Slashdot Mirror
AMD Athlon64 4000+ Underclocking
Bios_Hakr writes "PC Stats is running an article on their experiences underclocking an AMD 4000+ processor. Their goal was to try and reduce the voltage requirements and lower the heat output. They benchmark using 3dMark01, 3dMark05, as well as SuperPi. From the article: 'This got us thinking though; what about under-clocking? Most modern processors and motherboards can just as easily run under a rated speed as it can run over... but is there a point to this? Well possibly.'"
Their next article: how to remove 2 cylinders from your Ferrari's V12 engine.
Microsoft operating systems and software accomplish this without all the work.
Intolerance for ambiguity is the mark of the authoritarian personality.
The point is...
Suspense = more clickthroughs = more ad views = more revenue.
If you are designing a system for high reliability, under temperature extremes and such (military environments for example) underclocking is the way to go - you can minimize power and heat loads as well as potentially avoid timing instabilites that occur when you push a processor to the performance margins.
I thought what they are testing is the whole point of AMD Cool'n'Quiet technology, but they don't even mention it in the article! Nice try reinventing the bicycle. I'm already underclocking my Athlon 64 right now, thank you.
RTFA. From the article:
Case fans can generate some audible noise in an environment designed for quiet, and is this really acceptable? Professional studios can acoustically isolate computers, making this a moot point, but home audio enthusiasts don't have this luxury. The question is, how much heat and voltage can be dumped by underclocking a given processor (down) while still retaining acceptable processing performance?
The purpose of this article is to take a very fast, very hot modern processor (in this case an AMD Athlon 64 4000+) and underclock it with an eye to comparing performance to levels of heat and voltage at below stock speeds. The Athlon 64 is currently the fastest available desktop processor, so we reasoned that reducing its speed to the point where it could be operated silently with a passive cooling system should still leave us with a powerful machine for everyday tasks.
"The legitimate powers of government extend only to such acts as are injurious to others." Thomas Jefferson.
You would think the whole point gets moot - the system certainly won't be quiet. (I believe there is an actual need for quiet systems eg in recording studios etc - which make the article interesting, but not great).
I have a really elegant proof for Fermat's last theorem. If this sig was only a bit longer...
I think the point is that these "better" processors were built with that larger processor speed in mind, and if you underclock it, you still get the added benefit of somethign that's supposed to cool and use a bigger processor for a smaller one. It's like... AMD when they build the 4000+ over the 3200+, attempt to make the 4000+ as calm and quiet as possible, within limits, and go farther with the 4000+ than with the 3200+. Now if you underclock the 4000+, you still take advantage of that extra technology...
I'm probably wrong.
If you want low power you can buy systems specifically designed to perform well on low power supply.
The article is about researching how to build such systems out of cheap commodity parts, unlike the proprietary, often Windows-only parts found in laptop computers.
I think it is somewhat useful information. While most people are thinking about how much faster you can process, many of us are looking to reduce the noise of fans blowing. I recall when 800MHz was a fantastic speed... hell, for that matter, 300MHz was pretty nice too depending on how far back you go.
And are we really using all of those cycles? Not really. Right now, a system's performance (IMHO) is largely the responsibility of the quality of RAM, Video and system board stuff. After all, what "feels" fast must be fast. If I've got a slow hard drive, then it's a slow system and if I can accellerate the video, then it's a slow system. What good is 4GHz if you've got a slow everything else... and by the same token, if you've got a fast everything else, a 2GHz processor is probably plenty.
I decided to underclock some 1U systems (~XP 2500) to ensure that they would never overheat (longevity was more important than performance).
It works perfectly: a drop of 20% in core clockspeed greatly reduced the heat output, the core temperature dropped by almost 10 degrees C.
TODO: 753) write sig.
They listed the drop in temperature from 33.5 to 26.9 as a 20% drop. However, they didn't mention the ambient temperature. If you take 20 degrees, then this drop is more like 50%. That would also mean that it was consuming well under half the power. (I'm assuming watts->degrees is exponential.)
As a secondary matter, the person who got me interested in BSD, as a rule, made his servers with whatever was the cheapest AMD-K6, underclocked to 350MHZ. Bulletproof boxes with long lifetimes. I'm sure there are still some churning out the bits around this town.
On all of my personal must-stay-up servers, I get a processor that is too beefy for the task it's to do, then clock it down. It's usually rock solid and runs very cool. In some cases I've been able to get by using only passive cooling and still keeping the processor very cool, making the system solid, cool, and nearly silent.
If you want a vision of the future, imagine a youtube comments section scrolling - forever.
If they want to underclock a 4000+, they could just swap me my 3000+.. I wouldn't complain.
--
I have an Athlon 2100+ which I bought back in 2002 when it was brand new. After installing it, I experienced frequent lockups as the CPU overheated under heavy loads. I bought a new heatsink/fan combo (a Thermaltake Volcano 9, which was pretty good at the time) to replace the standard AMD one, but it sounded like a jet turbine at full speed and it only alleviated the problem a little. After that, I underclocked my FSB by only 3MHz (133MHz to 130MHz) and I haven't had a lockup in over a year. The associated drop in performance is unnoticable.
OLPC Australia
I don't get it. If you wanted a low power system, why would you buy the top chip and underclock it?
/sys), you can undervolt the chip even more than AMD's driver allows. My Athlon64 3200+ spends most of its time at 1 volt, 1 GHz, and it runs at ambient temperature. That's right, the heat generated is so little that on a stock cooler, the processor does not raise its own temperature significantly. And if a 3800+ with CnQ is too powerful, you can back down to the 3500, 3200, or 3000 models, depending on your exact needs.
The Athlon64 4000+ is a Hammer chip, iirc. It's manufactured on a 0.13nm process, and is a slightly older core. Instead of using that chip, why not use the Athlon 3800+? The Winchester and Venice cores are both 0.09nm chips, and run much, much cooler and dissipate much less heat than the Hammer cores. And you can use AMD's Cool n Quiet (aka PowerNow) technology to back off the processor speed to half speed automatically, when the processor isn't seeing heavy use. Coupled with a utility like RMClock on windows or a custom written utility on Linux (if your 2.6 kernel has the necessary options enabled, you simply have to write to some files in
Problem solved, and without several pages of blathering about underclocking.
In this context, talking about a 20% drop in temperature in degrees celsius makes no sense for comparison purposes. They go on to state that "a 43% drop in voltage producing a 20% drop in heat seems more reasonable", but this is assuming that the temperature drop corresponds to a equal reduction in heat output.
- Brian.
-
- - You can't take something off the Internet! That's like trying to take pee out of a swimming pool.
They have measured the (absolute) celcius temperature of a well-cooled system, without quoting the ambient temperature. Then, concluding that the temperature hasn't droppped much, they assume the power hasn't dropped much.
The correct measurement is the *difference* in temperature between the CPU and the ambient air. Power dissipation is linearly proportional to this.
a very fast, very hot modern processor (in this case an AMD Athlon 64 4000+)
Very hot? If you haven't already bought one, just make sure to get one with a Winchester or Venus core.
Using C&Q, mine (only a 3000, but "close enough" to make my point) could probably get away with purely passive cooling. Using a meter at the plug, it draws a whopping 54 watts average, with 48W idle (C&Q engaged) and 65W max.
Thanks to modern CPU power saving technologies as implemented in all newer Athlons and Opterons, or Pentium M, you really don't need to sacrifice peak performance for the sake of power and heat. They deal with usually sitting there idle fairly well, by throttling back, without needing to resort to such (relatively) drastic measures as "suspend" and "hibernate".
I do, however, see one possible use for underclocking... When you keep your CPU always pegged at 100% (running Seti @home or the like, for example). Then, underclocking would allow you to trade a little bit of performance for a lot of power and heat reduction.
From 99-03 I always underclocked my AMD CPUs for gaming, the minimal requirements necessary for playing everyones favorite Counter-Strike, and my undying urge to make my CPU last ages kept my system running smooth, my latest underclocking was from a 2800+ 2ghz AMD down to a 1.87ghz miracle machine that ran on a minimal fan system, never overheated and my record uptime was 3 months and 2 weeks ;D
it draws a whopping 54 watts average
Oops... Just to clarify, the entire system, including power supply losses, draws that much. The CPU itself, from what I've read (published numbers seem to vary a LOT, and I'd love to see some hard data on the min, mean, and max draw of the 90nm Athlon 64s), only eats between 7 and 35W (for comparison, the Pentium III line came in at the low 30s) with a theoretical max somewhere in the 60W range.
Kinda funny, actually... When everyone talks about needing bigger and better power supplies, with 400W considered a bare minimum and 600W not all that uncommon these days, I upgraded from an old P-III system and the total power consumption of the system dropped by half.
Well, the whole point for Cool & Quiet is that this can be done dynamically during run-time. It's quick and painless for your fast processor to slow down when you don't need the horsepower. The slower processor can't suddenly become faster if you need it.
Use CrystalCPUID to manage your AMD64 CPU's speed and voltage rather than the default Cool 'n Quiet power management (set your Power Scheme to "Always On" to disable that, definitely leave the CnQ driver installed). On most HP AMD64 notebooks we've found that you can usually safely set the core voltage at about 0.2V below stock at full speed. Judging by the AMD Thermal Design Guide, that's enough to cut power consumption nearly in half. I swapped in a Mobile-class Athlon 64 3200+ into my Pavillion zv5000z in place of the stock DTR-class chip and have been running 1GHz at 0.8V, 1.6GHz at 1.025V, and 2GHz at 1.225V for months. That puts the full speed power consumption at slightly above AMD Turion ML levels. For the stock DTR chips, 1.3V at full speed is popular.
Of course, in average use, the standard AMD Cool 'n Quiet behavior of running 800MHz at 0.95V while idle will give you battery life that's almost as good as an undervolted setup. 3-4 hours of battery life with a 12 cell battery is common, versus a fraction of that for the poor bastards who bought the P4-based zv5000 series (HP wisely dropped Intel CPUs from their zv6000 line). Undervolting does wonders under heavy CPU load though.
MobileMeter is my favorite way to monitor CPU speed and temperature, and Hot CPU Tester Pro verifies that I didn't go too far.
My system is never idle. It runs seti@home and/or folding@home 24/7 in the background. So I don't think the power saving features will work for me if they depend on the processor being idle. I bought a Dell 500SC for home. It has been rock solid, but the fan is very noisy, and the DMA on the secondary IDE is busted (chipset bug). When I upgrade, I don't care about bleeding edge performance, I want it to be quiet. Wouldn't you know, after I bought the 500SC, Dell came out with the 400SC, which I've installed at several customers. That thing is quiet as a mouse. Sigh. I thought about switching and telling them, "See, 500 is better 400!"
If you haven't already bought (an AMD Athlon 64 4000+), just make sure to get one with a Winchester or Venus core.
/. ID. It's a law.)
Nitpick: the 90nm 4000+ is a San Diego core. 1MB L2 cache is San Diego, 512MB L2 90nm E3 core is Venice, D-series core is Winchester (older 3000+ to 3500+). (You have to be this geeky to get a 4-digit
I did the same thing you did. I've got a Winchester core 3000+ in my 64-bit Fedora Core server. You can cut power consumption even more with a high efficiency power supply, Seasonic S12's being the absolute best (Newegg carries them). They made a very noticible difference over the Antecs I used to use. Using a 6600GT rather than a 6800GT video card made a huge difference too.
on an AsRock motherboard, yes? I've got two
2600+ systems with identical heatsink/fans.
The one in a MSI KM2M motherboard is rock-solid
stable at full FSB speed (133MHz), but the one
in the AsRock K7VT2 has to be underclocked to
130MHz FSB or else it constantly locks up.
I'm guessing your problem was never cooling,
it was getting stuck with a cheap mobo, as I did.
Ah, bitter dregs.
what I don't see very often is reviews address all the other sounds in a case, like the damn hard drives. I never hear my fans, system is water-cooled with two 120mm fans at 5v, but all nite all i hear is GRINNNNNNNNNNNNNNNNNNNNNND of the damn hard drives. Why doesn't someone address this issue and do a REAL review on how to get rid of hard drive sounds? Sure silentpc has done a few, but everyone else is like "yeah, i hear like, a fan, sometimes, so i'm gonna run my new 4000+ processor at 800mhz".
talk about unoriginal....
my karma will be here long after I'm gone
I think the power supply requirements are more based on a gamer rig than your everyday Joe workstation. Incredible video cards, multiple hard drives (possibly in RAID configuration striped for speed,) overclocking, cooling systems, and then bling (flourescent lights, etc) all suck down power. I doubt that a power supply will always draw it's peak power, so having a litttle headway is worth it to keep the system a little stable.
I'll never make that mistake again, reading the experts' opinions. - Feynman
PC Stats actually took this idea from me. I took the same processeor, unplugged the computer (to lower power consumption) and then stuck it in a freezer (to lower temperature). Wow! Also nil power consumption (the freezer uses some to keep it cool) and the temp is really low! Next I'm going to try this with my laptop because the bottom of it gets really hot!
Not a bad strategy, but it'd be much more efficient if you watercooled the CPU rather than aircooling it (which is what you effectively do when you put it in a freezer). Just take your case and place it in the bathtub (you can even soak in the bath with it if you want!). As far as lower power consumption goes, you'll notice a huge power spike for the first tenth of a second or so, but afterwards power consumption effectively drops to zero (along as your heart rate, if you're in the tub).
Just my $0.01999999974981
--
The article states:
Unless the computer and participants were in a frozen room (at 0 degrees celcius), their analogy is flawed. The amount of heat generated is directly preportional to the temperature INCREASE above the ambient temperature. Let's assume that the test occurred at "room temperature" (70F deg or 21C deg). The chart would look more like the one below:The article should have stated:
"For a 66% drop in speed, there was a 53% drop in added temperature."
"a 43% drop in voltage produced a 53% drop in in heat seems more reasonable."
My observation from that data above:
"A drop of only 400MHz (17%) and 0.15V (11%) showed a significant drop in the amount of heat generated (25%)."
I am extremely glad to see "underclocking" gaining in popularity.
I design systems that run applications across hundreds or thousands of servers. Many of my applications are bound by items such as connections, long before processor becomes a bottleneck.
As a case example, I will have an application that utilizes 55% of the proc across two processors. I use two processors to keep response time down (multi threads). Intel gives me a new processor. I get to spend more money to power the new processor, but now I get the amazing advantage of the new, faster, more power hungry machine now being 30% utilized.
More money down the drain, but I am not getting much for it. The worst abuse of this is static content web servers. I run into connection issues and network latency issues long before I run out of processor.
With the new HE processors from AMD, I can turn down the processor clock and cut my power consumption by as much as 50% across the board. This translates into real savings on power and cooling infrastructure.
A fair comparison of processors would be to calculate how much processing power you get in some benchmark per Watt of power dissipated: Fan noise for getting rid of CPU heat should be lowest when the least heat needs to be gotten rid of. Probably the processor with the best performance per Watt at full speed also has the lowest disspation per Watt at lower speeds.
Thank you for letting me share this old-timer drivelling on slashdot.
Panurge has posted for the last time. Thanks for the positive moderations.
A bunch of years ago I experimented with underclocking. I had an old AST 'Bravo' 286 machine and there wasn't much interesting to do with it. (I was running Slackware on a constellation of 386sx and better machines to fool around with networking). It had a socketed 'crystal block' TTL oscillator. I had a bunch of other oscillators around so I started plugging them in.
The base machine was as slow as an AT gets, it was a 6 MHz. 286. I plugged in a 1 MHz oscillator to make it a 512 KHz '286 machine. It actually booted up, very veeeery slowly. You could count the actual steps as the BIOS did the traditional 'step the floppy drive to one end and back' sequence.
Very nice!
Then I tried some even lower-value oscillators. I have block oscillators down to a value of 32.764 KHz. The machine wouldn't boot up at all at lower frequencies.
This is because the memory on the motherboard, and indeed the registers inside the CPU themselves, are dynamically refreshed. If the chip isn't run fast enough, it crashes.
There are processors that can run down to zero hertz, with an all-static CMOS design. The Intersil/Harris 6100 processor has this characteristic. You can use a knife switch as your clock if you wish.
A slower processor, say 1/2 the clock speed, would
not necessarily draw less power than the faster
processor running at that same speed. If you were
to compare 2 processors on the same die size, the
power required at a specific clock rate determines
what speed the manufacturer rates it for. Most
modern NMOS-type chip designs draw the most power
on the rising and falling clock edges.
The cleaner and sharper rise and fall times that
the processor clock runs at, the lower the power
requirements and the faster the clock could run.
Via and trace densities inside the chip determine
what the absolute maximum power can be drawn,
without melting (like a buss fuse). The faster
processor runing at a lower clock rate should
still have the steeper clock transition times,
drawing less power.
One of the requirements of a good chip design is
the use of a clock signal distributed well. So
long as a slower clock rate can still sync up
properly between on-chip modules (like caches),
a faster processor should draw considerably less
power than the slower processor, given the same
clock speed. Manufacturing tolerances determine
what a specific 6 inch or 9 inch silicon wafer
can produce, speed-wise. Of course, the more
faster chips that can be produced reliably from
a given wafer, the more $$$ the manufacturer can
make.
Processors designed for portable, low power use
already can make use of a slower clock when in
sleep mode. Desktop systems could also make use
of the same technology to save energy. A faster
processor that is running at a slower speed may
not even require a fan, if quiet operation is
desired. The motherboard design, mb support chips,
and the BIOS must support under-clocking for this
to work.
Just my rapidly depreciating $00.02 worth.