How many watts actually goes into computing and not thermal loss?

[Hulk]

I'm wondering how many watts from a typical processor actually goes into getting computing work done and how many watts is just heat?

Seems like you'd have to put the entire computer in a adiabatic chamber and load the cpu, measure the power into driving the computer, and then measure how much heat would have to be removed to keep the control volume at a constant temperature.

That would show the true efficiency of the processor right?

 

[dflynchimp]

there's no real standard for how much "work" gets done because you're trying to convert physical work (thermal/electric energy) into digital work, which runs on entirely different and incompatible scales.

What we can say for sure is that performance/watt efficiency is indefinitely improving. Archeteture and software improvements will hopefully never run out of new tricks. Like a couple of years ago when everyone was talking about the Pentium 4 scaling to 10 gigahertz, when it never broke 4GHz (non-overclocked). then BAM, the A64 comes along and opens a can of whoopass on the P4, then we get the move to dual core, and now quad core. Performance per gigahertz and core have also gone up dramatically. Is there an end in sight, who knows...hopefully not

 

[mgambrell]

There is a certain theoretical upper limit to the amount of computing you can get done per unit of energy or mass (look up bremermann's limit.) But I dont think the computing counts as work at all. I am no expert, but I think all the energy turns to heat eventually no matter what.

But thats not as strange as it sounds. A fire burns. Your body metabolises. Each just turns chemicals into heat, but one way creates a system of ongoing marvelous complexity and the other does not. Just as with different processors--each creates different levels of complexity with the same amount of heat. A 3ghz chip using 100W produces twice as much complexity as a 1.5ghz chip using 100W and they each produce 100W of heat. A human produces 100W and he is much more complex than either cpu! He can reason and move himself a mile.

Does anyone judge the efficiency of a car by miles per degree? No! It is miles per gallon. The gallons do not turn into miles. They turn into heat. It is just that along the way more useful work was done than an uncontrolled explosion. In the case of the cpu, we are clever enough to funnel the energy and heat through a series of "tubes" which create useful patterns.

 

[myocardia]

Originally posted by: mgambrell
There is a certain theoretical upper limit to the amount of computing you can get done per unit of energy or mass (look up bremermann's limit.) But I dont think the computing counts as work at all. I am no expert, but I think all the energy turns to heat eventually no matter what.

But thats not as strange as it sounds. A fire burns. Your body metabolises. Each just turns chemicals into heat, but one way creates a system of ongoing marvelous complexity and the other does not. Just as with different processors--each creates different levels of complexity with the same amount of heat. A 3ghz chip using 100W produces twice as much complexity as a 1.5ghz chip using 100W and they each produce 100W of heat. A human produces 100W and he is much more complex than either cpu! He can reason and move himself a mile.

Does anyone judge the efficiency of a car by miles per degree? No! It is miles per gallon. The gallons do not turn into miles. They turn into heat. It is just that along the way more useful work was done than an uncontrolled explosion. In the case of the cpu, we are clever enough to funnel the energy and heat through a series of "tubes" which create useful patterns.

Whoa, that's the best first post I've ever seen. Welcome to anandtech.

What I'm surprised to see not have been mentioned so far is the size of the process used in the processors. This is by far the biggest difference in how much work gets done per watt of energy consumed. Using your reference to fire, here's an example: I build a fire with 5 sticks, and so do you, yet mine produces 50x as much heat. How is that possible? It's very easy, you used twigs, with an average diameter of ¼", while I used "sticks" with an average diameter of 6".

Here's an example using processors. At one time, I had an Athlon XP 1800. It was a 1.53 Ghz Palomino core, which has a 180nm process. It idled @ 48-49°C using a better than stock heatsink, although I'm not sure exactly how hot it got under 100% load, but it definitely got quite warm, even though it was only running @ 1.53 Ghz, with 256KB of L2 cache. I replaced it with an Athlon XP-M 2600, which had a Barton Core, which is a 130nm process, with 512KB of L2 cache. It's stock speed was 2.0 Ghz, and it idled @ 37°C, at 2.5 Ghz, with considerably above the stock vcore, and was less than 45°C under 100% load.

Now, skip a year or two, and the last processor that I owned, which I just replaced, was a dual-core Opteron 170. It has 2x1MB of L2 cache, and is built on a 90nm process, and is 2.0 Ghz per core at stock speed. At 2.8 Ghz, it idled @ 34°C, with a 100% load temp of ~50°C. It was replaced with a 65nm Q6600, with 2x4MB of L2 cache, 2.4 Ghz quad-core. Even with only the stock heatsink, it idles @ 31C.

Now, the reason I noted each processor's amount of cache is because that cache has to be cooled also, since it's on-die. And the performance of each processor upgrade (I had others between, but just used these as examples) was very dramatic, because each was getting considerably more work done per second/hour/whatever than it's predecessor, even with single theaded apps. With SMP-enabled apps, the dual-core Opteron was getting upwards of 10x as much work done per minute/hour as the Palomino core, and the Q6600 does more than twice as much work per minute/hour as the overclocked Athlon, as long as the software I'm using can take advantage of all 4 cores.

 

[MarcVenice]

I think his initial question was about efficiency though, like, if you feed a CPU 100w, how much of that will turn into heat. With cars you can measure the miles you travel for each gallon you use, how are we to do this for cpu's though? Number of calculations ? That would be nice, but it wouldn't mean a damn thing. I think you misunderstood his question though, a carwill turn a gallon of gas into let's say 30 miles. But 50-60% or so of the energy in that gallon of gas is wasted and gets turned into heat. In cpu's there is no such thing though, unless you take into account the time a cpu idles, doing virtually nothing, yet consumes energy. In the end, the answer I think he was looking for, would have been, all energy you feed to a CPU, be it 50w or 150w, all of it will turn into heat.

 

[magreen]

What I'm surprised about is that nobody yet has said "I have a degree in electrical engineering and you don't, so you suck," which is what usually happens in this type of thread.

"...a system of ongoing marvelous complexity..." -- I like that. Welcome to the forums, btw.

 

[SerpentRoyal]

Add potential energy (height/stored energy) and kinetic energy (velocity) and we have work. There is no WORK without displacement. In general, the work done in a CPU is the movement of electrons. Power is work divided by time.

 

[wwswimming]

I think his initial question was about efficiency though, like, if you feed a CPU 100w, how much of that will turn into heat.

i would like to see Anandtech do more articles that include this kind of metric, a MIPS - per - watt kind of metric.

though, admittedly, MIPS are an older unit.

maybe ... Folding @ Home calculations per watt ?

What I'm surprised about is that nobody yet has said "I have a degree in electrical engineering and you don't, so you suck," which is what usually happens in this type of thread.

i went to a junior university near Palo Alto and worked in Silicon Valley as a design engineer, mostly in R&D, hands-on stuff. in college and at work, along the way, i inevitably met and worked with freakishly talented individuals (people like Anand). i worked with maybe ten that were absolutely "blow you away good" at metalworking or finite element analysis or electromagnetics. they all had a characteristic of trusting their own observations, loving their work, and having a lot of hands-on experience. most of them also had a serious exercise practice.

one or 2 of them had that kind of attitude, kind of like a French chef ?, wanting to control every detail. this would mean wanting me to do mechanical engineering their way.

i'm pretty sure the guy that did the original hardware design on the Mac started as a tech, & did not have an engineering degree. (Burrell ?)

the most talented hands-on stone-carving guy i know considers himself to be illiterate, though from hanging out with him i can tell that he reads a little bit.

i guess what i'm saying is, i really admire the self-taught, can-do techies that i've met.

 

[Hulk]

Originally posted by: MarcVenice
I think his initial question was about efficiency though, like, if you feed a CPU 100w, how much of that will turn into heat. With cars you can measure the miles you travel for each gallon you use, how are we to do this for cpu's though? Number of calculations ? That would be nice, but it wouldn't mean a damn thing. I think you misunderstood his question though, a carwill turn a gallon of gas into let's say 30 miles. But 50-60% or so of the energy in that gallon of gas is wasted and gets turned into heat. In cpu's there is no such thing though, unless you take into account the time a cpu idles, doing virtually nothing, yet consumes energy. In the end, the answer I think he was looking for, would have been, all energy you feed to a CPU, be it 50w or 150w, all of it will turn into heat.

Exactly. Actually the efficiency of an automobile engine IS determined by the temperature differences but that is another discussion.

The point is that the "thing" you're trying to do with a car is turn the crankshaft. Everything else is wasted. Heat, noise, friction losses, etc...

The "thing" you're trying to do with a computer is process instructions, anything else is just waste and that waste happens to be heat or watts if you will.

So if I'm reading this correctly if you load a certain computer and it requires say 100watts of input, and you measure the heat output it will also be 100watts. That would make sense since a cpu is a purely electrical device.

 

[mgambrell]

First let me start with a silly but actually quite pertinent analogy. It has been said that humans burn like a 100 watt bulb, mostly because they consume around 100 watts of power and emit that much heat. This is mostly invariant. But surely you know some folks that.. do a lot more _useful work_ with that 100 watts than some other folks do, eh? I'm talking about thinking, not moving pianos. How much useful work a person does has not so much to do with how much heat they emit. A genius man might help create a technology or a civilization. A worthless man might sit in a chair and pick his nose all day. All for the same heat.

---

I suppose I see how you guys are looking at it now. The problem is that a cpu is a much more closed system than a car.

With a car, even the energy not lost directly to the heat from combustion turns to heat eventually after doing its useful work through friction with the air and road and the chemical and physical decomposition of your tires. If you take a long enough view, eventually it all turns to heat.

With a cpu, there is no place to hide that heat temporarily so as to discard it where you might not notice. It just sits in your box and gets hot.

Ahh but I can think of a few ways. For example, perhaps you have noticed noise on your sound card line out on a laptop (or even an old desktop) whenever you hammer your cpu particularly hard? The electrons in the cpu dont just heat up directly as they pass through the resistance in the materials, but they also emit EM radiation which is capable of.. moving your speaker cones! You can imagine how that eventually dissipates in the room as heat. That radiation can even heat things directly, or perhaps cause your fluorescent cables to glow.

One way or another all that energy gets turned to heat. But here is the key: while moving the electrons X distance and creating the corresponding heat, you might produce N computations. But with a process shrink or a logic optimization, you might produce N*2 computations in the same amount of space and thus with the same amount of heat. There is no way to get around producing that heat when you move that many electrons in that amount of space. But you can be more efficient in what you get for that heat. The only relevant question here is how many computations you can perform per watt.

* I may be wrong or grotesquely oversimplified on some of these points, but this is just an exercise in thinking thermodynamically where one approach can clearly be more sophisticated than another.