Given a CPU-bound task that normally completes in time T, which of the following scenarios do you think uses less power?
1. CPU overclocked to complete the same job in 0.9T
2. CPU underclocked to compete the same job in 1.1T
Let's assume VCore is the same in both cases.
Which uses less power?
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
-
theycallmebruce
- Posts: 292
- Joined: Sat Jul 14, 2007 10:11 am
- Location: Perth, Western Australia
I don't know, but my hunch is the the underclocked processor will consume less energy (it will definitely consume less power at any given instant, but I'm pretty sure you mean energy) to perform the same task, despite taking longer.
The reason I say this is that although power consumption of a processor ideally grows linearly in proportion to clock speed, I've seen extreme overclocking tests where they show that this doesn't really hold true. Power consumption grows faster than linearly as clock speed is increased.
The reason I say this is that although power consumption of a processor ideally grows linearly in proportion to clock speed, I've seen extreme overclocking tests where they show that this doesn't really hold true. Power consumption grows faster than linearly as clock speed is increased.
When you think about this you have to think about how much power the rest of the system uses. A fairly typical value for the amount the motherboard, video card, and hard drive is say 50W. Let's say your processor uses 50W at full load. If the Vcore remains the same, the CPU overclocked will use 1.1*50W = 55W and underclocked .9*50W = 45W, as the power drawn by the processor is (approximately) proportional to its clockspeed.
This means the overclocked system uses 105W and the underclocked uses 95W.
The overclocked system will use 105*.9T = 94.5TW
The underclocked system will use 95*1.1T = 104.5TW
So here you can clearly see, in this case the overclocked system will use less power.
However if you use the case of a typical user, who only uses their processor at full load an estimated 10% of the time, the results switch and the underclocked processor will use less power.
This means the overclocked system uses 105W and the underclocked uses 95W.
The overclocked system will use 105*.9T = 94.5TW
The underclocked system will use 95*1.1T = 104.5TW
So here you can clearly see, in this case the overclocked system will use less power.
However if you use the case of a typical user, who only uses their processor at full load an estimated 10% of the time, the results switch and the underclocked processor will use less power.
If you mean which one will take less total energy to complete the task, I don't think you gave enough information to get meaningful answers. (Too many things left unspecified.)
What happens between 0.9T and 1.1T? Is the overclocked machine left on, but idle, is it turned off? If left on, what are the systems idle characteristics?
(e.g. Newer systems have low power idles, older systems (e.g. Win 9x) didn't so much.)
Should we assume that T is a relatively long time?
(At very short T, the higher power involved with the overclocked machine will swamp the effect of the slightly longer time interval, and the lower power system will use less energy. (Think of the limiting case of infinitesimal time, the energy approaches the power.))
If I were to hazard a guess
Assuming:
The machine is turned off when the task is done
T is relatively long
There is a reasonable amount of other (non-CPU) hardware.
I would guess that the overclocked system might use less energy (because all the constant load would be switched off sooner).
But:
There are enough variables in this that may not get a meaningful answer.
The big energy savings seem to come from undervolting, rather than just underclocking (theoretically a better than linear effect on power). So there may be a better question to ask than the original question.
"In theory there is no difference between theory and practice, but in practice there is."
(It may be easier to answer this question with a Kill-A-Watt meter and a computer than by thinking.)
A vaguely related question that I have wondered about (I don't have a Kill-A-Watt meter or an undervoltable system to find some answers).
Assuming you have an unlimited amount of work, where will you get the most computation done for the least power input. (Not assuming a fixed Vcore.)
Speed-step etc. kick up the Vcore and clock when there is work to be done, and drop it for idle. Which makes sense if you need to get something done soon. But if there is an unlimited supply of work, then you will never finish, so where is the sweet spot in terms of computations per watt?
Obviously idle is not the best (unless you can shut everything down), since you get nothing done, but still invest energy.
Semi-practical application: Distributed computing jobs (folding, etc.)
Is running it flat out the best bang for the buck?
Is there/should there be an intermediate state for running tar-baby jobs like this that optimizes instructions/watt?
What happens between 0.9T and 1.1T? Is the overclocked machine left on, but idle, is it turned off? If left on, what are the systems idle characteristics?
(e.g. Newer systems have low power idles, older systems (e.g. Win 9x) didn't so much.)
Should we assume that T is a relatively long time?
(At very short T, the higher power involved with the overclocked machine will swamp the effect of the slightly longer time interval, and the lower power system will use less energy. (Think of the limiting case of infinitesimal time, the energy approaches the power.))
If I were to hazard a guess
Assuming:
The machine is turned off when the task is done
T is relatively long
There is a reasonable amount of other (non-CPU) hardware.
I would guess that the overclocked system might use less energy (because all the constant load would be switched off sooner).
But:
There are enough variables in this that may not get a meaningful answer.
The big energy savings seem to come from undervolting, rather than just underclocking (theoretically a better than linear effect on power). So there may be a better question to ask than the original question.
"In theory there is no difference between theory and practice, but in practice there is."
(It may be easier to answer this question with a Kill-A-Watt meter and a computer than by thinking.)
A vaguely related question that I have wondered about (I don't have a Kill-A-Watt meter or an undervoltable system to find some answers).
Assuming you have an unlimited amount of work, where will you get the most computation done for the least power input. (Not assuming a fixed Vcore.)
Speed-step etc. kick up the Vcore and clock when there is work to be done, and drop it for idle. Which makes sense if you need to get something done soon. But if there is an unlimited supply of work, then you will never finish, so where is the sweet spot in terms of computations per watt?
Obviously idle is not the best (unless you can shut everything down), since you get nothing done, but still invest energy.
Semi-practical application: Distributed computing jobs (folding, etc.)
Is running it flat out the best bang for the buck?
Is there/should there be an intermediate state for running tar-baby jobs like this that optimizes instructions/watt?
Last edited by scdr on Mon Jan 07, 2008 11:31 pm, edited 1 time in total.
There is a possible analogy with a car engine, basically going faster ends up making power consumption exponential.
Also you have to consider the other component that need power and there usage according to the load, else you just need to find out the best energy per instruction ratio.
Alos computers, not unlike those who make them, are just lazy bastard bragging about what they can do but sitting on the silicon ass 99% of the time.
Also you have to consider the other component that need power and there usage according to the load, else you just need to find out the best energy per instruction ratio.
Alos computers, not unlike those who make them, are just lazy bastard bragging about what they can do but sitting on the silicon ass 99% of the time.
If both machines start at the same idle state, do the work, and then return to the same idle state, the the power usage should be about the same. The faster one will use more power, proportional to clock speed (with significant overclock, it goes non-linear, but 10% should be close to linear) but also for a proportionately shorted time.
If you have to turn off power savings modes to get the +/- 10%, and the machine is idle for long periods of time, then actually, 1.0 T with power savings modes on may use the least power.
Many CPUs will let you lower the voltage more if they are underclocked, etc. so really, this question can vary even from chip to chip.
If you have to turn off power savings modes to get the +/- 10%, and the machine is idle for long periods of time, then actually, 1.0 T with power savings modes on may use the least power.
Many CPUs will let you lower the voltage more if they are underclocked, etc. so really, this question can vary even from chip to chip.
Wow, thanks for all the insights. I think jwickers analogy to car engine makes sense.
Here's what I'm trying to do:
On a regular basis I run some video transcoding batch jobs overnight. The system shuts down after the jobs are done. I'm trying to determine if I should 1. overclock(with the necessary overvolting) as much as possible and get the jobs done sooner, or 2. underclock (and undervolt) and run at lower power state for longer period of time.
My original constraint of keeping the Vcore the same was an attempt to simplify the comparison. I'm sure removing this constraint will make the comparison more complicated.
Here's what I'm trying to do:
On a regular basis I run some video transcoding batch jobs overnight. The system shuts down after the jobs are done. I'm trying to determine if I should 1. overclock(with the necessary overvolting) as much as possible and get the jobs done sooner, or 2. underclock (and undervolt) and run at lower power state for longer period of time.
My original constraint of keeping the Vcore the same was an attempt to simplify the comparison. I'm sure removing this constraint will make the comparison more complicated.
at least with the current crop of Core 2's, this is not really true. most are very overclockable, even on stock voltage. most overclocking sites have charts which show clockspeed increases at stock voltage do not really increase power consumption very much at all. another example is the latest Core 2 Quads, which can be significantly undervolted at stock speed. so you save a lot of energy over the stock configuration, even though the number of computations performed is the same. even a slight undervolt + overclock should be possible, this should give you the highest performance per watt.There is a possible analogy with a car engine, basically going faster ends up making power consumption exponential.
Another item to consider - to what extent is the video transcoding job compute bound? (Seems like it could have a reasonably hefty I/O component as well.)
i.e. do you need more than a 10% overclock/underclock to get a 10%
speedup (or slowdown respectively). (And how much of an overvolt/undervolt do you have to do to get that clock speed.)
Given reasonable amount of I/O and changing voltage - I would guess that running at stock and undervolted, or underclocked/undervolted might win out. But it also depends on what sort of CPU you have (is it a Prescott P4, or a Core 2), and how many disks are you using (Raid 5 of 10k+ RPM monsters, vs. SSD
.
i.e. do you need more than a 10% overclock/underclock to get a 10%
speedup (or slowdown respectively). (And how much of an overvolt/undervolt do you have to do to get that clock speed.)
Given reasonable amount of I/O and changing voltage - I would guess that running at stock and undervolted, or underclocked/undervolted might win out. But it also depends on what sort of CPU you have (is it a Prescott P4, or a Core 2), and how many disks are you using (Raid 5 of 10k+ RPM monsters, vs. SSD
.-
smilingcrow
- *Lifetime Patron*
- Posts: 1809
- Joined: Sat Apr 24, 2004 1:45 am
- Location: At Home
It’s a lot easier to make accurate assessments when you know the actual job in hand as there are no one size fits all solutions.frank2003 wrote:Here's what I'm trying to do:
On a regular basis I run some video transcoding batch jobs overnight. The system shuts down after the jobs are done. I'm trying to determine if I should 1. overclock(with the necessary overvolting) as much as possible and get the jobs done sooner, or 2. underclock (and undervolt) and run at lower power state for longer period of time.
My original constraint of keeping the Vcore the same was an attempt to simplify the comparison. I'm sure removing this constraint will make the comparison more complicated.
But you need to be even more precise as it will depend on the codec used and how many cores it can utilise and how well it scales.
If it scales well with 4 cores and with higher clock speeds then I would start with looking at Quad core CPU that is over-clocked at stock voltage.
For something like DivX that utilises SSE4 you really want a Penryn chip. A fast dual-core with SSE4 might be more efficient than a quad core with SSE3.
The other equation is money. If that’s no object then a high end C2Q will tend to over-clock higher at stock VCore than the cheaper ones.
The transcoder program I'm using is multithreaded and pegs the two cores for 99.9% of the time. Yes, the program does take a breather to do some IO but that's an insignificant portion of the time in a multi-pass mpeg2 to divx transcoding session. In short, my task is purely CPU-bound.
It sounds like the real answer can be obtained only through experimentation. I'll take down my AM2 dual core based PVR (it's running 24/7) one of these weekends and attach it to a Kill-A-Watt and measure away. I will report back on my findings.
Thanks for all the responses.
It sounds like the real answer can be obtained only through experimentation. I'll take down my AM2 dual core based PVR (it's running 24/7) one of these weekends and attach it to a Kill-A-Watt and measure away. I will report back on my findings.
Thanks for all the responses.
Without changing voltages: rougly the same energy taken by the processor since power is roughly linear in frequency. Including the rest of the system, the overclocked will take less power since time is shorter.
Including voltage adjustments: now probably the underclocked system takes less power, since undervolting has a big effect on power. Instead of linear cpu power versus frequency is roughly square to cubic.
Including voltage adjustments: now probably the underclocked system takes less power, since undervolting has a big effect on power. Instead of linear cpu power versus frequency is roughly square to cubic.
I ran a simple test to convert an HDTV mepg2 recording to Divx. Here are the results.
Config:
Biostart TA690G, 1mb RAM, 2.5" notebook drive.
At 1.0ghz @ 0.784V vcore:
34W for 180 minutes, or 102 Watt-hours.
At 1.8ghz @ 0.976V vcore:
46W for 100 minutes, or 76.7 Watt-hours.
So on the surface, it's seems like "get the job done as soon as possible" strategy uses the least amount of power.
Config:
Biostart TA690G, 1mb RAM, 2.5" notebook drive.
At 1.0ghz @ 0.784V vcore:
34W for 180 minutes, or 102 Watt-hours.
At 1.8ghz @ 0.976V vcore:
46W for 100 minutes, or 76.7 Watt-hours.
So on the surface, it's seems like "get the job done as soon as possible" strategy uses the least amount of power.