XP-M vs A64 - cpu speed per wat calculation completed

[VERiON]

{Forgive me my english, but english is not my native language}


I was wondering - should I buy XP-M or A64 for my silent HTPC, and which cpu is faster at the same level of power consumption.




STEP 1: 20W = passive cooling
==============================

I have a running fanless PIII733, which discipates ~20W of heat.
I've assumed that I can passively cool both XP-M and A64 running @ 20W



STEP 2: software
================

I've downloaded CPU Power by Kostik @ http://www.silentpcreview.com/Web_Links ... d-137.html
and edited file cpupower.txt to add XP-M and A64 3000+ specs.

Just cut and paste text below:

[AMD Athlon Mobile XP-M]
Athlon XP-M 2500+ (1.45); 1.45; 45; 1867
[AMD 64]
Athlon 64 Winchester 3000+ (1.4); 1.4; 67; 1800

you can add any CPU you like - the convention is: NAME; DEFAULT_VOLTAGE; MAX_POWER_IN_WATT; RUNNING_FREQUENCY



STEP 3: XP-M & A64 @ 20W
========================

I've assumed that i can easily lower voltage to 1,15V (I know you can go lower than that, but I have to make som assumptions)

So, according to CPU Power by Kostik, you can achieve 20W at this settings:

Athlon XP-M 2500+ @ 1,15V @ 1327MHz = 20W
Athlon 64 Winchester 3000+ @ 1,15V @ 800MHz = 20W



STEP 4: how to start with no data?
==================================

Since there is no CPU underclocked and undervolted test (at lest one that fits me ) i used Mikael idea
from http://forums.silentpcreview.com/viewto ... highlight= to LINEARY scale cpu power.

Mikael said: "They didn't test an 800MHz Athlon64. What I did was to extract the results from the 1.8GHz 3000+ by multiplying the results from THG with 0.4444..... Since scaling isn't completely linear, the real results from an 800MHz A64 would likely be somewhat better than what you get with this method. It should still be quite accurate, especially for applications not depending on graphicscard performance (i.e. games)."

scaling:

XP-M @ 1327Mhz; 1,15V = 71% * 1867 (oryginal frequency)
A64 @ 800; 1,15V = 44% * 1800 (oryginal frequency)

assumption:

underclocked & undervolted XP-M @ 20W has 71% of oryginal XP-M power
underclocked & undervolted A64 @ 20W has 44% of oryginal A64 power



STEP 5: data from tomshardware.com

==================================
I'll be comparing 4 cpus test results from http://www.tomshardware.com/cpu/20041221/index.html

1. PIII800EB (the closest to my 20W PIII733)
2. Celeron 1.2 Tualatin (my previous CPU - for personal speed comparsion)
3. XP-M
4. A64


STEP 6: normalization
=====================

i want to have good comparsion, so i've normalized the THG results to PIII800EB (PIII800EB result is allways 100)



STEP 7: results
===============
convention:
CPU_NAME | CPU_THG_SCORE * UNDERCLOCKED_CPU_FACTOR = UDERCLOCKED_CPU_SCORE = X_TIMES_SCORE_IS_BETTER_THEN_PIII_800

Code: Select all

SYNTHETIC SISOFT SANDRA MULTIMEDIA BENCH INTEGER         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-23.html         
         
PIII  800 |   100        
CEL  1200 |   154        
XP-M 2500 |   247 * 71,00% = 175 = 1,75 * PIII  800  
A64  3000 |   247 * 44,00% = 109 = 0,70 * PIII  800
         
         
SYNTHETIC SISOFT SANDRA MULTIMEDIA BENCH INTEGER         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-23.html         
         
PIII  800 | 100        
CEL  1200 | 154        
XP-M 2500 | 247 * 71% = 175 = 1,75 * PIII 800
A64  3000 | 247 * 44% = 109 = 1,09 * PIII 800
         
         
SYNTHETIC SISOFT SANDRA MULTIMEDIA BENCH FLOATING         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-23.html         
         
PIII  800 | 100        
CEL  1200 | 153        
XP-M 2500 | 221 * 71% = 157 = 1,57 * PIII 800
A64  3000 | 223 * 44% =  98 = 0,98 * PIII 800
         
         
PCMARK04 SYNTHETIC CPU POWER         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-25.html         
         
PIII  800 | 100        
CEL  1200 | 137        
XP-M 2500 | 249 * 71% = 177 = 1,77 * PIII 800
A64  3000 | 278 * 44% = 122 = 1,22 * PIII 800
         
         
UNREAL TOURNAMENT 2004 (direct X 9)         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-15.html#directx_9         
         
PIII  800 | 100        
CEL  1200 | 115        
XP-M 2500 | 264 * 71% = 188 = 1,88 * PIII 800
A64  3000 | 360 * 44% = 158 = 1,58 * PIII 800
         
         
FAR CRY (direct X 9)         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-15.html#directx_9         
         
PIII  800 | 100        
CEL  1200 | 113        
XP-M 2500 | 251 * 71% = 178 = 1,78 * PIII 800
A64  3000 | 355 * 44% = 156 = 1,56 * PIII 800
         
         
LAME mp3 encoder         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-25.html         
         
PIII  800 | 100        
CEL  1200 | 108        
XP-M 2500 | 216 * 71% = 154 = 1,54 * PIII 800
A64  3000 | 216 * 44% =  95 = 0,95 * PIII 800
         
         
WINRAR         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-21.html#application         
         
PIII  800 | 100        
CEL  1200 | 98        
XP-M 2500 | 126 * 71% = 89 = 0,89 * PIII 800
A64  3000 | 218 * 44% = 96 = 0,96 * PIII 800

STEP 8: conclusions
===================

Since it is not "scientific" method, and lot of "assumptions" was made - my results may not reflect real world CPU scaling at all
But it if I am correct (more or less) then Athlon XP-M is the best CPU for me (occasional GAMER).

In most cases XP-M @ 20W is 1,5 - 2 times faster than pIII800, which is more then enough for internet, office works, mp3, DivX and HTPC.
A64 (except games) @ 20W is as fast as PIII800.

So, when you exclude games, XP-M is +50% faster at the same level of power consumption (20W).

And extra bonus - XP-M+mobo is little cheaper then A64+mobo939.

[Linus]

My "long" response: I suspect there are issues with your calculation. However, considering you want to run fanless most (if not all) of the time, it makes very little sense to spend twice as much on an A64 system.

My short response: I agree with you.

[SometimesWarrior]

There is a significant problem with your calculation: your assumptions about peak power use are incorrect. Sure, the specs say the Athlon 64 can use up 67 watts, but that's not what has been measured.

Look at this site: http://www.goodwin.ee/sulo/Power2.htm

The guy took apart the PSU and found a way to accurately measure the true power consumption of about a dozen processors. His numbers look like they scale correctly: for example, he compares the Thoroughbred 1.8GHz (1.6V) to the Thoroughbred 1.6GHz (1.5V). That's 51.1W peak vs. 39.6W peak, both tested with CPUBurn. Measured ratio is about 1.29:1. If you do the math for the expected ratio: 1.8/1.6 * 1.6^2/1.5^2 (in relation to power use, clock rates scale linearly and voltages quadratically), you get a very close 1.28. So you can use either processor as a reference for your Mobile XP-M.

His Athlon 64 scales are pretty accurate, too, but they're thrown off a bit because of the on-die memory controller, which doesn't seem to scale with the CPU core. Still, they are very close: Going from 2.0GHz @ 1.5V to 1.0GHz at 1.1V should be a ratio of 3.72, and he measures 3.77.

Okay, armed with this new information, you can run a 1.0GHz Athlon 64 at 1.1V, and CPUBurn will only use up 12.6W for the core, and 2.2W for the memory controller (this is measured, so we know the chip is capable of running at such a speed). Let's just assume you can run 1.2GHz at 1.2V, this sounds do-able. That gives you 20.2W max power. Now all your A64 benchmark figures should be 50% higher!

Let's see if the Athlon-XP numbers should be increased too: Compare the 1.327GHz @ 1.15V to the 1.6GHz @ 1.5V, which is 39.2W. 1.327/1.6 * 1.15^2/1.5^2 * 39.2 = 19.3W. So that number is accurate, and you shouldn't re-figure the Athlon XP numbers, unless you find out that the processor is capable of running at higher clock speeds or lower voltages.

Now, all these calculations assume that each processor turns the same % of its power consumption into heat. This may not be correct, as I have learned just this week: for example, my tests with the 6600GT video card, which draws 50W of power, lead me to believe that it generates less heat than a 50W Athlon-XP processor, because the heatsink doesn't seem to be as hard to cool. The explanatio I've been given is "leakage current", a phenomenon I'm still trying to understand. I guess a rule of thumb might be, the smaller the process (130nm, 110nm, 90nm...), the smaller percentage of power consumption that is turned into heat. I think both processors are 130nm, so hopefully this isn't a factor.

Edit: Here's a thread where people talk about A64 underclocking/undervolting and temperatures: http://forums.silentpcreview.com/viewtopic.php?t=16206 Notice Jan Kivar says his processor runs 2.0GHz @ 1.3V, or 1.0GHz @ 0.85V. Those are encouraging figures: you might be able to do 1.3GHz @ 1.1V or something of that sort. I'm still looking for people's underclocking numbers for A-XP processors...

Also, keep in mind that A64's have Cool 'n Quiet ability, and if you don't mind having a slow-spinning fan turn on under extended CPU load (it could be inaudible over the sound of any 3.5" hard drive), you can basically run your A64 at full-speed and it will only turn on the fan if you load the processor for a long time (games, etc). Short loads won't turn the fan on, but your system will be as responsive as if the processor was at 2.0GHz (because it is, for short time periods). Maybe it's possible (I don't know) to configure CnQ to only go through a partial range, so it could ramp up an down within a range of values that are all passive-cooling capable.

Edit 2: Here's some undervolting figures on a Thoroughbred, which lists maximum clock speeds at different voltages. It's likely that your XP-M is capable of slightly more undervolting than the processor used in this article (but probably not a whole lot more), so the 1.3GHz @ 1.15V expectation seems accurate.

[tay]

I really like the analysis even if the numbers used were incorrect based on the sometimeswarriors post.

AFAIK,leakage current is dissipated as heat. If this was not the case prescott wouldnt be so bad. Of course it doesnt help that prescott has more baggage than northwood (rumored to be a disabled x86-64 featureset).

[mrzed]

Yeah, unless there is some really wierd quantum effects I'm unaware of in silicon chips, all power drawn must become heat.

[SometimesWarrior]

Yeah, unless there is some really wierd quantum effects I'm unaware of in silicon chips, all power drawn must become heat.

That's what I thought too, but I was told differently and now I'm confused, so I think my best option is just to keep quiet on the matter till I get things straightened out.

[StealthGirl]

Firstly, there is good old fasioned friction. Those transistor gates opening and shutting so fast builds up a fair bit of heat - they may be small but there's a lot of them.
Secondly, electron leakage. To some extent this is caused by the first, as it gets worse the hotter the chip is. Electrons "jump" circuits to where they should not be, and one thing you learn in physics is that any electron jump is always from a high potential plane to a lower one - which means a net loss of energy. Where does that energy go? Yep, you guessed it.
This is also the reason for processor instability at high temperatures. There is a temperature range between stable operation and permanent damage to the structure of the silicon where enough electron leakage occurs to disrupt the calculations the processor makes, so the processor starts making errors and programs crash.

ROFL! I've GOT to send this one to my friends at AMD! HA! That's good!

Good calcs though, nice work to both Verion and SomtimesWarrior. Thanks for the numbers!

[sthayashi]

Firstly, there is good old fasioned friction. Those transistor gates opening and shutting so fast builds up a fair bit of heat - they may be small but there's a lot of them.
Secondly, electron leakage. To some extent this is caused by the first, as it gets worse the hotter the chip is. Electrons "jump" circuits to where they should not be, and one thing you learn in physics is that any electron jump is always from a high potential plane to a lower one - which means a net loss of energy. Where does that energy go? Yep, you guessed it.
This is also the reason for processor instability at high temperatures. There is a temperature range between stable operation and permanent damage to the structure of the silicon where enough electron leakage occurs to disrupt the calculations the processor makes, so the processor starts making errors and programs crash.

This is a load of crap.

I'm inclined to agree, but I'm not entirely sure which bits are fact and which aren't (and I work with chips for a living).

Attempting to get back onto topic, it's safe(r) to assume that all energy drawn goes into producing heat. In actuality, some may go into producing light, some into sound. So Heat Output <= Power Drawn. We generally assume Heat Output = Power Drawn, because Power draw is relatively easy to measure and it's better to overshoot anyways.

[tay]

Anyway, if a processor is drawing 50W of power and giving of 50W of heat then you've not got a processor, you've got the world's most efficient electric heater!
Even electric heaters don't have 100% efficiency at converting electricity to heat, so if a processor was doing it...!

Ok I shouldve posted the bit that followed. I forgot out the best part. In any case, yes there is fact mixed in with fiction in there. Like all the little gates opening and closing causing friction. This is not why power draw increases with switching. Its all that source drain transistor crap that i've forgotten.

[Elixer]

The only problem is that a lot of athlons aren't able to cold start at a voltage that low. My mobile can only go as low as about 1.3V; any lower and it won't post on a cold start.

[merlyn]

Elixer - sounds to me like you need to drop your operating frequency. I currently have an XP-M 2500+ which starts and performs flawlessly. it runs at 166 x 9 = 1494MHz @ 1.115V. many people on this board have achieved similar results.

I promise to do a write-up soon as it's nearing retirement.

[meglamaniac]

This is a load of crap.

I have to agree. I think I had a brain freeze or something that day.
I didn't actually mean real friction, I was trying to put transistor capacitance into simpler terms (poorly).
The bit about electron leakage is broadly correct.

However, I think an all round "doh!" is called for heh.

[lm]

It's true that not all of the energy which is used by the cpu turns into heat inside the cpu. Part of it can turn to radiation of various wavelengths, for example radiowaves. These would be perceived as noise on any nearby device that receives this specific wavelength of radiation.

Now the part that turns into something else than heat should be very small, and even then, most objects that directly surround the cpu would absorb them and would eventually transform to heat very near the cpu, so effectively we can say that all of the cpus energy usage transforms into heat in the cpu.

Any errors in the calculations might be because we for example measure the voltages from points in the circuits that are somewhat away from the cpu and so part of the energy transforms into heat in the wires that transfer the electricity to the cpu or sth like that.

[VERiON]

There is a significant problem with your calculation: your assumptions about peak power use are incorrect. Sure, the specs say the Athlon 64 can use up 67 watts, but that's not what has been measured.

Yes, I'm aware of that. AMD simplified things a bit. It is impossible that 3000+, 3200+, 3500+ peak power is equal (67W). I think it simplifies cpu binning process. 3000+ can be as hot as 3500+ and still fits specification numbers (running @ 67W).

Let's just assume you can run 1.2GHz at 1.2V, this sounds do-able. That gives you 20.2W max power. Now all your A64 benchmark figures should be 50% higher!

Thanks for the numbers. Since OpenOffice calc is handling all calculation (great piece of free software BTW) I will replace A64 CPU_UNDERCLOCKED_CORRECTION FACTOR. I'll rise it from 44% to 64% oryginal CPU speed and post the results when i get back home.

It's safe(r) to assume that all energy drawn goes into producing heat. In actuality, some may go into producing light, some into sound. So Heat Output <= Power Drawn. We generally assume Heat Output = Power Drawn, because Power draw is relatively easy to measure and it's better to overshoot anyways.

I agree, it's better to be on a safe side

ok. i have the results

Code: Select all

SYNTHETIC SISOFT SANDRA MULTIMEDIA BENCH INTEGER         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-23.html         
         
PIII  800 |  100        
CEL  1200 |  154        
XP-M 2500 |  247 * 71% = 175 = 1,75 * PIII 800
A64  3000 |  247 * 64% = 158 = 1,58 * PIII 800
         
         
SYNTHETIC SISOFT SANDRA MULTIMEDIA BENCH FLOATING         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-23.html         
         
PIII  800 |  100        
CEL  1200 |  153        
XP-M 2500 |  221 * 71% = 157 = 1,57 * PIII 800
A64  3000 |  223 * 64% = 143 = 1,43 * PIII 800
         
         
PCMARK04 SYNTHETIC CPU POWER         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-25.html         
         
PIII  800 |  100        
CEL  1200 |  137        
XP-M 2500 |  249 * 71% = 177 = 1,77 * PIII 800
A64  3000 |  278 * 64% = 178 = 1,78 * PIII 800
         
         
FAR CRY (direct X 9)         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-15.html#directx_9         
         
PIII  800 |  100        
CEL  1200 |  113        
XP-M 2500 |  251 * 71% = 178 = 1,78 * PIII 800
A64  3000 |  355 * 64% = 227 = 2,27 * PIII 800
         
         
LAME mp3 encoder         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-25.html         
         
PIII  800 |  100        
CEL  1200 |  108        
XP-M 2500 |  216 * 71% = 154 = 1,54 * PIII 800
A64  3000 |  216 * 64% = 138 = 1,38 * PIII 800
         
         
WINRAR         
http://www6.tomshardware.com/cpu/20041221/cpu_charts-21.html#application         
         
PIII  800 |  100        
CEL  1200 |  98        
XP-M 2500 |  126 * 71% =  89 = 0,89 * PIII 800
A64  3000 |  218 * 64% = 140 = 1,40 * PIII 800

You can see that A64 (except games) is equal or slightly better then XP-M
Mayby i'll add 3500+ scores, because theoretical initial max cpu wattage (67W) is the same for both processors (3000+ & 3500+), and i'm trying to show which cpu is faster @ 20W. It will add another +25% to A64 score.

[SometimesWarrior]

You can see that A64 (except games) is equal or slightly better then XP-M
Mayby i'll add 3500+ scores, because theoretical initial max cpu wattage (67W) is the same for both processors (3000+ & 3500+), and i'm trying to show which cpu is faster @ 20W. It will add another +25% to A64 score.

You're right about the need to revise power figures for the 90nm Winchester-core 3500+. Since it's a different product series than the 130nm Newcastles, its power use figures are going to be different, and all the sites I've seen show lower power consumption. That means an upward adjustment in performance figures, but I'm not sure by how much. AMD says 25%, but once again that number may be affected by their plans for maximum CPU speed of the series.

I found this interesting test at Xbit-labs where they measure the 12V current going to the CPU as a way to gauge the relative reduction in power going from 130nm to 90nm. Although the results could be affected by efficiencies of power regulators between the two test boards, their numbers show a power reduction of 25% (after accounting for the core voltage drop to 1.4V). So you and AMD were right all along.

But another interesting observation, in that same article, is that core temperatures don't actually drop by 25%. Xbit-labs' explanation is that, because the A64 die is smaller in size, the thermal conductivity limitations (Watts/m^2) conspire to make the chip harder to cool. As evidence, they note that the heatsink on the new 90nm Winchester core feels significantly cooler than the older 130nm Newcastle core, despire the processor core temperatures being very close. If you combine this information with AMD's new maximum safe temperature spec (65C for the new chips, compared to 70C for the old chips), then you see that the decreased power consumption of the 90nm Athlon 64 may not help you.

It's a lot to think about, but since I'm in the market for a new A64 system of some sort, I'm trying to digest the information too.

[VERiON]

I've read that x-bits article. Very interesting.

They measured max power consumption of 3000+ = 40W
That confirms our previous assumption that 3000+ @ 1,2V @ 1200 MHz consumes 20W.

My first conclusion was: why buy A64 winchester, when you can have the same preformance @20W with cheaper XP-M platform.

In my country [A64 3000+ winchester & Gigabyte GA-K8NSNXP] is 400$ and [XP-M 2500+ & abit nf7-s 2.0] is 270$ which is 70% cheaper ... but within 3 months the prices will be in the same range.

I've come to another conclusion. You DON'T HAVE TO buy an XP-M to have that low temps. You can have the same temps with "modern" A64, which is great, because you don't have to stick to "old" socket A platform anymore (when you have spare money to go for socket 939/winchester).

[Jan Kivar]

Okay, armed with this new information, you can run a 1.0GHz Athlon 64 at 1.1V, and CPUBurn will only use up 12.6W for the core, and 2.2W for the memory controller (this is measured, so we know the chip is capable of running at such a speed). Let's just assume you can run 1.2GHz at 1.2V, this sounds do-able. That gives you 20.2W max power. Now all your A64 benchmark figures should be 50% higher!

For the record, I'm running my A64 also at 1600 MHz/1.1V, which results in 22.5W. 1400MHz/1.1V results under 20W.
[Calculated by entering 1000 MHz/1.1V/12.6W to the stock values, and then adding manually 2.2W for the memory controller after CPU Power calculation. This way, for max. stock speed (2000 MHz/1.5V) I get 49W, which is a bit different from the given 89W value...]

This is with an A64 3000+ 130nm CG-core.

Cheers,

Jan

[VERiON]

I've wrote earlier that using the same peak power for different processors simplifies cpu binning process. Here is a quote from "Athlon 64 for Quiet Power" thread with great explanation of that process.

As for the 89W thermal number for all K8 speed grades, it is fairly interesting. I can see a number of reasons for doing this. I assume that these are partly why this is this way:

o Because they can: PC manufacturers do not want to engineer an airflow for multiple thermal numbers. Especially in the build-to-order market, if a PC must ship with a 3.4GHz processor, chances are that the 2.4GHz version is identically cooled, so there is no advantage to giving more constrained numbers. Note that big OEMs might actually get specially binned parts with lower numbers if they care for a particular platform.

o If they do that, they can play a very interesting game. With process variations, fast parts are often very leaky too, so in addition to running hotter because they can run faster, they run even hotter because leakage is higher. Normally, these extra leaky parts must be thrown out because they exceed the thermal specs. However, if this fast part is rebinned in a lower-frequency bin, it can compensate the extra leakage with the lower frequency-based power. This actually is a common trick. If the thermal spec is the same accross the frequency bins, it makes it much easier to downbin leaky parts and improve yield.
Of course, this means that in a lower-frequency bin, you'll find "normal parts" that run fairly cool, but can't be overcloked too much, but also very hot parts that can be overclocked a lot. So you'll encounter a lot of variation in heat and OCability in those bins.

[Jan Kivar]

I've wrote earlier that using the same peak power for different processors simplifies cpu binning process. Here is a quote from "Athlon 64 for Quiet Power" thread with great explanation of that process.

As for the 89W thermal number for all K8 speed grades, it is fairly interesting. I can see a number of reasons for doing this. I assume that these are partly why this is this way:
...

AFAIK the reason for unified TDP is that every time AMD released a new CPU the XP lineup (Palomino core, actually even T-Bird should be included here), the new CPU had a higher TDP than the previous ones. This caused OEMs a lot of headache in order to keep the temperatures under limits, and even today the cheapest HSFs have texts like "AMD Athlon processors up to 2200+" or "Athlon 2000+ and above".

With the current method, if the HSF passes tests at 89W load, it's safe to put to any available CPU, which simplifies the OEM assembly lines. (Excluding the FX-55, which has 104W TDP... I wonder if it ships with a different HSF?)

Cheers,

Jan

[Mikael]

The Winchester would be a much better choice than the AXP-M here. I have seen indications that a stock S939 3000+ outputs something like 30W (or even lower) @ 1.8GHz and 1.4V. This would indicate that if you could lower the CPU voltage to 1.2V and retain the CPU frequency of 1.8GHz (which shouldn't be impossible), you would end up with a power figure of 20W, or maybe even less than that. The Winchester @ 1.8GHz would provide for significantly better performance-to-wattage rating that an AthlonXP-M. I'd say that it's definately worth going for the Winchester if you are able to do so.

[VERiON]

I have seen indications that a stock S939 3000+ outputs something like 30W (or even lower) @ 1.8GHz and 1.4V.

I'm sorry but this is too good to be truth. Can you post any link? Or more information

No offence.

[akio]

Hi there.
I found an interestingsite associated with this thread, although it's Japanese.

The current through 3.3V/5V/12V is measured with digital multi-meter and clamp-probe to gauge approx CPU power consumption.
Motherboard is MSI K8T Neo2-FIR for Athlon 64 3200+ (Winchester core).
Clock/voltage is set 1GHz/1.10V or 2GHz/1.40V with Crystal CPUID 4.0.3 tool.

Now that I saw the table, I would like to go Geode NX.

[Mariner]

I have seen indications that a stock S939 3000+ outputs something like 30W (or even lower) @ 1.8GHz and 1.4V.

I'm sorry but this is too good to be truth. Can you post any link? Or more information

No offence.

VERiON. Try the following link:

Toms Hardware

Not sure to the accuracy of these measurements but one or two of the knowledgeable people on this message board have indicated Tom's technique ought to be reasonably accurate.

[Mikael]

I'm sorry but this is too good to be truth. Can you post any link? Or more information

No offence.

Well, consider this then:

AthlonXP @ 1.8GHz 1.45V = 45W
AthlonXP @ 1.8GHz 1.20V = 31W

This reduction in power does not take into account that power output will increase non-linearly with increasing voltage. In other words, the real world figure would be even lower for the undervolted CPU, since leakage currents are decreased. The 1.8GHz AthlonXP on 1.2V will most certainly consume less than 30W.

Now, let's take a look at this Anandtech graph:



Note that those figures are for the whole system during load. It's pretty clear from those graphs that the Winchester isn't getting it's lower power consumtion only from the voltage decrease. The chip is simply more efficient even at the same frequency and voltage! The idle power consumtion of an A64 Winchester is reported to be around 10-15W with Cool 'n Quiet disabled. Check for example this test at Tom's. Let's say 15W to be on the "safe" side. With this information (and with Anandtech's figures) we can calculate the approximate power consumtion of a 2.2GHz Athlon64 Winchester:

86-15 = 71W (system power consumption minus the CPU)
114-71 = 43W (CPU power consumption during load)

We're now disregarding that the motherboard probably dissipates a few extra watts during load, so the real figure is probably around 40W for a 2.2GHz Winchester at 1.4V.

Extracting those results down to 1.8GHz and 1.2V:

(1.8/2.2) * (1.2/1.4)^2 * 40 = 24W

This is again without taking reduced current leakage into account, so close to 20W could be a realistic figure.

Looking at our initial AthlonXP calculations, those for the A64 Winchester also make sense. Since the Winchester is much more power efficient than the AthlonXP, it's reasonable to conclude that it outputs a lot less than the AthlonXP at the same settings. Which ever way you look at it, I wouldn't expect a 1.8GHz Winchester at 1.2V to output more than 25W. And as said, it's probably even less than that.

Ohhh, and I realise that there's an awful lot of calculations and that this method isn't entirely accurate. It should however provide a decent indication of the Winchesters thermal performance.

EDIT: One could off course also use Tom's load figures directly:

(1.2/1.4)^2 * 28.8 = 21W

Everything points to close to 20W @ 1.8GHz and 1.2V.

[Blappo]

Is the die size of the PIII near the same of the 90nm A64? If not, I don't think that it is a valid assumption that a Winchester using 20W can be passively cooled. Although I have no experience with this.

[Shriek]

... if you could lower the CPU voltage to 1.2V and retain the CPU frequency of 1.8GHz (which shouldn't be impossible) ...

While I can't speak for wattage and thermal dissipation, I can say that I recently got an A64 3000+ 90nm and an Asus A8V Deluxe, and they're performing quite well set to 1.8GHz @ 1.05V (Prime95 overnight--it errors out in less than ten seconds at 1.0V).

[VERiON]

@Mikael

that sound convincing, sorry for my disbelief
...and thanks for the data

[Mikael]

@Mikael

that sound convincing, sorry for my disbelief
...and thanks for the data

No problem! Ohh, and if you want even better power characteristics you should probably wait for the "Venice" core Athlon64 which is to be released soon. Also, Shriek's result with a 1.8GHz Winchester at 1.05V is totally awesome! Makes one wonder why AMD doesn't take a serious stab at Intel's Dothan... Maybe that's what the new Turion platform is supposed to accomplish?

I might add that cooling the 90nm CPUs passively might be harder than the PIII, because of what Blappo said: Die size. For example, my AthlonXP-M @ 1.2GHz and 1.3V easily climbs over 50C with zero to minimal case airflow. This is while being cooled by a Thermalright SI-97... Cooling it passively is probably not a problem at 1GHz and 1.1V, but the CPU is having problems cold booting at voltages that low...

[mczak]

Is the die size of the PIII near the same of the 90nm A64? If not, I don't think that it is a valid assumption that a Winchester using 20W can be passively cooled. Although I have no experience with this.

Die size doesn't make that much of a difference if you're willing to accept high temperatures anyway. Why? Because a small die size causes a _constant_ temperature difference vs. a larger die (with the same wattage) to a given heatsink. So, if you try to cool it down to almost room-temperature levels, this is a problem, but if the temperature difference between heatsink and air is much larger anyway this constant difference is (relatively) not that large.
Some example (all numbers completely pulled out of the air):
room temperature 20 degrees centigrade, small die causes 10 degrees difference to heatsink, large die 5 degrees. So if you want to cool the cpu to, say, 35 degrees, this is a lot harder with the small die, since the heatsink-air temperature gradient is only 5 degrees, whereas with a large die it would be 10 degrees.
But for passive cooling, you're usually accepting temperatures about 60 degrees. So those 5 degrees more from die-heatsink you get by using a smaller die isn't all that much, compared to the 30 degrees gradient you have between heatsink and air.

btw winchester die size isn't that much smaller compared to PIII coppermine anyway.
newcastle 144mm²
winchester 84mm²
coppermine 106mm² [edit: later steppings (cC0) 90mm²]
tualatin 80mm² [edit: this is with 512KB cache, supposedly 256KB cache versions are just as large, with half the cache deactivated]

[Mikael]

I've been running my AthlonXP-M @ 1GHz and 1.1V passive for a while now. Running it at this speed seems to be okay. I have a little airflow around it coming from the case fan and PSU, though. Temperatures are far from hot, with idle temp of 40C and load temp of 43C (Prime95).

EDIT: It's still not cold booting at these settings, though.

EDIT2: Removed back case exhaust and have been running with the case closed for the past 45 minutes. Temps rose a little but have now stabilised at 48C. Only airflow out of the case now is the fan in my Antec TruePower (which is modded with a lower speed fan). Looking pretty good. It's a shame that it won't cold boot and I suspect that the only way around that is to increase the voltage, right?