By: Eric (lol.delete@this.safetymail.info), April 23, 2012 12:33 pm
Room: Moderated Discussions
mpx (mpx@nomai.pl) on 4/23/12 wrote:
---------------------------
>EduardoS (no@spam.com) on 4/22/12 wrote:
>---------------------------
>>mpx (mpx@nomail.pl) on 4/22/12 wrote:
>>---------------------------
>>>"Steamroller" 3-rd generation modular core: greater parallelism.
>>
>>Hum... 256 bits AVX2?
>
>Perhaps plus a decoder and SMT per 'core'? With only FPU and branch parts shared.
>
>They only have +30% of additional estate while going from 32nm to 28nm with similar
>die sizes. These three seem like an economical enough way of spending such limited resources.
There's some new information (all right, it's almost a week old, but I didn't see it until now) about the estimated performance of the Steamroller core. AMD apparently projects that core performance will increase by 33% over that of Bulldozer, and according to Softpedia: "This doesn’t mean that the CPU won’t benefit from other performance enhancing developments, such as a better memory controller or more cache with lower latencies." What kind of improvements to the micro-architecture can we expect based on that information? I especially jumped on the part that apparently excludes improvements to caches and the memory controller from those 33% given that the current consensus seems to be that most of Bulldozer's performance problems are caused by its cache hierarchy.
http://news.softpedia.com/news/AMD-s-New-Steamroller-Architecture-to-Bring-Significant-Performance-264918.shtml
http://www.xbitlabs.com/news/cpu/display/20120416165428_AMD_Expects_Significant_Performance_Improvements_with_Steamroller_Microprocessors.html
---------------------------
>EduardoS (no@spam.com) on 4/22/12 wrote:
>---------------------------
>>mpx (mpx@nomail.pl) on 4/22/12 wrote:
>>---------------------------
>>>"Steamroller" 3-rd generation modular core: greater parallelism.
>>
>>Hum... 256 bits AVX2?
>
>Perhaps plus a decoder and SMT per 'core'? With only FPU and branch parts shared.
>
>They only have +30% of additional estate while going from 32nm to 28nm with similar
>die sizes. These three seem like an economical enough way of spending such limited resources.
There's some new information (all right, it's almost a week old, but I didn't see it until now) about the estimated performance of the Steamroller core. AMD apparently projects that core performance will increase by 33% over that of Bulldozer, and according to Softpedia: "This doesn’t mean that the CPU won’t benefit from other performance enhancing developments, such as a better memory controller or more cache with lower latencies." What kind of improvements to the micro-architecture can we expect based on that information? I especially jumped on the part that apparently excludes improvements to caches and the memory controller from those 33% given that the current consensus seems to be that most of Bulldozer's performance problems are caused by its cache hierarchy.
http://news.softpedia.com/news/AMD-s-New-Steamroller-Architecture-to-Bring-Significant-Performance-264918.shtml
http://www.xbitlabs.com/news/cpu/display/20120416165428_AMD_Expects_Significant_Performance_Improvements_with_Steamroller_Microprocessors.html



