests, we can properly evaluate the efficiency of both underclocking and undervolting (although I'm beginning to suspect that it is always efficient to do both).

8rdavcore scares me, but I'll consider it. Sorry that this post is so short, but I'm about to head out the door and I wanted to at least comment on your suggestion. I'll seriously consider it since price is one thing that's constantly on my mind.

That is exactly what we find. We also find the reverse when we increase clock speed by overclocking.



I don't know, but from experience we can see that the level is low enough to render it irrelevent as the processor would be so slow as to be virtually useless. What we do find is that just a little underclock yields excellent gains while leaving the cpu at a good clock speed. Hence the need to really start with a faster and more expensive cpu.




It would be cheaper to buy a slower cpu. But it's also cheaper to buy a regular Athlon 64 than it is to buy a mobile Athlon 64. Why? Because a 35W mobile Athon 64 2700 is for all intents and purposes an underclocked and volted Athon 64 3200. That is exaclty the reason mobile AMD's "overclock" so well.

Maybe if it were undervolted, but I doubt it. My Dell P3-1ghz consumed 61 watts A/C at idle and 81 watts when folding. Using sandra as the benchmark, it was equal in processing power to my Athlon running at just 850mhz, indicating the Athlon seems to do more work per clock cycle than the P3. Viewed from a watts standpoint, my Athlon system consumes about the same power under load at 1.7ghz, which would make it about twice as fast when both servers are consuming 80 watts. That's with my Athlon at 1.55v too! In theory, you should be able to run 1.4 or 1.35v at 1.7ghz for even lower power output, on conversely, at 1.45v you could probably generate that same 80 watts at 1.85 - 1.9ghz.

To see the value of underclocking and volting, look at the vlues reported by manero over on the SFF Tech forums.



His system with the highly regarded Pentium M pulls 41 watts idle and 60 watts loaded. My 3.06 @ 2.56 pulls about 46 with a 100% hyperthreaded cpu load. And I bet mine would whoop its ass in just about any catagory. The exception being that his graphics, I suspect, are considerably stronger than mine.

Given that the ati 9600 runs passive I can't believe it's adding all that much to his system's overall wattage, not compared to what my on-board ati 9100 adds as well.

AMD or Intel, a slight underclock combined with as much undervolt as possible brings about a very efficient machine.

It's an unfair comparison, isn't it? You're comparing an undervolted P4 with a Pentium M on stock voltage. What happens if you undervolt the Pentium M?

Good question. I know the Pentium M uses speed step, but I wonder if it can undervolt at full speed and still be stable. If it has to be underclocked to be undervolted then it would just be that much slower than the 3.06 @ 2.56 mhz.

I don't think the comparison is unfair if only because the Pentium M is regarded for its efficiency "out of the box." But I also think you are right in suggesting that a little tweaking is very likely to make the Pentium M even more efficient.

For the record, the Commell Pentium M board doesn't support speed step. So it's not going to be making the most efficient use of the Pentium M.

Those figures a little bit on the high side. Perhaps because of the FritzCard. German magazine C't got 37W idle figures with a standard Enermax 433W and 27W with a 60W Morex. The used the onboard VGA of the Lippert Thunderbird.

http://forums.silentpcreview.com/viewto ... highlight=

Hmm... I guess this discussion of processor efficiency is a lot deeper than I had thought. What with the diminishing returns on underclocking power efficiency not happening until they reach ridiculously low levels.

Back to my original intentions, I think it may be best if I purchase an XP-M processor and an 8rdavcore-compatible motherboard, as pod03 suggested. Presumably by getting that, my system can change dynamically, i.e. 300MHz when I'm posting to SPCR, and much faster when I play Homeworld 2. Basically, I don't want to have to reboot everytime I want/need a faster speed (thus the reason I've been skeptical about underclocking).

It CAN be changed dynamically, right?

Just comparing different CPU and how efficient they are compared to the wattage is very difficult as this thread has already prooven. The die size and use of heatspreaders or not also makes a difference making it even more complex.

What is more interesting is a few important heat steps:
1. CPU does not need a heatsink ~3Watts. Mostly important for embedded systems, aimed for systems with cost <$100.

2. CPU can be cooled with a cheap and light heatsink. ~10W. This is approximately where some VIA C3 CPU's are and a few AMD Geode CPU's.

3. Passively cooled CPU with just a larger and more expensive heatsink. ~20-25W

4. Aircooled CPU with noise level very low by SPCR standards. E.g. 30dBA. Hard to tell, but somewhere in the 60W area seems to be where we end up.

5. Cooling with more costly solutions. E.g. heatpipes to case, watercooling, peltiers etc. Noise level should still be acceptable. e.g. 40dBA. Here we end up somewhere over 150W.

The interesting thing is how fast CPU we can get and still be in one of these categories. For categories 1 and 2 price is rather important, so undervolting and underclocking an expencive CPU into this category proves very little.
Category 3 already has a special thread called something like "most wattage passively cooled".
Since category 4 and 5 need to have some dBA level to reach resonable sanity it becomes harder to put down the wattage we can take here. If will also be hard to judge a real winner since more people don't have equipment to measure properly. Judging if 3dBA less is better than 300MHz more and similar things complicate things even more.