In "The Ultimate Underclock and Undervolt Project" article, the editor disagrees with the author's power consumption equation for a CPU with the following statement:
>" ...(Editor's note: The relationship between CPU clock speed and
> power dissipation is linear...Leo's actual...heat drop is closer to 4.5%"
This is not correct. Every CPU (and digital circuit in general) has a certain steady-state power draw due to leakage current. Decreasing the clock rate on a modern cpu certainly does _not_ result in a linear decrease in power consumption, for this very reason. I work with embedded cpus on a daily basis, underclocked to allow for battery-powered operation. For one of these, lowering the clock rate from 16mhz (yes only 16mhz) to 8, and its power consumption is 79% of nominal (not 50%). Lowering it all the way to 1 mhz reduces power only to 45% of nominal, not the 6% a linear decrease would predict. And reducing it all the way to 0.1 mhz (about as slow as it can stably function), it still uses 41% of the power the original 16mhz processor does, while running 160 times slower.
For older ICs done on .25 micron or larger processes, the leakage current term was usually rather. However, it is has grown exponentially as transistor size decreased with smaller lithography. This is one of the hottest topics in modern EE design-- at nodes below the upcoming .09 micron that Intel is now designing chips for (the P4 is, of course, on a .13 micron process), leakage current will dominate, and instead of our digital circuits being filled with on-off "switches", they're made of leaky buckets.
Article Error...well, sort of.
Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee
-
MikeC
- Site Admin
- Posts: 12285
- Joined: Sun Aug 11, 2002 3:26 pm
- Location: Vancouver, BC, Canada
- Contact:
Thanks for spotting the error. I had based this comment on long what-if sessions with the utility Radiate mentioned in the article. At least with the algorithms for MPs, T-birds, and P3s in this utility, the relationship between clock speed and power dissipation appears linear. It was obviously too broad a generalization. So you're saying it is subject to processor-specific non-linearities & no quick rule of thumb can be applied, right?
Your final statement conjurs up a rather alarming image. For the mostly non-engineers here, what does that mean in plainer language and what is the consequence?
Your final statement conjurs up a rather alarming image. For the mostly non-engineers here, what does that mean in plainer language and what is the consequence?
MikeC,
A modern Intel/Athlon CPU should be close to linear at the top end of its performance range, but even for those chips, as you scale back the clock, you'll see that you were just on the flat end of a logarithmic curve. You're right, there isn't a rule of thumb that can be applied, especially since power consumption can vary not only by the chip and its stepping version, but even by the program(s) that are being run at the time.
Regarding the problems with leakage current, its means that Moore's Law is a dying bird. Sure we'll get around the problem, and get silicon down to 0.06 or even 0.03 microns....but its a lot harder and slower than people originally thought. The leakage problem is forcing a redesign on the actual shape of transistors (they're not being made just smaller, but with a different geometry on the N-P boundaries). Also, the smaller circuits are having immense problems with cross-talk and signal integrity, making not only the chips in need of redesign, but the software tools used to develop the chips as well.
Finally, we're also seeing some issues with the materials themselves. For instance, for 20 years, we've been used to the concept of "infant mortality" on chips. If they died, they died young...otherwise they would run forever. But copper is replacing aluminum as an interconnect within the chip itself-- a technology that was heralded as the 'savior' of Moore's Law, to allow us to reduce resistance and voltage quickly. Well it does, but copper has turned out to have the annoying problem of having migrating voids in the interconnects, a problem that may turn a good chip bad only after a few thousand hours of use.
All these problems will be solved. But we're not seeing the easy gains we once did. Look at Intel itself-- a year ago exactly they introduced a 2ghz P4. By Moore's Law, a 4ghz chip should be due in 6 months. Graphics processors (ala NVidia) were a couple years ago outpacing Moore's Law. But now, Nvidia is 3 months late in taping out their NV30, primarily because of yield issues on the 0.13 micron process.
A modern Intel/Athlon CPU should be close to linear at the top end of its performance range, but even for those chips, as you scale back the clock, you'll see that you were just on the flat end of a logarithmic curve. You're right, there isn't a rule of thumb that can be applied, especially since power consumption can vary not only by the chip and its stepping version, but even by the program(s) that are being run at the time.
Regarding the problems with leakage current, its means that Moore's Law is a dying bird. Sure we'll get around the problem, and get silicon down to 0.06 or even 0.03 microns....but its a lot harder and slower than people originally thought. The leakage problem is forcing a redesign on the actual shape of transistors (they're not being made just smaller, but with a different geometry on the N-P boundaries). Also, the smaller circuits are having immense problems with cross-talk and signal integrity, making not only the chips in need of redesign, but the software tools used to develop the chips as well.
Finally, we're also seeing some issues with the materials themselves. For instance, for 20 years, we've been used to the concept of "infant mortality" on chips. If they died, they died young...otherwise they would run forever. But copper is replacing aluminum as an interconnect within the chip itself-- a technology that was heralded as the 'savior' of Moore's Law, to allow us to reduce resistance and voltage quickly. Well it does, but copper has turned out to have the annoying problem of having migrating voids in the interconnects, a problem that may turn a good chip bad only after a few thousand hours of use.
All these problems will be solved. But we're not seeing the easy gains we once did. Look at Intel itself-- a year ago exactly they introduced a 2ghz P4. By Moore's Law, a 4ghz chip should be due in 6 months. Graphics processors (ala NVidia) were a couple years ago outpacing Moore's Law. But now, Nvidia is 3 months late in taping out their NV30, primarily because of yield issues on the 0.13 micron process.