Replies to post #52730 on Advanced Micro Devices Inc (AMD)

[wbmw]

Re: Then you have a misunderstanding of 'logic'.

Sorry, Doug, the misunderstanding is on your part. Let me help you out.

Re: I stated that the 2.8-3.2GHz Smithfields are to have 130W TDP.

Point #1: You admit Smithfield dual core is not 2x the power of single core Prescott, but rather around 55% higher, according to your estimate (based on the 84W stated TDP of the 3.2GHz processor).

Re: then dual core 90nm Prescott at 4GHz is about 300W TDP.

Point #2: Now you want to apply a 2x power increase.

Re: Single core 90nm Prescott at 4GHz would have 150W TDP, per Paul Otellini.

This is just ridiculous. You bring up an obscenely high value that you refuse to back up with proof and you try to apply it in a way that renders it meaningless.

Re: Now, at the same speed, how much power do you think 65nm will save over 90nm?

You can't estimate a power reduction from one process to the next using the same frequency, as easy as you can estimate a frequency improvement from one process to the next using the same power.

The reason is that leakage is now enough of a contributor to power dissipation that adding frequency increases power more than just linearly. Consider the step function between a 3.2GHz Prescott (rated at 84W TDP) and a 3.4GHz Prescott (rated at 115W TDP), followed by a complete flattening in power up to 3.8GHz (still rated at 115W).

This is because Intel bins their chips based on leakage levels. The 3.4GHz/84W parts are ultra low leakage parts, as are the 3.8GHz/115GHz parts. Meanwhile, the 2.8GHz/84W and 3.4GHz/115W parts are higher leakage parts. By moving to 65nm, the distribution of power is going to shift according to this graph:

ftp://download.intel.com/research/silicon/65nm_logic_press_briefing_0804.pdf (slide 12)

It shows that at the same leakage, drive strength (key component to switching speed, i.e. frequency) can increase from 0.6 to 0.7mA/uM (about 17% improvement). Meanwhile, leakage at the same drive strength decreases from 100 down to 20nA/uM (5x reduction). This shows us that you can reduce power a helluva lot more at the same frequency than you can increase frequency at the same power.

Like I said before, I expect a 25% or greater improvement in frequency at the same power from 90nm to 65nm. If Intel wanted, they could vastly reduce power at the same frequency, but that doesn't offer as much of a competitive prospect for them than increasing frequency.