Friday, August 18, 2006

Athlon 64 X2 4200+ at 25C under load

Interesting test at LegitReviews on Athlon 64 X2 4200+ (2.2GHZ). They lowered the voltage from 1.35v to 1.125v while keeping the same clockspeed. As a result, power consumption dropped by 28 watts (from 218 to 190) under load.

A number of Intelers asked me to tell them what this result means (which was obvious to me), so let me elaborate to them, using high school physics and elementary school math. Correctly applying these physics and math knowledge requires IQ. So the following can prove at least I have a high school diploma and sufficient IQ to interpret this LegitReview result. It will also prove that Intelers do not have sufficient IQ to do the same. But it can't prove my Ph. D -- that's actually very hard to do. However, I don't have the need to prove that to Intelers any way. It's like a five star general doesn't have to prove his rank to a foot soldier.

At a fixed frequency, both the CMOS switching power and resistive power are proportional to V^2. We have (1.125/1.35)^2 = 0.694

28/(1-0.694) = 91.6

Using an AC-DC conversion efficiency of 83%, we found AMD AM2 (89 watt)'s power dissipation under load at default voltage to be

91.6 *0.83 = 76 watts

Thus, when under volted to 1.125 v, the CPU power is 53 watts. One of our AMDers asserted that lowering vcore made the chips an EE version. He was absolutely right, demonstrating the inherent intelligence advantage of an AMDer over an Inteler.

Intel's Conroe is rated at 65 watts, but that was typical power, not max. Furthermore, Intel's Conroe chipset consumes about 20 watts TDP.

23 Comments:

Anonymous Anonymous said...

I guess they turned it into an EE chip by undervolting. That said any chip, Intel or AMD, can be undervolted so it isn't particularly revolutionary.

10:38 AM, August 18, 2006  
Blogger Sharikou, Ph. D said...

I guess they turned it into an EE chip by undervolting.

I think this is a simple trick that everyone can do-- drop the voltage a little bit and save some energy. If one is afraid of OCing and increasing voltage, one should not be afraid of lowering voltage.

10:43 AM, August 18, 2006  
Anonymous Anonymous said...

Well, I've done this for years on my laptop using Notebook Hardware Control. If one is afraid of instabilitu one should do massive stability testing before running 24/7.

11:22 AM, August 18, 2006  
Anonymous Zetro said...

I agree 100% Sharikou,

Not every processor can be undervolted obviously as a general rule of thumb. If your processor overclocks better chances are it undervolts better as well!

Undervolting is not dangerous to a processor at all. All that happens is well your computer is unstable when the VCORE is to low.

To the first poster, it is not revolutionary but with my experience with the Pentium 4 chips they could not be undervolted very much. My short experience with Core 2's have only been with overclocking so I will have to explore this further.

12:01 PM, August 18, 2006  
Anonymous Anonymous said...

Wow did you take physics the E&M part in Freshman before you failed out of colledge.

Power=V*I. Did you realize lower voltage would make the left hand side equation go down..

Now why didn't it clock slower. That is obvious to.

PhD pretender want to tell everyone vs. twsit the obvious?

3:44 PM, August 18, 2006  
Anonymous Edward said...

I don't think SuperPi or Prime95 is any indication of system stability, though.

The biggest and most complex part of modern microprocessors is the control logic of the datapath for superscalar execution, such as branch prediction, out-of-order load and store, cache set and way prediction, etc. Those mathematical benchmarks, however, do not stress these logic very much.

IMO, they should've compiled with the linux kernel and KDE/Gnome with -O3 optimization, and fully test the correctness of the results, before declaring the boxes "solid stable."

5:32 PM, August 18, 2006  
Anonymous Edward said...

"Power=V*I. Did you realize lower voltage would make the left hand side equation go down.. "

For digital ICs, current is a super-linear function of clock frequency. Thus lowering the voltage itself does not necessarily induce power reduction unless the clock frequency remains the same or becomes lower. However, using the same (stock) clock frequency inevitably reduces stability margin the system has - if it was not watercooled, or if it was put inside a closed case, or if it was running on a hot day, things could go bad without warning.

And frankly I don't think their "idle" comparison is fair. Apparently they didn't use cool&quiet for the regular-voltaged processor.

5:58 PM, August 18, 2006  
Anonymous Anonymous said...

The funny thing for Sun is that an high efficiency Opteron basically gives you the same performance as their T1000/T2000 boxes -- for half the price.

Sun may put out some Opteron boxes but they have not realized yet that there is no future for SPARC.

CoolThreads don't run cool and is no better than an Opteron. The total cost of ownership is outrageous.

Of course, no one ever accused Sun of having all their marbles...

6:03 PM, August 18, 2006  
Blogger pointer said...

I think this is a simple trick that everyone can do-- drop the voltage a little bit and save some energy. If one is afraid of OCing and increasing voltage, one should not be afraid of lowering voltage.

This can be tried out, but seriously need to follow what Edward said:
fully test the correctness of the results, before declaring the boxes "solid stable.

Not sure if you have a chance to shmoo (characterize) a chip before or not. Statistically, most chips work at the suggested voltage range (higher yield), with smaller portion of those shows better voltage toleration. From personal experience, the chip operating at out of suggested ranged voltage can pass 1000 tests but fail one. but sometime you will be in luck because the failing test is compensated by a better tolerance external component. take ATI chipset for example, it has not so good SATA electrical signal by design (read as not per SATA electical spec, but follow the digital one ), it still work normally because the normal harddisk under normal environment can cope with it. i strongly believe it might fail under extreme condition such as extreme cold or hot environment.

8:19 PM, August 18, 2006  
Anonymous Ho Ho said...

Is it just me or the link to the review does not work? All I get is 404

Anyway, temperatures mean almost nothing, you can't use only that to compare heat dissipation with different CPU.

E.g if you have a CPU with die size of 50mm2 that is 30C and CPU with 100mm2 with the same temperature they do not produce the same amount of heat. The smaller one produces quite a bit less but all the energy is consentrated on a smaller area.


"My short experience with Core 2's have only been with overclocking so I will have to explore this further."

I don't know about downclocking them either but I know several people who run their 2.13G 6400's at 3.2G+ with default voltage. I would assume that as they OC that well they work rather nicely with lowered voltage too.

Mostly the limiting factor of OC'ing C2D's is the motherboard. Those low-end Core2's need 450+MHz FSB to reach their maximum OC. That means >1.8GHz FSB. Too bad such motherboards cost so much.

2:23 AM, August 19, 2006  
Anonymous Zetro said...

Well I am cranking a 6400 currently on a Gigabyte DS3 and it is going quite well as you said. It runs at 3.2 on the stock vcore and with a little bump (1.3625 from 1.325) it happily does 3.5 with 24+ hours of Prime and 3dMark01 stability.

I dropped it down to stock and I have been fiddling with the voltages, currently with the board only having 0.995 as the lowest selectable voltage I have been able to get close to a volt and it "feels" stable. Since it has only been 10 hours since I have started fiddling I obviously have not stressed tested enough to be sure.

7:31 AM, August 19, 2006  
Anonymous Anonymous said...

Since the PhD pretender can't aswner the riddle

There are two elements to capability to over clock. Notice how the PhD pretender never provide the had data but only partially teases in the hope of making AMD look good and INTEL look bad. Never tells the whole story.

THe Pretender would make a good AMD markerter or Landis/Bonds PR man. Lies like the best of them.

Lower the voltage reduces static power since P=VI.. I will also go down due to lower V driven gate oxide and transistor leakage. This helps compensate for running a higher clock and more dynamic power from the same capacitance switching faster.

All chips have some stability head room to overclock. Kind of like all cars can run 110 MPH to maybe even 140 MPH but aren't spec'd to run there...

Its really too bad that the PhD pretender doesn't become more level headed in his prospective instead of always sucking AMD. You'd get more credibility if you didn't plain lie and distort facts.

Why don't you have a more level propspective? Why.. you have nothing to lose and everything to gain..

Got fired by INTEL or something... sour grapes... didn't get a job at INTEL? Tell us! To gutless?

Whats the reason. You spend enough of your life blogging. WHy not get some real credibility. The potential is there.. the attitude is so sad...

9:14 AM, August 19, 2006  
Anonymous Anonymous said...

Judging by Mikes many Maricopa’s, Dell must have learned something from the good folks at Alienware after there purchase.

It’s getting extremely painful to watch. Accounting problems, battery explosions, poor customer service issues, profits down, shirking market share and pissed off share holders. Christmas season looks questionable without the release of Vista. Low margins and the Intel channels stuffed to the max makes 4th Q 2006 bleak for Dell. Increased pressure from HP and others is compounding Dell’s missteps.

Dell’s only ace in the hole is its business model, but with a huge cash burn-rate unchecked I think the AMD deal may be too little to late.

I hope this great American institution has a quick fix I don’t know about.

11:56 PM, August 19, 2006  
Blogger N4CR said...

Good to see, but they were running WC so ain't a totally 'normal' figure. Stock HSF would be right or maybe I just misread it? Tired today..

Not to dissimilar to what they are doing with the 'EE' models, same thing. I can remember undervolting my dads 130nm hammer and being impressed. Also when it ran at just above room temps with C 'n C running.. sheesh.

My venice did well too, my current oppytron 165 runs 2.8 at stock volts. That's a 1ghz oc at stock volts, not any different to an early conroe funnily enough ;) High end air though.

If only CnC worked with OC's then i'd be a happy man (and not running CPDN all the time).

1:09 AM, August 20, 2006  
Anonymous Anonymous said...

Whats the reason. You spend enough of your life blogging. WHy not get some real credibility. The potential is there.. the attitude is so sad...

That would require a real PhD; half-baked knowledge of yield and zero knowledge of computer architecture won't help. Seeing how ignorant Sharikou is on the simple facts of Comp Arch and VLSI, I don't think he has PhD in any field.

Heck, my PhD is in Game Theory, and I know more about Comp Arch and VLSI than this guy does!!

10:10 AM, August 20, 2006  
Blogger Sharikou, Ph. D said...

That would require a real PhD; half-baked knowledge of yield and zero knowledge of computer architecture won't help. Seeing how ignorant Sharikou is on the simple facts of Comp Arch and VLSI, I don't think he has PhD in any field.

I updated the article with some high school physics. You know, Eistein derived special relativity with no more than high school physics and math, which was kinda amazing.

But your comments completely missed the point, this test has zero with computer architecture or VLSI design, it's a question of device physics, an area in which Dr. Craig B was an expert and seem have forgotten after becoming CEO.

4:54 PM, August 20, 2006  
Anonymous Edward said...

"Heck, my PhD is in Game Theory, and I know more about Comp Arch and VLSI than this guy does!!"

YOU MUST BE KIDDING!! When does economics have anything to do with computer architecture and VLSI?

Since I started to visit this blog I've been amazed by people's ability to make up claims with seemingly overwheming confidence without any slight bit of proof.

Just because you are a PC enthusiast, or you follow the PC industry stock market, or you visit those amateur benchmarketing sites, doesn't make you a bit more proficient in computer architecture nor in VLSI.

9:44 PM, August 20, 2006  
Anonymous Edward said...

"Lower the voltage reduces static power since P=VI.. I will also go down due to lower V driven gate oxide and transistor leakage. This helps compensate for running a higher clock and more dynamic power from the same capacitance switching faster."

You don't mix up static and dynamic power together. One does not "compensate" the other.

Dynamic power is due to switching, which is affected by frequency and voltage in a complicated way. One needs SPICE simulation to estimate dynamic power, but it is not impossible for a multistage circuit to consume more dynamic power at a lower supply voltage (if the output of the previous-stage transistors rises too slowly, due to its low supply voltage, it can put the later-stage transistors in the active region for too long and uses up more electric energy). Running it at a higher frequency only makes things worse.

Lower voltage however will definitely lower static power - which is the whole point of speedstep and cool&quiet. Of course, in such cases the clock is also reduced, but since the processor is supposedly under low load, there would be few gate switching beside the clock tree.

10:32 PM, August 20, 2006  
Anonymous Anonymous said...

YOU MUST BE KIDDING!! When does economics have anything to do with computer architecture and VLSI?

Nash was a mathematician, not Economist. You would be surprised to see how many Electrical Engineering PhDs are given out in Game Theory. You know, there is a whole different world of Communications and Control Theory.

12:06 PM, August 21, 2006  
Anonymous Anonymous said...

Let's not mention the fact that Conroe is capable of being undervolted to almost 1.0V. Seriously, you cannot compare an undervolted CPU to a stock CPU and say the undervolted one provides more performance/watt. Undervolt that Conroe as well, THEN we can talk. ;)

3:06 PM, August 21, 2006  
Anonymous Anonymous said...

Edward seems to be pwning everyone here.

Anyone else care to step up?

3:31 PM, August 21, 2006  
Anonymous Anonymous said...

Sharikou, you have just stooped to new lows. Comparing an undervolted X2 to a stock C2D? Err, hello, ever heard the phrase apples to apples?

Overall system power is almost idential between an X2 3800+ and E6300.

Source:
http://techreport.com/reviews/2006q3/e6300-vs-sff/index.x?pg=15

Performance/watt is obviously in the E6300's favour, I want to see how you're gonna put a spin on that one.

6:25 PM, August 21, 2006  
Anonymous Anonymous said...

Does Conroe hrottle?
http://www.digit-life.com/articles2/cpu/intel-core2-duo-x6800.html

"So, that's what the extreme Core 2 Duo processor is like: rather fast (no one has ever doubted that) and... rather hot. This hot summer has positively spoilt the triumph of the new Intel processor. In return, we have learnt lots of interesting things about throttling in a CPU with the new core — a simple fact of its presence there and a whole palette of results demonstrated by a slowed-down overheated processor in comparison with a more or less cooled processor."

4:50 AM, August 23, 2006  

Post a Comment

Links to this post:

Create a Link

<< Home