SI
SI
discoversearch

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor.  We ask that you disable ad blocking while on Silicon Investor in the best interests of our community.  If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.
Politics : Formerly About Advanced Micro Devices

 Public ReplyPrvt ReplyMark as Last ReadFilePrevious 10Next 10PreviousNext  
To: Mani1 who wrote (102474)4/6/2000 1:42:00 AM
From: Jim Davis  Read Replies (1) of 1575424
 
Mani,

Re: "I still do not understand why there is more "branch miss-prediction penalty"
if the CPU is more deeply pipelined. At what point it is no longer beneficial
to have a deeper pipeline?"

The branch mispredict penalty starts when the processor realizes it executed
the wrong code branch. It ends when the correct instruction is finally executed.
Since everything in the pipeline is thrown away when the mispredict is detected
nothing comes out of the execution unit until the correct instruction gets
executed. This is called a "bubble" and is the penalty in cycles. Deeper pipelines
have more cycles between detecting a mispredict and recovering to execute
the correct instruction. This is all in general terms. No need to go into
all the special cases.

You can do a SIMPLE tradeoff on the pipeline depth using the 3 factors I mentioned.

Branch performance = (1 - prediction accuracy) * mispredict penalty * 1/frequency

Lower is better. Real computer architects would cringe of course. They do
thousands of tricks to optimize performance for their pipeline. OK, maybe
just hundreds.

This only accounts for branches. You then have to decide the percentage of branches
in the code (15-25%) calculate total performance of branches and straight code combined.

JD
Report TOU ViolationShare This Post
 Public ReplyPrvt ReplyMark as Last ReadFilePrevious 10Next 10PreviousNext