Article: AMD's Mobile Strategy
By: anon (anon.delete@this.anon.com), December 17, 2011 4:19 pm
Room: Moderated Discussions
mpx (mpx@nomail.pl) on 12/17/11 wrote:
---------------------------
>anon (anon@anon.com) on 12/16/11 wrote:
>---------------------------
>
>>That means the same core can be used for server and client, and also the core can
>>run *real* server work (not just partitioned tpcc) well too.
>>
>
>In the future there will be more transistors at disposal (process technology going
>forward), but the same or even lower (economy, growing prices of energy sources)
>power budgets. This does not bode well for universal cores.
>
>Hybrid CPUs look better. By hybrid I mean both various types of cores optimal for
>various types of tasks, as well as lots of accelerators. This way for a given task
>done only processing elements that do it optimally will run it, the rest will be
>turned off or slowed down to keep power low.
>
>Eg. for encryption, compression, transcoding - mostly acclelerators, for troughput
>optimizing codes - throughput cores (lots of narrow ones with hordes of threads),
>for single-thread dependant cores - fewer larger cores will be awakend. For array
>operations GPGPU portions of chips could be used.
I'm talking about the general-purpose compute core for high performance server/client workloads, of course.
Even if you drop the client part of it, server is still far from EP, so SMT approach is superior to lots of small cores, IMO.
---------------------------
>anon (anon@anon.com) on 12/16/11 wrote:
>---------------------------
>
>>That means the same core can be used for server and client, and also the core can
>>run *real* server work (not just partitioned tpcc) well too.
>>
>
>In the future there will be more transistors at disposal (process technology going
>forward), but the same or even lower (economy, growing prices of energy sources)
>power budgets. This does not bode well for universal cores.
>
>Hybrid CPUs look better. By hybrid I mean both various types of cores optimal for
>various types of tasks, as well as lots of accelerators. This way for a given task
>done only processing elements that do it optimally will run it, the rest will be
>turned off or slowed down to keep power low.
>
>Eg. for encryption, compression, transcoding - mostly acclelerators, for troughput
>optimizing codes - throughput cores (lots of narrow ones with hordes of threads),
>for single-thread dependant cores - fewer larger cores will be awakend. For array
>operations GPGPU portions of chips could be used.
I'm talking about the general-purpose compute core for high performance server/client workloads, of course.
Even if you drop the client part of it, server is still far from EP, so SMT approach is superior to lots of small cores, IMO.