re: Intel Server Strategy Shift with Sandy Bridge EN & EP

Clock$peedy — Thu, 24 May 2012 16:43:11 GMT

Hi Joe,

To the extent that you seem to be suggesting that Intel might replace the E7 line (Westmere-EX) with high-end implememtations of Sandy-Bridge EP (E5), I think you're mistaken and missed the main point of E7. In a world where computational power is no longer the bottleneck for the overwhelming majority of enterprise apps (and has not been for years), 'In-memory database' technologies and techniques represent the future.

http://blogs.technet.com/b/dataplatforminsider/archive/2012/04/09/the-coming-in-memory-database-tipping-point.aspx

In order to do IMDB reliably you need serious RAS features in the memory subsystem. E7 has them and E5 does not.

On the other hand, four-socket Sandy-Bridge EP systems are really good for HPC!!!

re: Intel Server Strategy Shift with Sandy Bridge EN & EP

jchang — Fri, 25 May 2012 20:59:49 GMT

Intel has replaced the E7 with the E5. The E5 will take the majority of 4-way system, which is the large majority the E7 volume. I am further inclined to think 2-way E7 volume (both system and processor) is higher than 8-way. So what else defines E7 once common 4-way moves to E5.

1) glue-less 8-way

2) MCA & other RAS

3) memory capacity (4 channels per socket + SMB splitter for 16 DIMMs/socket)

Like I said, 8-way volume is very small (but important).

I really don't think people understand the E7 MCA + RAS to be relavent.

Now that E5 has 4 MC, the difference in memory capacity is not important.

My opinion is that 8-way could be a glued system, ie, a QPI cross-bar.

If MCA+high RAS are important, it could be a once every other year or even every third year release. So E7 on Westmere skips Sandy Bridge, and gets a refresh with Ivy Bridge or even Haswell.

Sure, database engine really needs a complete architecture overhaul, which would reach into both processor and system level. But I do not agree that a 4-way system needs more than 32-64 DIMM sockets. If MS or some one else can do IMDB, they can also do clustered IMDB, which means memory capacity is distributed across many nodes. The general idea is $20K of CPU is paired with about $20K of memory, not 20/200 split.