anon (a@b.c) on 2/7/09 wrote:
---------------------------
>>Here's gratuitously piling on:
>>
>>1) IPF QPI and Xeon QPI don't use the same coherency
>>protocol (Xeon uses a simplified snoopy scheme suitable
>>for 2s/4s, IPF uses a more complex scalable distributed
>>directory system). Where do you think most of the "high
>>level" complexity is.
>>
>
>This is hard to believe.
>
>a) Nehalem MPs will have 8 cores/socket. It need more scalable interconnect even
>at 2-4 sockets.

More scalable than what? How are you quantifying "need"?

> Isn't the whole point of moving to QPI is to get rid of the old broadcast snoopy scheme?

No. There are several elements to it. For x86 sized systems, IMO the most important thing is low latency ODMC, and scaling memory capability per socket. P2P interconnect is simply the obvious choice for small system sizes after going to memory controller per socket.

Snoopy cc is surely the fastest and cheapest choice for 1-4s systems (maybe even to more sockets, given a well connected architecture and sufficient snoop bandwidth).

>b) If there two design centers working on IPF, why wouldn't Intel reuse the work
>in Xeon? A few thousand people cost real $$. It doesn't seem possible to sustain that many people
>on a low volume product.

I find it hard to believe you are so familiar with Intel's finances as to make that judgement.