Re: For those who missed it -another PC milestone

From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Sun, 29 Jun 2008 20:00:22 +0100 :

Dirk Bruere at NeoPax wrote:
FACE wrote:
I asked a simple question and you could not back up your dumbass
claim that " So PCs just jumped in effective power by about a factor
of 50."
Because it is so obvious.
The processing power of the GPU is now available for general purpose
use as a vector processor delivering single and double precision
floating point operations. Which are the hardest to do. The offering
increases the computational power available on a PC by around 50x.

Above, do you mean "single and double precision" or "single or double
precision"?

I still think you made a massive overstatement. You are talking about
a discrete component.

See Amdahl's Law.

Calling it "obvious" is an exercise in arrogance.

OK - I'll assume you're not playing stupid to wind me up.
http://en.wikipedia.org/wiki/Floating_point

Releasing the power in GPUs to the user, rather than limiting it to
graphics calculations, means that almost any simulation will run at what
was once supercomputer speeds. No doubt you read the Dr Dobbs article on
relative performance.

Almost all realworld simulations need the precision arithmetic of
floating point, which is the most complex instruction executed by a
general purpose CPU.

In fact, one researcher found his program running faster on his PC with
CUDA than it did on his university supercomputer. It has effectively
reduced the price of TFLOP computation from around $10K to a couple of
hundred.

You should also bear in mind that as little as 10 years ago that
performance would have been world class for supercomputers, and would
have cost millions.

Here's one from 2004
http://www.geek.com/dalco-to-build-2-tflop-supercomputer-to-model-race-cars/


India in 2003
http://www.smh.com.au/articles/2003/01/01/1041196677610.html

In 1993 it would have been state-of-the-art no expense spared.


A mere 3 years ago:
http://news.taborcommunications.com/msgget.jsp?mid=406641&xsl=story.xsl

"
The Leading Source for Global News and Information Covering the
Ecosystem of High Productivity Computing / June 24, 2005

...
As predicted several years ago by the research team behind the TOP500
listing, only systems exceeding the 1 TFlop/s mark on the Linpack were
qualified to enter the list this time. The system in No. 500 spot
reached 1.166 TFlop/s... "


I started out with a simple question, not to wind you up.

Just a by the way point, I note that you say TFLOP, PFLOP, etc. You do
realize that the "S" is an integral part of the acronyms that introduces the
time element and not a pluralisation. Usage of it is almost a personal
choice though........

I hope you read up on Amdahl's Law.

FACE

.

Relevant Pages

  • Re: For those who missed it -another PC milestone
    ... " So PCs just jumped in effective power by about a factor of 50." ... The processing power of the GPU is now available for general purpose use as a vector processor delivering single and double precision floating point operations. ...
    (uk.politics.misc)
  • Re: For those who missed it -another PC milestone
    ... The processing power of the GPU is now available for general purpose use ... do you mean "single and double precision" or "single or double ...
    (uk.politics.misc)
  • Re: So far OT its over the horizon: Anyone into 3D modelling?
    ... > screws, 2.5 metre ball screws have a certain minimum cost, even if you ... > cut out ALL the middle men you'll be looking at IRO 500 quid per axis ... you could use precision threaded bar and a couple of nuts to ... > eliminate backlash, but it's gonna wear quickly, take more power to ...
    (uk.rec.motorcycles)
  • Re: So far OT its over the horizon: Anyone into 3D modelling?
    ... screws, 2.5 metre ball screws have a certain minimum cost, even if you ... cut out ALL the middle men you'll be looking at IRO 500 quid per axis ... you could use precision threaded bar and a couple of nuts to ... eliminate backlash, but it's gonna wear quickly, take more power to ...
    (uk.rec.motorcycles)
  • Re: My Homework: how to calculate natural log, e?
    ... > So GT's claim of 1ulp precision are demonstrably false. ... with pwr equal to a power of two) throws away a lot ... couple of hundred ULPs when the input is around 700. ... no deficiencies and the other way is to make it so complicated ...
    (comp.lang.fortran)
Loading