By: EduardoS (no.delete@this.spam.com), April 28, 2012 9:46 am
Room: Moderated Discussions
Paul A. Clayton (paaronclayton@gmail.com) on 4/28/12 wrote:
---------------------------
>Based on http://semiaccurate.com/forums/showpost.php?s=4177be421d79a8dbc2b3a7ceb3e5cd7d&p=143476&postcount=16
>he appears to mean the issues of memory
>dependency, particularly with respect to unknown store
>addresses. Providing store alias prediction 'weakens' this
>"dependency model".
>
>(The relevant quote: "The current strong dependency model
>of x86 doesn't permit high efficiencies if you go above 2
>pipes for scaler execution, and in BD you have already 4
>pipes for this INt. If you employ some data speculative
>technics and some value locality technics and ease those
>dependencies with prediction (alias/other) is quite
>possible you'll manage to fill 4 or 5 pipes with good
>efficiency")
Weren't this supposed to be solved by speculative loads?
Well... For the BD case, it is actually only able to sustain 2 IPC, no need to worry about 4 or 5 "pipes", the AGUs are way too limited...
---------------------------
>Based on http://semiaccurate.com/forums/showpost.php?s=4177be421d79a8dbc2b3a7ceb3e5cd7d&p=143476&postcount=16
>he appears to mean the issues of memory
>dependency, particularly with respect to unknown store
>addresses. Providing store alias prediction 'weakens' this
>"dependency model".
>
>(The relevant quote: "The current strong dependency model
>of x86 doesn't permit high efficiencies if you go above 2
>pipes for scaler execution, and in BD you have already 4
>pipes for this INt. If you employ some data speculative
>technics and some value locality technics and ease those
>dependencies with prediction (alias/other) is quite
>possible you'll manage to fill 4 or 5 pipes with good
>efficiency")
Weren't this supposed to be solved by speculative loads?
Well... For the BD case, it is actually only able to sustain 2 IPC, no need to worry about 4 or 5 "pipes", the AGUs are way too limited...



