Slashdot Mirror
Chipmakers Admit Your Power May Vary
Dylan Knight Rogers writes to mention a News.com story discussing the realities of chip power consumption. From the article: "Assessing only pure performance is passe. The debate these days is about performance-per-watt, which seems like it should be a simple miles-per-gallon type of calculation. However, miles are miles, and gallons are gallons. There's no one simple way to measure processor performance, and measuring the amount of power output by today's chips is proving just as difficult."
Comment removed based on user account deletion
Performance being difficult to measure is well known- you can't go by clock speed, or even clock speed*instructions per clcok since these will differ based on instruction mix. For power, a simple inverter will use different amounts of power depending on if its on or off- exact power for a chip is impossible to guess. This is all old news.
I still have more fans than freaks. WTF is wrong with you people?
Comment removed based on user account deletion
So it's exactly like the miles-per-gallon on new cars.
"I use a Mac because I'm just better than you are."
This is what benchmarks are for. Compare the performance of two systems with other variables held as constant as possible. This has been going on for years, has it not? If I want a computer to play games, I see what different CPU configurations yield in, say, HL2 with the same ram and video card.
Is this perfectly scientific? No.
Is it practical? Hell yes!
BTW, EPA mpg are measured without using real mile on real roads.
Engineering is the art of compromise.
Our nation is one of conviences, not of caring if our grandchildren have conviences.
May Heisenberg will protect us!
Stop! Dremel time!
Miles per gallon are hardly constant either. Uphill? Downhill? 10mph or 100? Highway or city? Same difference.
---- Den ene knappen er powerknapp, den andre er Bender voice knapp "Bite My Shiny Metal Ass"
For most users (i.e. not power-users doing heavy calculations for some scientific purpose, or high-quality video editing, or raytracing), most processors provide way more power than needed, and have done so for years. Or at least, they *would* provide all that power if the software running on top of it wasn't bloated and unnecessarily complex, unoptimized and badly written. And no, I'm not just talking about Windows, I'm including Linux, MacOS and all the others in the bag.
The best proof that modern software makes modern hardware suck is that, back in the mid-eighties, I used an Atari ST to do desktop publishing, and it wasn't all that different from what I can do now with a simple PC that would look like a supercomputer back then.
"A door is what a dog is perpetually on the wrong side of" - Ogden Nash
You need a stats class, badly.
No offense, but nothing is perfect. This is why we have a thing called "standard deviation".
Me hitting the letter "e" will probably not take the same amount of energy to process twice. But I bet over 1000 e's the standard deviation could be found and would indicate that 66% of the time it's "x J +/- y" and so on...
So you sample something like "building the linux kernel to a ram drive" 100 times, find the deviation and use that. The tighter AND lower the better. The wider and higher the worse.
Tom
Someday, I'll have a real sig.
So how does the electric company know how much you've used if power is impossible to measure?
As per my other post under this article the author is confusing issues of measurment with issues of prediction.
KFG
I'd think both AMD and Intel are well aware of the MIPS/Watt challenge. It's not new. Problem is CUSTOMERS still want a bazillion Ghz attached to the processor because they think it will make it faster or better or something.
I've got two x85 class Opterons sitting here at 1Ghz most of the time. That's ~35W vs. ~95W. AMD seems to care about power. Intel is no worse off with the Pentium M and "core" series (netburst was a mistake).
Tom
Someday, I'll have a real sig.
Dvorak admits he trolls, Chipmakers admit your power may be vary, what's next?
Looks like everyone's coming out of the closet today.
The SUV of CPUs.
I don't want it to be like buying industrial lighting where you have to compare 60W bulbs with 800 lumens to 75W bulbs with 1000 lumens.
I'm not confused by simple linear ratios: We just don't need naming conventions and measurements whose only purpose is to obfuscate easy
comparisons that would allow for simple commodity pricing of a consumer good (which is what you are purchasing -- operations per second).
Why, for a given chip, power consumption raises with clock speed? I know there's corelation, but I'd like to know the physical relation between the two variables.
It seems some want to eliminate the time component from speed measurements, so you'd only care that one machine got to 110,000 calculations versus another getting to 120,000 calculations.
With desktop machines, just hook up each computer to a 1000VA battery backup UPS and see how FAR each gets ... not whether one got to 100,000 calcs in six minutes versus twelve minutes.
At first glance, I thought the headline said "Chipmunks" instead of "Chipmakers." For a split second I thought maybe Alvin, Simon, and Theodore had a tech column going.
This sig, aah-ah, is comin' like a ghost-sig...
All K8 processors have had frequency scaling, most K7s had it too, was just disabled in the desktop parts. The P4s had frequency scaling too.
The problem before hand was that the designs were just inefficient. It took your K6-2 or P2 running at full tilt to keep up with demand. Scaling didn't make too much sense. Now a 500Mhz K8 can cope with most usage, playing mp3s takes less than 1% of the cputime where it used to take more than 80% on a 486...
I wouldn't call the power savings as a "new scam" or trick.
Tom
Someday, I'll have a real sig.
Power is still costing me about the same as it did years ago. I think what happened is that
A: Companies started caring more, since they have to pay twice the electricity (once in their server farm, and again to remove the resulting heat).
B: Portable computing keeps gaining in popularity, including cellphones and PDA type devices. Less power demand increases battery life and reduces weight.
C: CPU's just started getting so hot that more and more elaborate measures were needed to cool them. Reducing power demand is ultimately cheaper than having to use liquid cooling or AC systems for every person's computer.
I don't read AC A human right
Semprons don't, which came as a bit of a surprise when I tried to configure clockspeed control and found that it wasn't working.
20 January 2017: the End of an Error.
Maybe not the K7 semprons but I thought K8s did. Did your BIOS recognize the CPU?
Tom
Someday, I'll have a real sig.
Only the server guys ever cared about CPU usage. That's why a SCSI operation utilizing 20% over 60 seconds, beat an IDE operation using 30% for 50 seconds (in this case 3 seconds total time saved)
The same exact processor can exhibit up to 50% variation in average power usage. Manufacturing variability.
The BIOS recognizes it just fine. For whatever reason, AMD either disabled clockspeed control in Socket 754 Semprons or didn't include the necessary circuitry for it (depending on whether Semprons are Athlon 64s with certain functionality (such as half of the cache) turned off or whether they're a completely different design.)
20 January 2017: the End of an Error.
My point was that scaling a 200Mhz chip doesn't usually make sense since you needed the speed to use the damn box. Sure you could idle it but the savings wouldn't be as important.
Only when we started getting into designs like the K7 and Core processors did the speed become excessive. A 2Ghz K7 core was way more than capable of playing mp3s or video files while not killing the box.
Tom
Someday, I'll have a real sig.
Then I want to know the next step. I want to know how many tens/hundreds/thousands/millions of instructions per cycle (I henceforth trademark the analogy GIPC/MIPC/KPC for any processor performance comparison!!!!) the processor can handle, plus the comparative cycle rate (I.E. Speed) of the processor. Then I'd compare that per watt. I'd think that would be better - specialized instructions or not.
Still waiting on Serviscope_minor to wake up to fucking reality and realize that Jessica Price isn't going to fuck him.
They just about guesstimate.
My solution, that is:
A standardised code segment with broad instruction type usage and a long time to complete(to minimize differences/errors). If everybody is using the same reference instruction sequence on all processors for the same amount of time, no debate ensues. Right?
Why hasn't anyone combined the concepts of heat pipe, heat sink, and calorimeter? Anyone who's taken high school physics knows the concept.
The foundation's already there, in water cooling systems. A rudimentary system could be built by dropping a thermometer in the reservoir, and turning off the radiator at the beginning of each test.
Granted, you're only measuring waste heat, but how much power does a CPU pump through data busses?
tasks(723) drafts(105) languages(484) examples(29106)
The reason that the frequency scaling on the desktop came about is for two reasons:
1. When the TDP of chips hit over about 80W, heat became a problem and frequency scaling helped abate that.
2. The tech is needed in laptop chips. With the exception of the Pentium 4, all notebook chips are just modified desktop parts. So the manufacturers made one core and saw no reason to disable that feature, especially since #1 is true.
So frequency scaling just piggybacked its way onto the desktop. What's good for the goose (notebook) is good for the gander (desktop) and I am surprised that this hadn't shown up before.
Just "gittin-r-done," day after day.
He had a point though; if you constantly do video editing with your PC, your personal measurements of performance will be different from someone who uses only Word and Firefox all the time.
... etc.
There's a reason I ask people what they intend to do with their PC before selling them one -- do they need more ram, or more drive space or more drives or a bigger video card
Very rarely does CPU speed come into the equation; the slowest CPU available at retail is quite fast enough for most people, most of the time, but if they say "Play 3D video games", it might not be.
- Michael T. Babcock (Yes, I blog)
The article glosses over the real problems.
The first real problem is that blade servers are so small now, but require so much power, that companies can easily fit way more compute power in a server room than can be reasonably cooled. So they need more power-efficient servers to use their server space effectively.
And the problem isn't that power can't be measured--it can be measured just as easily as performance. Which is the problem hinted at in the article--firms focusing on the positive results they have and pushing that way of measuring power. Which is what they've been doing with performance for decades. Everyone can measure power, but what "benchmark" should we use?
And to correct other comments, chip dynamic power utilization is proportional to fCV^2, where f=frequency, C=capacitance, V=voltage. Reducing a chip's frequency from 2GHz to 1.5GHz will only at best save you 25% of the power. But, circuit speed is also proportional to voltage, so if a chip at 1.2V can operate at 2GHz, there's a good chance it might operate at 1.5GHz at 1.0V (or maybe 1.1V). So the real power savings is in the voltage reduction: 1.2V at 2GHz use almost twice the power as at 1.0V at 1.5GHz. But chips waste power even at 0Hz, especially at 90nm and below, so it's not quite that good. I believe AMD and Intel both use voltage reduction to save power in their reduced power modes.
Having the idea of maximizing efficiency is one thing, actually doing it is another. Transmeta chips were just too slow.
Why?
Me hitting the letter "e" will probably not take the same amount of energy to process twice. But I bet over 1000 e's the standard deviation could be found and would indicate that 66% of the time it's "x J +/- y" and so on...
But with different usage-patterns you *will* get consinstant differences..so sure, you'd get data, and they'd be valid, but that doesn't mean GP needs a stat-class. Regardless of how much people refuse to believe it, even in todays 'massage-the-data-until-you-find-something(anythin g-will-do)'-world, the ol' trusted "garbage in, garbage out" still holds.
"" How about taking the safety labels off everything, and let the stupidity-problem solve itself? """
I'm sorry, but under no conditions will "compile this C file" vary by an unmanageable amount. If you expect the power to vary by 500 Watts each time you compile something... you're sadly mistaken.
... you're wrong.
Most likely with the CPU/memory under full load the Wh deviation is less than 10% of the mean usage. On a typical desktop the Wh rating is about 200-250 at full load. If you see a variance of more than say +/- 20Wh something is wrong or the test isn't reproducible. If you think things like differing occurences of interrupts and cache misses will make a difference in Wattage
So if you can say box A and box B compile [or do work] in T units of time, let's compare the power. Box A takes 100Wh +/- 15Wh and box B takes 175Wh +/- 20Wh. Which one do you suppose takes more power? Is that "numbers-saying-anything-they-want?"
That's the whole MIPS/Watt thing. If AMD and Intel hit an empasse where both are just as IPC efficient the next question is the power they take to get there. Sure if Box A takes 3x the time the results need adjusting (namely it takes 3x the power) but we say MIPS PER WATT for a reason.
Tom
Someday, I'll have a real sig.
Why?
;-)
Because, as the parent post pointed out, the problem of how to describe the measurement of a varying quantity does not actually pose a problem.
But with different usage-patterns you *will* get consinstant differences.
I agree the "hitting E" example seems a bit odd, but I would guess the parent just needs a computer science course badly.
Instead, do the same with a few typical real-world usage patterns (arranged into a repeatable suite). Then divide the performance value by the watts value, and you have a ballpark measure of performance-per-watt. Repeat the test a few hundred times, and you can get the mean and standard deviation.
Now, your objection seems to center on the idea that not everyone will use a computer with a similar usage pattern. I grant that as true, almost trivially. But if you want a one-number rating, you need to accept some averaging.
Personally, as someone who writes his own CPU heavy code, I might only care about how efficiently it can process an inner loop of "movq, psllq, pxor, pand, paddd, movq, psrlq, pxor, pand, paddd" (a real example from a program I toyed with last year, which probably single-handedly accounted for over half of my main PC's CPU time for several months). But even having a highly unusual usage pattern, I still get numbers in the same ballpark as published benchmarks. Why? Because chipwide power consumption depends far more on architectural considerations than it does on individual instruction-with-context consumption.
Dropping to 90nm SOI gave AMD a huge leap over Intel when they introduced the Winchester core Athlon 64s - Then an additional boost with the "dual stress liner" Venice core. And now Intel has come back with the 25W Core Duo (The M, Yonah's predecessor, did better than the Athlon 64s, but cost an arm and a leg), which seems poised to take the crown of MIPS-per-watt for a while. And how did it do this? By further architectural changes such as the drop to 65nm and reduction in the number of FPU pipelines. So although it might have poor FP performance, the specific instruction mix won't matter as much as overall "FP-heavy"ness - And then, similarly FP-heavy apps will get similar MFLOPS-per-watt.
Hall effect sensors can easily measure current with no resistance added or interfering with the circuit. Many IC manufacturers test VERY complex integrated systems every day with NO issues -- and IC test setups use nowhere near 1ohm for current tests. Maybe 0.1 ohm. I think the real problem here is the lack of a standardized setup and method for measuring. Having different supply points or loads on a single chip is pretty trivial. The biggest difference is the performance -- what the chip is doing -- when it is measured. I'd almost think you'd want a ramp up of processor performance and measure current use while this happens. There is no guarantee with todays processors that performace is linear with power consumption. Either that or measure at standard "benchmark" levels.
Charge up a new battery, see how many mAh it has, and use the computer. See how long it lasts with the way you use a computer. There, that's the only number that really matters.
My God, it's Full of Source!
OUTSIDE_IP=$(dig +short my.ip @outsideip.net)
That’s not even true: for a start, a UK gallon is some 20% bigger than a US gallon, and there are many more types of gallon in existence...
Need to type accents and special characters in Windows? Use FrKeys