From: Jaffray Woodriff (jaffray48@yahoo.com)
Date: Tue Jul 03 2001 - 10:50:32 MDT
Also, one has to remember just how much tailoring is
going into IBM's Blue Gene. Some are saying that the
job for which it is being constructed is the only one
that it will ever work on, and that job is only one of
a variety of approaches to analyzing gene and protein
data.
Therefore, from Blue Gene one cannot extrapolate the
computing cost per flop and compare that to the
computing cost of a Beowulf cluster without a huge
caveat. Beowulfs are more flexible by magnitudes.
[This is more in response to a somewhat earlier post.]
-Jaffray Woodriff
--- Dani Eder <danielravennest@yahoo.com> wrote:
> There are two trends that are leading to the Blue
> Gene type architecture. The first is the speed
> imbalance between the CPU and memory chips, which
> is getting to be 10:1 for the highest performance
> CPUs. Blue Gene type chips incorporate memory
> and CPU core on the same chip.
>
> If you look at current CPU chips, a good portion of
> the chip is already devoted to memory in the form of
> the L1 and L2 caches. So putting more of the main
> memory on the chip is an extension of this trend.
>
> The other component that is separate on today's
> chips
> is the controller that talks between the CPU and
> memory, and other peripherals.
>
> By merging the contoller onto the CPU chip, you now
> have a System-on-chip (SOC), as the usual buzzword.
>
> The other trend is to put more execution units onto
> a single chip. Currently you get 2-3 Floating Point
> Units per chip. The Blue Gene project aims to put
> 32 complete CPU cores on one chip. These will be
> simplified cores (probably 1 FPU, and without the
> boatload of specialized instructions current chips
> have).
>
> You still have a cluster of computers, but the
> clustering is on the chip and board level, rather
> than the box level of a Beowolf Cluster.
>
> What makes a Beowolf Cluster cheap is that they
> use cheap commodity parts. If an SOC chip becomes
> a commodity part (i.e. used for home PCs), then
> you can continue to make cheap clusters of them.
> Otherwise, a Blue Gene chip, which is targeted at
> 1 GHz x 32 cores, may cost $2000 per chip because
> it is not a high volume item. Meanwhile the
> high volume Intel and AMD chips of 2004 may be
> running at 5 GHz x 4 FPUs and only cost $200, which
> will still be a cost winner.
>
> Dani
>
>
>
> __________________________________________________
> Do You Yahoo!?
> Get personalized email addresses from Yahoo! Mail
> http://personal.mail.yahoo.com/
__________________________________________________
Do You Yahoo!?
Get personalized email addresses from Yahoo! Mail
http://personal.mail.yahoo.com/
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT