backtop


Print E-mail del.icio.us 154 comment(s) - last by Setsunayaki.. on Dec 16 at 6:23 PM

AMD's power consumption figures advertised and republished for customers just don't add up

Those familiar with the processor business will be familiar with the term Thermal Design Power (TDP), which refers to the number of watts that a cooling system must dissipate under a maximum real program load.  AMD, however, has adopted a different power measurement system to rate its processors.

The new system, dubbed Average CPU Power (ACP), was championed by AMD for the first time with the release of its K10 Phenom family.  AMD claims ACP is comparable to TDP measurements in Intel's benchmarks; and is in fact a superior measuring stick for power consumption.  ACP is described as being calculated based on an "average" daily use scenario for desktop users or for a server.

Despite AMD's claims that ACP is directly comparable with TDP, many began to have doubts when AMD struggled to raise the speeds on its Phenom processors. This could indicate, among other things, thermal issues due to high power consumption.

The results are startling.  The story is best explained by two versions of a white paper ironically titled "ACP – The Truth About Power Consumption Starts Here," -- the old version can be found here and the new version here.  The key information in these papers can be seen in the table below.

 K10 Family ACP and TDP
Processor
TDP Whitepaper 1 TDP Whitepaper 2
ACP Whitepaper 1
ACP Whitepaper 2*
Quad-core AMD Opteron SE
120 W 137 W
105 W
105 W*
Quad-core AMD Opteron
95 W
115 W
75 W
75 W*
Quad-core AMD Opteron HE
68 W
79 W
55 W
55 W*

Look at those numbers closely.  The first thing to notice is that TDP measurements are significantly higher than ACP.  When AMD compared its power consumption figures to Intel's TDP, ACP measurements significantly underestimate power consumption.   TDP differed between the two versions of the white paper by as much as 20 W, which is a 21% increase in the case of the quad-core Opteron.   AMD did not increased its ACP estimates, emphasized in bold, despite the TDP increase.

Either the ACP is an arbitrarily measured system, and AMD changed it at will for its convenience, or AMD's document team failed to update the document properly.   There is no other feasible explanation of why a 20 Watt TDP increase would be accompanied by no increase in ACP. 

Whitepapers are a very important tool for an electronics firm to inform the public of its products' performance.  Glaring inaccuracies such as these certainly reflects negatively on AMD.  Furthermore, the admission that ACP can be as much as 20 Watts lower than the TDP on the same processor indicates that AMD is spreading misinformation using these ACP numbers in its advertising.

To put this information in perspective, a 3.16 GHz Xeon X5460 from Intel squeaks in at a still weighty 120 W.  While AMD failed to disclose in the white paper on what frequencies its selected processors operate, it is almost surely 3.0 GHz or lower, as 3.0 GHz is the highest speed K10 processor currently demonstrated.  The best case scenario is that a 166 MHz slower AMD processor consumes 17 more watts.  If this was truly the case, it would not be the end of the world, but it would indicate AMD falling behind further in performance. 

However, if the samples tested were lower than 3.0 GHz, obviously the picture becomes far worse.  And since AMD's 2008 roadmap states that its 2.4 GHz processors are rated at 125 Watts TDP, this is almost certainly the case.  Architecture and design advantages aside, K10 is a chip that is almost a gigahertz slower but with a significantly higher power consumption rating. 

AMD's Phenom processor already underwent  lackluster reception at launch, and the company faces ever-higher-piling losses.   The problems have led AMD to try to focus on expanding its K8 platform offerings, instead of addressing the compounding issues of its K10 platform. 

The ACP/TDP issues are just one of many issues with the K10, but have many ramifications -- everything from increased rate of chip failure, to simple higher energy bills at no additional speed benefit.  Expect AMD to fix its documents soon, but it will take far longer for AMD to salvage the physical hardware of its K10 platform.

Updated 12/10/2007
AMD has clarified its position on ACP and TDP. Read here for further information.



Comments     Threshold


This article is over a month old, voting and posting comments is disabled

nothing wrong about this, right?
By cgrecu77 on 12/10/2007 11:53:33 AM , Rating: 5
So what's wrong with AMD adopting a similar policy (with Intel) of measuring power consumption? If Intel is actually quoting average power, then so should AMD - to level the playing field.

Of course, if AMD's way of measuring the average is significantly different than Intel's, that that could be considered "cheating", but I can't tell from this article.

Is it me, or lately there's a slight bias agains AMD at anandtech/dailytech?




RE: nothing wrong about this, right?
By Spuke on 12/10/07, Rating: -1
RE: nothing wrong about this, right?
By System48 on 12/10/2007 12:25:26 PM , Rating: 2
Intel's TDP isn't an average, it's the max. AMD's ACP is what they think the average server is loaded to not what the max load may produce. To me this almost seems dangerous, if you design your cooling solutions based on the ACP power but then you have your system loaded to 100% for extended periods it's more likely that the processor would overheat and fail.


RE: nothing wrong about this, right?
By FITCamaro on 12/10/2007 1:11:03 PM , Rating: 3
Yes I agree. Processor thermal ratings should be measured at max load. You need to design a cooling solution around that, not the average load. That'd be like designing a cars radiator for the average temperature. It would overheat on extremely hot days.


RE: nothing wrong about this, right?
By TomZ on 12/10/2007 1:22:42 PM , Rating: 2
That would be like putting transmission specifications in the spec sheet for an engine. It makes no sense since the engine (or CPU) can be used with any number of transmissions (or chipsets).


RE: nothing wrong about this, right?
By KristopherKubicki (blog) on 12/10/2007 1:25:32 PM , Rating: 2
This is why you only see Analogy used as an argumentative form on Boston Legal. Though FIT's example was a little more relevant than most I've seen.


RE: nothing wrong about this, right?
By theapparition on 12/11/2007 10:37:43 AM , Rating: 4
You just publicly admitted to watching Boston Legal?


By KristopherKubicki (blog) on 12/11/2007 1:03:01 PM , Rating: 4
Damn... Why don't we have an edit button.... :-P


RE: nothing wrong about this, right?
By TomZ on 12/10/2007 1:32:04 PM , Rating: 2
Sorry, I attached this reply to the wrong post. I'll copy to the correct one. Sorry ahead of time for the duplicate.


By Marty007 on 12/11/2007 12:43:13 AM , Rating: 1
all true, your cooling should be able to take the max load under any circumstance,
however.... especially with liquid cooling systems,
you dont want to go TOO cool if your cpu is idling/average(amd power estimate)
you might get condensation and all the troubles with that,
and fact is, some cpu's have serious issues with sub-zero temperatures,

i think intel's way of measuring is the more usefull for designing,
but i think amd's measuring will be more accurate over a normal day of usage .. but its strange that the difference is so much however..

they should really just create 3 stats for new cpu's
- idle power drain
- average power drain
- max power drain

and also, inform at wich voltages/speeds these values are benched (so they have to make these stats for every single product they release.. )

its not like you got your cpu working at full power 24/7
and an few hours of labor shouldnt be a real problem for proper cooling, as most cooling is semi-intelligent by default they will spin up when things get warm


RE: nothing wrong about this, right?
By Dactyl on 12/10/2007 1:35:24 PM , Rating: 2
You're not supposed to design cooling systems based on the average.

But:

1 - the average can tell you about power costs of running the system

2 - in a place with lots and lots of CPUs, that needs to be air conditioned, the average power use can tell you something about how much cooling capacity the entire room normally needs (it's unlikely that all CPUs would be running at max capacity).


By Proteusza on 12/10/2007 2:33:17 PM , Rating: 5
If you are running a server room, the average is little to no use to you, because it is so imprecise.

If you plan to have your CPU's on average of 75% load, what kind of cooling would you need to install? Difficult to tell with this "average" figure. Most server rooms would have quite high CPU loads, so this average isnt much use to them.

Besides, I dont like the use of the word "average" here. It implies they did statistical research of real world Barcelona chips and discovered that the mean, or average, TDP was a certain wattage, and then termed this new figure ACP. I dont think they could have done that, so a better term would be estimated.

But hey, when your products are inferior, use marketing, it worked for Intel for so many years.


RE: nothing wrong about this, right?
By mars777 on 12/10/2007 7:53:01 PM , Rating: 4
quote:
Intel's TDP isn't an average, it's the max. AMD's ACP is what they think the average server is loaded


While you are correct regarding AMD, you are just partially correct about Intel.

Intel declared TDP as the maximum power drawn from 100% usage of the CPU through standard application load (defined by Intel).

AMD TDP stands for the maximum power draw a processor can handle before it burns out (basically every gate on the CPU drawing power - which is never achieved).

And there is more to this:

If ACP is, say 80% load, then there can exists processors with different TDP with the same frequency and the same ACP. But only if using AMD representation of TDP. Those processors would be better in overclocking under 80% load because they produce less leakage under 80% than the other one. But it would mean that the manufacturing process is producing more leakage on 100% load while the silicon is able to sustain it without producing calculation errors. Of course overclocking on 100% usage would produce system crashes... and nobody wants to overclock the CPU to use only 80% of it...

Sorry for the language, i'm not anglo-saxon / english.


RE: nothing wrong about this, right?
By lemonadesoda on 12/10/2007 8:14:48 PM , Rating: 1
This brings us back to the "old" AMD... the days of fried CPUs. Happened to two of my K6-III's. And they were nasty expensive in thier day. After that experience I switched away from AMD.


By bryanW1995 on 12/10/2007 11:06:31 PM , Rating: 3
good switch. you moved to intel just in time to take advantage of that great p4 feeling.