ISSCC: 90nm PowerTune revealed
According to Heise online (one of Europes leading tech mags), IBM has revealed PowerTune details of the 90nm PPC 970.
According to the article, a PowerTune 970 can scale back its frequency in three steps to 1/2, 1/4 and 1/64th of max clockspeed. Additionally, in the 50% and 25% stages, the chip can lower the core voltage to 1V to conserve energy.
In the 50% and 25% speed setting, the chip can go to "nap mode", whereas the 1/64th speed is a sleepmode drawing minimal power but keeping the functional units ready to go to full speed without much delay.
Chip size is 62mm^2 (down from 118mm^2), typical power consumption is 50W @ 2.5Ghz (down from 66W @ 2Ghz). Running on the 25% speed (625Mhz) setting with lower core voltage, it consumes just 12W. This is for the 2.5Ghz part.
Link (german only, sorry).
According to the article, a PowerTune 970 can scale back its frequency in three steps to 1/2, 1/4 and 1/64th of max clockspeed. Additionally, in the 50% and 25% stages, the chip can lower the core voltage to 1V to conserve energy.
In the 50% and 25% speed setting, the chip can go to "nap mode", whereas the 1/64th speed is a sleepmode drawing minimal power but keeping the functional units ready to go to full speed without much delay.
Chip size is 62mm^2 (down from 118mm^2), typical power consumption is 50W @ 2.5Ghz (down from 66W @ 2Ghz). Running on the 25% speed (625Mhz) setting with lower core voltage, it consumes just 12W. This is for the 2.5Ghz part.
Link (german only, sorry).
Comments
Originally posted by Smircle
Running on the 25% speed (625Mhz) setting with lower core voltage, it consumes just 12W. This is for the 2.5Ghz part.
Some estimation for the 2 GHz part at 25% speed?
Originally posted by Smircle
... whereas the 1/64th speed is a sleepmode drawing minimal power but keeping the functional units ready to go to full speed without much delay.
Now, what's that? Does it means that if the computer does almost nothing (for example when you are reading a static document, scrolling from time to time), the processor almost shuts down, but it is ready to raise its performance if needed? If it is like that, then this would make an awesome Powerbook.
Originally posted by PB
Now, what's that? Does it means that if the computer does almost nothing (for example when you are reading a static document, scrolling from time to time), the processor almost shuts down, but it is ready to raise its performance if needed? If it is like that, then this would make an awesome Powerbook.
The newsticker-item I got the info from is not very clear on how this feature is going to be useful, but it could also be targeted at routers and other network gear. If no traffic is detected, the CPU falls to the nap-mode and does nothing but regularly check the ports for incoming traffic and come back to live if it detects a packet.
I am not so sure if a Unix-like OS can make use of this except for putting the computer to sleep and pulse the LED - however it sure is interesting that the article talked about waking from nap-mode being almost instantanous - this could make it indeed useful for reading text on a Notebook.
Originally posted by costique
This must be a good compromise between better performance and lower heat dissipation. At least, in 'normal' (or average use) the CPU will have some time to laze.
I thought I had read about PowerTune including tech to individually shut down non-active functional units? Like if all instructions in que were FP, shut down all non-FP parts of the chip.
Originally posted by Smircle
. . . I am not so sure if a Unix-like OS can make use of this except for putting the computer to sleep and pulse the LED - however it sure is interesting that the article talked about waking from nap-mode being almost instantanous - this could make it indeed useful for reading text on a Notebook.
Someone else suggested that Mac OS 10.3.3 will take advantage of new features in the 970FX, but did not say or know what these features are.
Originally posted by Smircle
According to Heise online (one of Europes leading tech mags), IBM has revealed PowerTune details of the 90nm PPC 970. . .
From the details revealed, we can figure out some things about the 970FX and how it is used. The chip running at 2.5 GHz is operating at a higher voltage than the chip reported on earlier, which indicated between 24 and 25 Watts at 2.0 GHz and 1 Volt. According this article, the chip can reduce voltage to 1 V to conserve energy, meaning that it requires greater than 1 V to achieve 2.5 GHz. The Xserve appears to run the 970FX at 1 volt to reduce power in its thin 1U server enclosure, but the Power Mac should have no problem running it wide open at 2.5 GHz and 50 W power dissipation.
The article also says at 1 V and 1/4 clock rate (625 MHz) the power drops to 12 W. Combining this data point with the earlier report, we can estimate that leakage current dissipates about 6 W with a 1 V supply. So we can get other power estimates, like 18 W at 1 V and 1/2 clock rate, or 1.25 GHz.
If we make a guess that the higher voltage is 1.2 V, this gives us a leakage current dissipation of 16 Watts at this voltage. I don't know whether this a reasonable number or not. ( If the voltage is higher than this, the calculation would give lower leakage to get the 50 W stated for 2.5 GHz.) Using this guess, however, the power dissipation at 3.0 GHz would be just under 60 W, meaning that the Power Mac would have no problem when IBM gets chips that yield this performance.
Originally posted by snoopy
From the details revealed, we can figure out some things about the 970FX and how it is used. The chip running at 2.5 GHz is operating at a higher voltage than the chip reported on earlier, which indicated between 24 and 25 Watts at 2.0 GHz and 1 Volt. According this article, the chip can reduce voltage to 1 V to conserve energy, meaning that it requires greater than 1 V to achieve 2.5 GHz. The Xserve appears to run the 970FX at 1 volt to reduce power in its thin 1U server enclosure, but the Power Mac should have no problem running it wide open at 2.5 GHz and 50 W power dissipation.
The article also says at 1 V and 1/4 clock rate (625 MHz) the power drops to 12 W. Combining this data point with the earlier report, we can estimate that leakage current dissipates about 6 W with a 1 V supply. So we can get other power estimates, like 18 W at 1 V and 1/2 clock rate, or 1.25 GHz.
If we make a guess that the higher voltage is 1.2 V, this gives us a leakage current dissipation of 16 Watts at this voltage. I don't know whether this a reasonable number or not. ( If the voltage is higher than this, the calculation would give lower leakage to get the 50 W stated for 2.5 GHz.) Using this guess, however, the power dissipation at 3.0 GHz would be just under 60 W, meaning that the Power Mac would have no problem when IBM gets chips that yield this performance.
Well, Intel's 90nm Prescott is using 30% of its ~100 Watt power budget as leakage current. 16 Watts out of 60 is in that same ballpark.
Originally posted by snoopy
After posting, I thought I'd better check it out to confirm my memory, but I cannot find the report.
You can't find it because it's not there.
IBM did not report the voltage to achieve 12.3W @ 1.4GHz. The 0.85V is an estimation made here in the AI forum.
Here is IBM's reference (page 9).
We knew that it could scale frequency, although I wasn't expecting anything quite that variable.
But of course, it can also scale its power consumption in specific areas of the chip depending on what's being used.
It's actually when you combine all of the 'innovation' in 970FX, you start to wonder when it ONE will appear in a Powerbook, but you also start to wonder how long it will be before Apple tries to stuff FOUR of them into something - either workstation or server, or possibly a toaster.
Originally posted by Mark- Card Carrying FanaticRealist
But of course, it can also scale its power consumption in specific areas of the chip depending on what's being used.
How do you know that?
It's actually when you combine all of the 'innovation' in 970FX, you start to wonder when it ONE will appear in a Powerbook...
Now, can anyone explain to us how possible is to put a 970FX with PowerTune on a Powerbook with or without a SoC design? How it would compare against today's Powerbooks in power consumption and performance? Against Pentium-M and Pentium 4-M?
Originally posted by smalM
. . . IBM did not report the voltage to achieve 12.3W @ 1.4GHz. The 0.85V is an estimation made here in the AI forum.
Here is IBM's reference (page 9).
Thanks for the link to that report. The data point I used for the calculations looks okay however, the 24.5 W at 2.0 GHz and 1.0 V. If the 12 W at 625 MHz and 1.0 V is a fairly good number, my estimate of 6 Watts leakage dissipation seems reasonable.
The report does give us a clue about the higher voltage level, to get to 2.5 GHz. It is likely 1.3 V rather than 1.2 V, which I simply guessed. That being the case, the leakage power dissipation at 1.3 V would be about 12 W, not 16 W. This assumes that 50W at 2.5 GHz is also a good number. The estimate of just under 60 W at 3.0 GHz still holds.