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: Scumbria who wrote (94208)2/19/2000 1:50:00 AM
From: Tenchusatsu  Read Replies (1) of 1570817
 
Scumbria, <It is nonsensical these days to do anything else other than a deep pipeline.>

That didn't seem to be the theme during last year's MPF. I got the impression that long pipelines were looked down upon because of the overhead.

<A deep pipeline will have more bubbles, but each bubble is of shorter duration, so the branch mispredict penalty is not much different in a deep pipeline than it is in a short pipeline.>

I'm just assuming that when you double the length of the pipeline, you double the size of the pipeline bubbles and double the cycle penalty incurred by branch mispredicts. I'm also assuming that doubling the length means shortening the clock cycle by 33% (not 50% because of the overhead). Taking those assumptions into account, the time penalty will increase by 33%.

Like I said before, that's a good reason why Intel had to implement features to reduce the penalties in any way possible. Sure, it's fun to jack up the clock speed by going with a longer pipeline. And MHz (or in this case, GHz) is a great seller. But if the processor doesn't scale well in actual performance because of the overhead and penalties, that would be a huge embarassment.

After all, I sure don't want to hear someone say this, "Oh look, Willamette at 1.5 GHz can hardly outperform an Athlon at 1.1 GHz!"

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