Wbmw:
Leakage is also extremely dependent on Transistor parameters like channel length, channel length to width ratio and Ft the maximum switching speed of the circuits. All else being equal, slightly longer channels decrease leakage exponentially. A 10% channel length increase can decrease leakage by half. The disadvantages are slighty lower on current of about 4 to 8% which slows the Ft by about 4 to 8%.
Dothan is designed for lower power operation and probably needs only get to 2.2 to 2.4GHz in the long run. It also has more logic delays per stage and thus the transistors need only be 80 to 90% as fast as the desktop P4s. Thus they can trade a 20% longer channel and still get to their target speeds. This cuts leakage to about 25%. The dynamic switching and such probably cuts this further. Opteron is in the same area. It probably can handle a channel length increase in the 20% range, but increases channel width by 10% and gets back most of its speed. It also gets low leakages 25% of 110% is still low at 27.5% compared to 100%. You give up 20% more die area, but you gain lower power operation and higher clock rates in return. In the caches, it probably allows for even a longer channel length which decreases its density, but reduces leakage even more.
Prescott however, was designed for maximum clock rate. It has 31 stages over the 20 stages in the Northwood execution pipeline. It probably needed 10% shorter channel lengths to get the speed up and it doubled the leakage. Now you have a sieve for the quick CPU. Originally the process guys probably stated that the leakage constant was half that of what it turned out to be. If it was half like they thought, Prescott would have been well past 4GHz by now. Now, they are lucky to match Northwood with reasonable power usage.
As you can see design plays a big part in this. AMD went about this at two times so far. TbredA to TbredB (5% more die area) and going to SOI (10% more die area). Yes their chips were larger, but power was not a problem. In fact it is so good, that 130nm Athlon 64s have as little power usage as Banias at similar speeds with vastly more performance. They are near enough to Dothan to make Intel sweat (256KB Mobile Semperon K8 may be there at roughly the same performance (faster in games and office apps, slower in some encoding and decoding work)). Intel is really sweating AMD's 90nm performance and power usages and that is before 64 bit and NX.
Pete
