Article: AMD's Mobile Strategy
By: Exophase (exophase.delete@this.gmail.com), December 20, 2011 9:05 am
Room: Moderated Discussions
Bill Henkel (noemail@yahoo.com) on 12/19/11 wrote:
---------------------------
>Tom P (tp5@gmail.com) on 12/19/11 wrote:
>>Are you sure it is possible to make a 6 GHz dual core chip?
>
>If it is possible to make a 3.1 GHz 8-core 150W 32nm chip (SandyBridge-EP), it
>surely must be possible to make a 6 GHz 2-core 200-300W 28nm chip. Doubling the
>number of pipe stages would be the simple way to do it but since you would have
>4x the transistors per core, it wouldn't be necessary to do that. AMD could make
>a lot more profit selling $1K processors to gamers instead of $100 mainstream processors
>even though the market for $1K processors is 1/10th as big. AMD could sell an ultra
>low voltage version of this core (perhaps with some adjustments to threshold voltages) in the mobile market.
>
Current Bulldozer CPUs could probably be binned to hit 6GHz on water cooling with at least 1 module (2 cores). AMD was able to exceed 8GHz with one module on liquid helium. But I wouldn't be that surprised if even with only one module it takes way over 200W. Look at how power consumption explodes with much more modest overclocks of the platform.
The market for high end processors for gamers is probably much smaller than 10% because you'd need at the very least a very expensive GPU setup to not be GPU limited in most games (with a much weaker CPU than what you're suggesting). And a lot of games now can at least take significant advantage from more than two cores, if not really approach a balance over 4+.
Optimizing for frequency doesn't strike me as good for perf/W because at lower clock speeds you suffer worse IPC and end up burning more power on the extra transistors spent on allowing higher clock speed. That's why I question AMD's strategy of designing BD for high clocks but deploying it with low clocks/many cores in the server world, with the ability to turbo to increase single threaded performance.
---------------------------
>Tom P (tp5@gmail.com) on 12/19/11 wrote:
>>Are you sure it is possible to make a 6 GHz dual core chip?
>
>If it is possible to make a 3.1 GHz 8-core 150W 32nm chip (SandyBridge-EP), it
>surely must be possible to make a 6 GHz 2-core 200-300W 28nm chip. Doubling the
>number of pipe stages would be the simple way to do it but since you would have
>4x the transistors per core, it wouldn't be necessary to do that. AMD could make
>a lot more profit selling $1K processors to gamers instead of $100 mainstream processors
>even though the market for $1K processors is 1/10th as big. AMD could sell an ultra
>low voltage version of this core (perhaps with some adjustments to threshold voltages) in the mobile market.
>
Current Bulldozer CPUs could probably be binned to hit 6GHz on water cooling with at least 1 module (2 cores). AMD was able to exceed 8GHz with one module on liquid helium. But I wouldn't be that surprised if even with only one module it takes way over 200W. Look at how power consumption explodes with much more modest overclocks of the platform.
The market for high end processors for gamers is probably much smaller than 10% because you'd need at the very least a very expensive GPU setup to not be GPU limited in most games (with a much weaker CPU than what you're suggesting). And a lot of games now can at least take significant advantage from more than two cores, if not really approach a balance over 4+.
Optimizing for frequency doesn't strike me as good for perf/W because at lower clock speeds you suffer worse IPC and end up burning more power on the extra transistors spent on allowing higher clock speed. That's why I question AMD's strategy of designing BD for high clocks but deploying it with low clocks/many cores in the server world, with the ability to turbo to increase single threaded performance.
