Ok then.

 

There has been incredible amount of speculation floating around the web about how much power this or that processor consume. Unfortunately, only a very select few people have actually taken the time and measured something. Then again, those who have, have used very, very inaccurate means of doing so. The only credible measurements we have so far have measured the power consumed at the wall outlet.

 

While it is good enough for showing that K8 cores consume way less than P4 cores for similar kinds of performance, it tells very little about exact number of watts consumed by the CPU itself in different load conditions.

 

Of course those who haven’t measured anything, still brandish TDP figures back and forth, with many not really grasping what TDP actually means. And use TDP figures as some strange kind of weapons.

 

Ever since the gossip about K8 power consumption started way back when using a K8 based computer was just a wet dream of mine and everybody else I vowed to myself that I will measure how much exactly does this beast consume. Then again, living in a somewhat far away place where living standards aren’t really comparable to that of G7 countries, I had to wait till I got my hands on some. Even though some select A64 and Opteron rigs floated around me, most of the owners really weren’t much into what I wanted to do. So even now, I got my hands on the very lowest end of K8 line, the Sempron 3100+. The first one. There are actually 17 more sitting in boxes around me that I am going to unpack and set up soon. But before that I was very curious, and just wanted to see the numbers.

 

So, even though I can measure the lowest end of K8 line, I can actually do a pretty comprehensive study on the variability between chips.

 

 

The technique.

 

So, how do I measure the consumed power? No, I don’t think using some contorted water cooling systems is going to help much. What I did is much simpler. I just measure the current flowing out of a PSU, on all 3 rails that actually mean something, +5, +12 and +3.3V rails. After thinking about it, I didn’t even add any resistors in line with power rails – the power supply contains a filter choke in series with all rails and all these chokes have a very finite and fixed DC resistance and the voltage drop across this DC resistance can be measured very accurately.

 

So, to measure this voltage drop, I connected a RC low-pass filter network to filter the AC noise out of measurement voltage from every choke:

 

 

After that I had to find out the DC resistance values of all these 3 chokes. To do that, I used various “test loads” out of several resistors in parallel – the resistance of which I could accurately measure connected those to the lines and measured the voltage drops over those three chokes. I used three different “test loads” and averaged the calculated resistance of the chokes. The resistances turned out to be 4.7milliohms in 5V filter, 28.8milliohms in 12V filter and 12.6milliohms in 3.3V filter. The voltage drops across these chokes was large enough (tens of millivolts most of the time) that it could be measured with only a few % accuracy.

 

So, after rigging it all up, I made quite a few tests, with the aim of isolating separate power consuming items as best as I could. What I did not do, unfortunately, somewhat due to laziness on my part was that I did not separate the current drawn by the separate 12V connector used by CPU core regulator. I hoped, which in hindsight turned out to be a pretty valid hope, that very, very few items in my simple office computer actually consume anything from 12V rail beside CPU voltage regulator and a HDD.

 

So, here come the numbers. I show the currents drawn from rails, and calculate the total power consumed by these three rails together. I did not try to measure the consumption at the wall socket – so I don’t have any idea of how efficient power supply is with these kinds of loads. What should be kept in mind however is that I don’t have any “unnecessary” loads in my computer. I don’t have any high performance graphics cards, I just use an integrated video controller of a MSI K8MM-ILSR uATX mobo for measurements. The hard drive used was Seagate 7200.7 80GB Serial ATA drive. There was also 512MB of Corsair PC3200 memory. Integrated video ran in 1024x768x32bpp @ 75Hz feeding a 15” LCD monitor. I was actually unable to see any effect of keyboard and mouse on the power consumed from these three rails, which I guess means that these items run off a +5VSB rail (which I did not measure). The OS is WinXP Pro SP2.

 

 

The numbers.

 

I @ +5V

I @ +12V

P @ +12V

I @ +3.3V

P total

Description

0.79 A

0.78 A

9.3 W

3.47 A

24.7 W

Sempron 1.0GHz @ 1.1V CnQ, Windows Idle

0.79 A

1.12 A

13.4 W

3.48 A

28.9 W

Sempron 1.8GHz @ 1.4V, Windows Idle

0.79 A

2.41 A

29.0 W

3.75 A

45.3 W

Sempron 1.8GHz @ 1.4V in BIOS menu

0.79 A

3.24 A

38.9 W

3.47 A

54.3 W

Sempron 1.8GHz @ 1.4V, 1000000! in WinCalc

0.79 A

3.82 A

45.8 W

3.47 A

61.2 W

Sempron 1.8GHz @ 1.4V, BurnK7

 

Let me just say before commenting on anything else – the first thing I noticed with this computer was that it felt cold. And I mean cold. A small uATX case (from Chieftec), but a quite hefty cooler – Thermaltake SilentBoost, which I promptly “underclocked” down from 2500rpm to some 1750rpm by way of adding a 4.2V zener in series with a fan. PSU fan turns at roughly 1080rpm even after after 30minutes of 1000000! in WinCalc and barely moving air from the PSU still feels only slightly warm.

 

I have not felt the kind of “coldness” since the days of Coppermine Celerons running below 1GHz. Even Duron 1600 runs quite a bit hotter than this as you will see soon.

 

But back to the Sempron. The above testbed still had a harddrive humming in all these tests. I tried to make it stop using power management but all my attempts were futile. Couldn’t force it to any shorter waiting times than 3min and it was simply impossible to make it sit still for 3min. Killed all processes I could think of, though one Mcaffee virus shield related process didn’t disappear, regardless of how many times I tried to kill it. No messages, no nothing, the process stayed on.

 

Anyhow. It is very obvious that the above +12V power rail numbers include a humming HDD which does draw something from +12V. But now, the power consumption figure measured in BIOS menu comes to the rescue:

 

I @ +5V

I @ +12V

P @ +12V

I @ +3.3V

P total

Description

0.79 A

2.41 A

29.0 W

3.75 A

45.3 W

Sempron 1.8GHz @ 1.4V in BIOS menu

0.11 A

2.22 A

26.6 W

3.69 A

39.3 W

Sempron 1.8GHz @ 1.4V in BIOS, no HDD

0.68 A

0.19 A

2.3 W

?

5.7 W

deduced idle HDD power consumption

 

The current measured in 3.3V line varied somewhat over time. At times it was very stable, at times if was slightly larger and varied. I don’t know why that is. It mostly tended to be slightly higher when I was in BIOS, but then sometimes it wasn’t. The currents measured from +5V and +12V rails were very repeatable. Though +12V current rose somewhat (+2%) over time as the computer warmed up a little. That is still a very little change.

 

When we now subtract HDD power consumption numbers from the first table we get:

 

I @ +12V

P @ +12V

“Pfans”

“P @ Vcore”

Description

0.59 A

7.1 W

1.8 W

4.6 W

Sempron 1.0GHz @ 1.1V CnQ, Windows Idle

0.93 A

11.1 W

1.8 W

8.4 W

Sempron 1.8GHz @ 1.4V, Windows Idle

2.22 A

26.7 W

1.8 W

22.4 W

Sempron 1.8GHz @ 1.4V in BIOS menu

3.05 A

36.6 W

1.8 W

31.3 W

Sempron 1.8GHz @ 1.4V, 1000000! in WinCalc

3.63 A

43.6 W

1.8 W

37.6 W

Sempron 1.8GHz @ 1.4V, BurnK7

 

In this table I also calculated a hypothetical Vcore power consumed by this particular Sempron 3100+, based on a bunch of assumptions:

 

-     that a Vcore regulator has 90% efficiency (which I think this 2phase regulator could be pretty close to)

 

-     that the two fans running in this system, which both consume power from +12V rail take 1.8W between both of them. They are running quite slowly, roughly 1750rpm both. But even the CPU fan socket is in a practically unreachable spot. Long story, the small uATX case is really cramped given the large SilentBoost cooler.

 

-     that nothing else consumes anything from +12V rail.

 

Of course this power forms only the Vcore part of dissipated power of a Sempron. Besides Vcore, the CPU also has an I/O voltage which I am not measuring at the moment. That seems to be included in the +3.3V “bill”, though it is difficult to measure how much exactly is consumed. Considering that I/O power is quoted of being 2.2W regardless of P-state in Processor Power and Thermal Data Sheet, I think it is safe to say that this particular CPU never consumes (and dissipates) more than roughly 41-42W.

 

The CPU has a 62W TDP, so we see now first hand with what kind of power budget margins AMD is operating at. Makes me wonder what kind of ridiculously low numbers I am going to see with 90nm cores.

 

Before somebody says that these numbers seem impossibly low, I think it is wise to ponder about huge power consumption differences between A64 and Prescott chips when measured at the wall. The K8 numbers have to be incredibly low if the Prescott does indeed fulfil its ca 100W TDP promise.

 

 

The Duron 1600.

 

The other, quite comparable computer (actually lot of) I have here are Duron 1600 based office desktops, in similar Chieftec uATX case, EPOX 8KMM3I uATX mobo, also with (horribly flickering) integrated video. They usually run with an AGP card to make up for that, but I removed it now for power measurement purposes. The also box contains 256MB of Corsair PC2700 and a Seagate 7200.7 40GB drive. As is obvious from below, most (but not all!) K7 boards feed the CPU from +5V rail. This one definitely does. Duron 1.6GHz has a Vcore of 1.5V. The CPU fan runs at 2500rpm. OS is Win2k.

 

I @ +5V

P @ +5V

I @ +12V

I @ +3.3V

P total

Description

1.62 A

8.1 W

0.55 A

3.33 A

25.7 W

Duron 1.6GHz, Windows idle w/bus disconnect

6.34 A

31.7 W

0.55 A

3.37 A

49.4 W

Duron 1.6GHz, Windows idle (default)

7.02 A

35.1 W

0.54 A

3.68 A

53.8 W

Duron 1.6GHz in BIOS menu

8.53 A

42.7 W

0.55 A

3.73 A

61.6 W

Duron 1.6GHz, 1000000! in WinCalc

10.3 A

51.5 W

0.55 A

3.38 A

69.3 W

Duron 1.6GHz, BurnK7

 

One thing that has always bothered me with K7 based systems is that most boards feed the CPU from +5V rail and since most power supplies can’t take huge current fluctuations on +5V rail without system instability and therefore most mobos have disabled “bus disconnect” during Windows idle. Or in other words, they don’t actually “let” the CPU “go to sleep” at all. This means that K7 based systems are way hotter than older PIII based and newer K8 based systems. I don’t know exactly how much Willamette and Northwood systems consume on Idle, but I think it is also less than what any K7 systems consume (because of the lack of true idle). Fortunately, by choosing quality power supplies it is actually possible to run them with “bus disconnect”, which I invariably do with no significant ill effects (instabilities).

 

I also measured the power consumption without HDD and CPU fan in the BIOS menu:

 

I @ +5V

P @ +5V

I @ +12V

I @ +3.3V

P total

Description

6.26 A

31.3 W

0.20 A

3.71 A

45.9 W

Duron 1.6GHz in BIOS menu, no HDD, no fan

6.26 A

31.3 W

0.32 A

3.71 A

47.4 W

Duron 1.6GHz in BIOS menu, no HDD

7.02 A

35.1 W

0.54 A

3.68 A

53.8 W

Duron 1.6GHz in BIOS menu

 

 

0.12 A

 

1.4 W

One 2500rpm CoolerMaster fan

0.76 A

3.8 W

2.64 A

?

6.4 W

Seagate 7200.7 40G ATA133

 

Subtracting the above HDD consumption and 0.11A more from the whole +5V rail consumption, and also assuming 90% regulator efficiency we get:

 

I @ +5V

P @ +5V

“P @ Vcore”

Description

0.75 A

3.7 W

3.4 W

Duron 1.6GHz, Windows idle w/bus disconnect

5.47 A

27.4 W

24.6 W

Duron 1.6GHz, Windows idle (default)

6.15 A

30.8 W

27.7 W

Duron 1.6GHz in BIOS menu

7.66 A

38.3 W

34.5 W

Duron 1.6GHz, 1000000! in WinCalc

9.43 A

47.1 W

42.4 W

Duron 1.6GHz, BurnK7

 

Why 0.11A more? It is difficult to believe that absolutely nothing else consumes +5V except a CPU regulator and a HDD. Come to think of it, at least a floppy was also connected to +5V. And in my Sempron test, there seemed to be a 0.11A residual power drain on +5V rail, when hard drive was disconnected (and CPU just doesn’t use +5V in K8 systems).

 

Typical thermal power is quoted as 48W for this part and maximum thermal power as 57W. It is quite obvious that K7 comes quite a bit closer to its advertised maximum power than K8 does. 75% from TDP in case of this Duron and 60% in case of my first Sempron 3100+.

 

Ok, the numbers from Duron 1600 don’t seem that much higher than Sepmron 3100+ numbers, but there is one one other thing to keep in mind: PSU efficiency is quite a bit lower on +5V rail currents than it is on +12V rail currents. That is because the rectifier diodes have ca >0.5V voltage drop anyway, and that creates relatively larger losses on lower voltage lines. So, at the wall the power consumption difference would be noticeably bigger between Sempron 3100+ and Duron 1600, than the way I measured here, in +5V/+12V rails.

 

 

The other measurements.

 

Coppermine Celeron 800 running on a Tyan Tsunami mobo (440BX chipset), with 3 sticks of single sided 64MB PC100 SDRAM, a 80G Maxtor DiamondMax drive with 8MB buffer. Video card is ancient (but lovely) ATI 3Dcharger with 4MB of VRAM and a SMC 1255TX network card. Running a Win2k.

 

I @ +5V

P @ +5V

I @ +12V

I @ +3.3V

P total

Description

2.02 A

10.1 W

0.49 A

1.92 A

22.3 W

Celeron 800, Windows idle

4.81 A

24.0 W

0.53 A

2.25 A

37.8 W

Celeron 800, BIOS menu

5.96 A

29.8 W

0.50 A

2.37 A

43.6 W

Celeron 800, 1000000! in WinCalc

 

And my current home computer: 1800+ Thoroughbred running at 1.66GHz (overclocked), 1.5V (default voltage), 166MHz FSB on MSI KT3 Ultra, with 256MB of el-cheapo PQI PC2700 memory. 2pcs 7200rpm harddisks (60G Seagate BATA4 and Maxtor 80G w/8MB buffer), ancient 12x DVDROM, DVD+-RW burner. Matrox G450 16MB DDR, ADSL PCI card with Itex chip, SMC 1255TX network card (or something similar), a Hollywood+ MPEG2 decoder card, a cheap Firewire adapter with VIA chip and a CMI based PCI add-on soundcard (i.e. all 5 PCI slots filled).

 

I @ +5V

P @ +5V

I @ +12V

I @ +3.3V

P total

Description

4.62 A

23.1 W

1.31 A

4.13 A

52.4 W

AXP variant, Windows idle w/bus disconnect

8.51 A

42.6 W

1.28 A

4.29 A

72.1 W

AXP variant, BIOS menu

10.7 A

53.6 W

1.31 A

4.19 A

83.2 W

AXP variant, 1000000! in WinCalc