Slashdot Mirror
Cooling Down Hot Processors
DonnaMai writes "Face it: the only scorching hot thing you want with a chip is salsa. Any other overheating is potentially counterproductive, and can be downright damaging to the microprocessor -- or other components. This article uncovers potential ways to chill the chips."
We already have that. It's called Speedstep and it's on Pentium M processors.
E = m c^3 Don't drink and derive E = m c^3
It's called the Crusoe Processor by Transmeta. And yes, it is an x86 processor.
Fly me to the moon Let me sing among those stars Let me see what spring is like On jupiter and mars
I guess that is available, you have nothing to wait.
The AMD Athlon64 3500 consumes 20W of power on low load and 69W when you stress it. Most of the time it will be pretty cool.
Andrew
That sounds like a bunch of extra logic at each flipflop. Which will reduce speed and increase the size of the chips.
Their problem is the megahertz myth...
They've been pushing this for so long, that they can't look back now. Yes, the Pentium M's perform great, but they're still only around 1.7 GHz.
While that out-performs P4's with MUCH higher clock-speeds, what are they going to say? Buy this CPU, it's 1.7GHz! Joe Sixpack would say "But I can buy this here Penteeyum Four with 4 GHz... 4 is better than 1.7."
AMD has been rating their CPUs on performance to keep competitive with Intel's. If anything, Intel will have to follow their lead and do the same if they really want to push the Pentium M's to the masses.
But, as the voltage levels drop, the leakage current through the transistors increases. At some point, dropping the voltage does not reduce the power. I think that we are pretty close to this point already.
"-1 Troll" is the apparently the same as "-1 I disagree with you."
It would --- but there would be other problems.
The first one is the most simple: silicon's expensive. Really expensive. The more units you can slice off that wafer the cheaper the units are. Making the die bigger simply for thermal reasons isn't going to wash with the chip manufacturers. They already glue the die to a metal backing plate, which gives you much the same effect anyway.
The second one, however, is the most crucial one. Electricity is slow. Electrical impulses travel at about 2/3 c through copper and a touch less through silicon (IIRC, I can't find the figures to check). This means that the bigger your die is, the longer it takes the impulses to travel from one side of it to the other.
A 1GHz clock fires every 10^-9 seconds; since the speed of light is 3x10^8 m/s, this means that the impulses are going to travel about twenty centimetres between clock pulses. For a 4GHz clock, it'll be about 5cm. There's a lot more wiring than that folded up inside the die; and it gets worse --- particular things happen at particular times throughout the clock cycle, and where you are in the clock cycle now depends on how long the wire is that connects you to the clock. Making sure everything happens in sync is a nightmare.
There are solutions to all of this; asynchronous designs which don't use clocks, offloading functionality to special-purpose processors like GPUs so you don't need as fast a main processor, radically different approaches like Cell, optical transports so you can route signals through each other, etc, but basically there are loads of good reasons why you need the die to be as small as humanly possible.
Sadly, we live in a world where the OPERATING SYSTEM will soon require a 3D card to even function. (Windows Longhorn)
:-)
And that's not even an original idea from Microsoft.
OSX 10.3 (Panther) already utilizes 3D acceleration on video cards and treats all windows as textures. That's how the nifty "expose" feature works. The first time I saw that feature it was one of those "why didn't anybody think of this before" moments. I don't even own a mac, yet... but my mini is on its way.
10.4 (Tiger) will take it to the next level with Core Image, which requires a video card with programmable pixel shaders (DirectX 9 equivalent from the Windows world). This will allow, among other things, realtime filters applied to videos, images, and of course some new OS eye candy. But it's mainly there to speed up image and video processing, which is what macs are supposed to be good at.
However, a 3D card is still not "required" as the OS will automatically scale back the features that aren't supported (you will still be able to run Tiger on a 400MHz G3 iMac).
I don't know whether Longhorn will actually require 3D cards, but saying "soon" in reference to Longhorn is a tad bit optimistic, don't you think?
-CausticPuppy "Of all the people I know, you're certainly one of them." -Somebody I don't know
- Interface with USB peripherals, and hence the majority of devices which people want/use.
- Have long filenames
- Access large HDs
- Have limited access users without some novell type extension
And with word 2.0 you couldn'tThe bottom line is, despite significant advances in hardware, the "User Experience" still feels as sluggish and slow as it did in the days of Windows 3.1 on a 386.
Why is this the bottom line? The UI needs to be acceptably fast, not 1ms fast.
Can you IMAGINE how fast Windows 3.1 would be on modern hardware if the drivers existed?
And who do you think would actually care?
My Pentium-M (733) has 5 clock-speed steps.
Primer at www.silentpcreview.com
std. disclaimer: i am just a fan. ba dum bum.
When I was a kid, we only had one Darth.
Bzzt, actully you're both wrong.
I'm also an EE. True, as technology improves the voltage is lowed. The Vt must also be lowered, thus increasing leakage. But for A GIVEN PROCESS is in the method described by GP, drain-source voltage has minimal effect on subthreshold leakage. The standard equation is, assuming Vds > 100mV
Id_sub=Id0*W/L*exp[q(Vgs-Vt)/(N0*KT)]
Id0 begin the current at Vgs-Vt and N0 being the subthreshold slope
As it turns out, Anandtech mentioned this today.
ANANDTECH ARTICLE
"-1 Troll" is the apparently the same as "-1 I disagree with you."