[Fsf-friends] Re: Desktop supercomputers
Fri Aug 5 07:11:50 IST 2005
,----[ Ramanraj K <email@example.com> ]
| Suppose I want my application to run 40 to 200 queries on a db, all
| concurrently, to take a simple decision, brain like parallel
| processing schemes/devices seems useful. The recent 64 bit
| processors seem to come with capacities for very large RAM, etc, and
| I am just curious if brain emulators are possible now.
| "Supercomputers" seem to be more focused on increasing clock speeds,
| but it may be worthwhile to also focus on increasing the number of
| threads/processors we could use to the range of million millions
| trading off with clock rate.
| I am not sure about technical feasibility etc., but this is one of
| the "unsolved" areas of cs, that may be of interest to researchers
| Ref: http://www.electricminds.org/ussclueless/essays/futurecs.htm
Supercomputer manufacturers are *not* focussed on increasing clock
speed, instead they try to maximize FLOPs (Floating Point Operations)
/ second for a given budget.
If you look at Top500.org, they consider the following factors to
grade a supercomputer.
Rmax Maximal LINPACK performance achieved
Rpeak Theoretical peak performance
Nmax Problem size for achieving Rmax
N1/2 Problem size for achieving half of Rmax
To calculate RPEAK use the following formula:
Rpeak ~= Number of Processors * Clock Speed * FLOPS/Cycle (or Number of
4096 x 1.4GHz x 4 ~= 23 TFlops.
CDC6440 has 4FPUs
Rpeak/Rmax ~= Efficiency.
Top500.org's way of ranking a supercomputer is not adequate. They fail
to consider some of the important factors like efficiency,
price/performance ratio, manageability and TCO. Actual design of
supercomputer is very much dependent on the application needs. Best
way to benchmark is to use the real world applications and measure
performance. For example applications that are embarrassingly parallel
doesn't really need an expensive high performance low latency
interconnect such as Infiniband. GigE will do fine.
On the other side, if you are taking about processor clock speed..
recently processor manufacturers have realized it makes more sense to
maximize the number of cores in a processor than frequency. Dual core
procs are already out there it the market. Quad and Octa are in
development. Your argument is valid and Industry is certainly moving
in that direction.
Speaking about supercomputers replacing human brain, we really don't
know enough about human brain. We don't even have a good acceptable
definition for Artificial Intelligence. Its a long way to go to for
computers to achieve human like intelligence (30-50 years may be).
But with the capabilities of today's computing hardware, we can sure
replace lot of routine human tasks smartly. Mostly in the field of
automation. Software for AI is vastly behind. I very much agree with
you on this subject. Software engineers are more concerned about
parsing XML stuff or writing web applications, instead of solving the
real problem of making machines intelligent and they perform the tasks
themselves instead of we programming them every time.
The GNU project has entered the field of AI too. Look at
GPG Key ID: 0x62E15A31
The GNU Operating System [http://www.gnu.org]
More information about the Fsf-friends