Marcin Niewiadomski (marcin.niewiadomski@gmail.com) on 11/9/07 wrote:
---------------------------
>Hi David,
>
>I've an impression that I just saw first "benchmark-style" >article on RealWorldTech.
>It looks quite nice I think, but I have a bit mixed >feelings...

So it's not actually the first, I wrote up an article on Dempsey, Woodcrest, Clovertown and now Harpertown.

>1. On one side it could be analyzed in more detail (please see remarks below),
>but on the other hand - I like RWT for being "very >technical" site (especially I
>like the very good architectural drawings).

The drawings are a lot of work, but they sure help to write the article.

>2. You're right that it would be better compare a bit >different processors, as
>2.33 Clovertown has 7x multiplier, while 3.0 Hampertown has >7.5x, additionally we
>would need to ensure that memory latencies are same (in >terms of cycles - maybe
>it would be good to add latency data to table on page 3?). >Still I think that the
>difference in multiplier is not big enough to question Your >estimates.

So the issue isn't the multipliers, it's the clock speed relative to memory.

>3. As I'm not entirely sure how snoop filters are >implemented - does 24MB snoop
>filter means it is capable of work with 24MB of L2 cache >total or it actually has
>a 24MB buffer inside? (btw: it would be also nice to know >what is a real-world impact
>of four affinity groups in snoop filter)

It means that the snoop filter can index 24MB of cache data. Basically the snoop filter replicated the tags of all the L2 caches, with status information (MESI).

The real world impact is that it reduces the number of snoops. When you miss in cache, you have to snoop all other caches in the system to see if they have the data.

The snoop filter acts as a directory for the caches in the system...one of my previous articles on Blackford had more detail:

http://www.realworldtech.com/page.cfm?ArticleID=RWT110805135916&p=3

Cheers,


David