While Intel (INTC) has stopped trying to maximize CPU clock speeds in favor of adding cores and...

This news story has 4 comments:

Register or Login to rate comments »

AlSabawi Comments (29)
 
I think these processors are liquid cooled and massive in terms of floor footprints. Their sale is limited to government research centers and big university. The fact is CPU raw speed follows the law of diminishing marginal returns; it is more efficient to add processor count and use distributed/parallel software than to increase each processor's speed.
3 Aug 2012, 07:00 PM Reply Like

AM1000 Comments (7)
 
I think it's cool to use 100 powerfull UNIX / mainframe servers instead of 100000 or more x86 servers for parralel processing. This will save power, real estate, administrative overhead, and engineering support efforts.
3 Aug 2012, 07:19 PM Reply Like

eh.geee@gmail Comments (14)
 
All of these devices have to exist not in a laboratory or a thesis paper, but in the real world of an energy conscious server farm facility or in a laptop or desktop machine. Moore is not yet "over" but the "next" curve of technology price/performance will need need horses to replace the humble electron with something less bulky.
4 Aug 2012, 03:47 AM Reply Like

jtemple Comments (3)
 
Well, Alsabawi and AM1000 make valid points in the context of their point of view.

I think water cooling is irrelevant to the point here which is why rather than how this is done. Also I think you will find that machines using these chips are in large commercial rather than large research facilities. The fact is that clouds depend on virtualization consolidation for their effectiveness. Virtualization consolidation effectiveness depends on thread capacity which is a combination of thread speed and cache per thread. If you build more cache you have less design space for threads, if you have fewer threads it is worth more to have faster threads.

The second point I would like to make is that higher thread speed is required when code contains serializations. Code which shares data or resources inherently contains serializations. This will vary with the work to be done. Blue Gene, Watson, Pure, and Sysplex are all parallel clustering solutions. There is very good reason for the differences in the way the machines are designed. Otherwise, no one would buy them and IBM would have been out of the hardware business a long time ago.
4 Aug 2012, 04:56 PM Reply Like