By: Rumi Naik (easy.to.delete@this.guess.anyway), May 6, 2012 7:45 pm
Room: Moderated Discussions
Let's be civil. I strongly doubt that any of those factors matter to the degree you seem ti believe. The reason is that nearly all mostly feedforward network optimizations experience diminishing returns when the cost if communication is nonzero. The shorthand textbook form if this is Amdahl's law, but it's far deeper than that.
Wilco (Wilco.Dijkstra@ntlworld.com) on 5/6/12 wrote:
---------------------------
>anonymous (no@spam.com) on 5/5/12 wrote:
>---------------------------
>>>>40% better IPC on which kind of benchmark?
>>>>Or is it only theoretical IPC?
>>>
>>>The 40% is an average claimed by ARM, likely based on Spec.
>>
>>In English that's called making s*** up.
>>
>>Show me the SPEC results... or STFU.
>
>So you can't grasp that doubling fetch bandwith, increasing decode width by 50%,
>quadrupling reorder queues, doubling load/store bandwidth and significantly improving
>L2 could possibly give 40% performance?
>
>If you have evidence the gain is less than 40%, show it, otherwise just STFU.
>
>Wilco
Wilco (Wilco.Dijkstra@ntlworld.com) on 5/6/12 wrote:
---------------------------
>anonymous (no@spam.com) on 5/5/12 wrote:
>---------------------------
>>>>40% better IPC on which kind of benchmark?
>>>>Or is it only theoretical IPC?
>>>
>>>The 40% is an average claimed by ARM, likely based on Spec.
>>
>>In English that's called making s*** up.
>>
>>Show me the SPEC results... or STFU.
>
>So you can't grasp that doubling fetch bandwith, increasing decode width by 50%,
>quadrupling reorder queues, doubling load/store bandwidth and significantly improving
>L2 could possibly give 40% performance?
>
>If you have evidence the gain is less than 40%, show it, otherwise just STFU.
>
>Wilco
