Re: For those who missed it -another PC milestone



Dirk Bruere at NeoPax wrote:

johannes wrote:

Dirk Bruere at NeoPax wrote:
johannes wrote:
FACE wrote:
From johannes <johs@xxxxxxxxxxxxxxxxxxxxxxxxxx>, in uk.politics.misc on
Sun, 29 Jun 2008 16:26:48 +0100 :

FACE wrote:
From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Fri, 27 Jun 2008 22:33:36 +0100 :

FACE wrote:
From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Fri, 27 Jun 2008 21:28:20 +0100 :

FACE wrote:
From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Fri, 27 Jun 2008 20:53:41 +0100 :

FACE wrote:
From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Fri, 27 Jun 2008 14:39:25 +0100 :

http://tinyurl.com/4y9ktt

AMD is shipping a graphics chip for use in PCs that can run at 1TFLOP ie
one million million floating point operations per second. And their
rival nVidia has released software to make this generally available as a
PC resource on their own chips. So PCs just jumped in effective power by
about a factor of 50.
From article lead:

"AMD Ships Teraflops Line Of Graphics Cards

AMD's new line of graphics cards is capable of a trillion calculations per
second, double the computing power of the company's previous generation of
high-end cards. "

In any event, this is a graphics card, not a processor. In fact, the
performance of the graphics card will be dependent on relayed information
emanating from central processor unit input.....

So where did your comment "So PCs just jumped in effective power by
about a factor of 50." come from?

(Or just put a check mark here __ to agree that you overstated all of
this........)
Because nVidia have also supplied a programming language extension, CUDA
that makes the parallel processing power available to applications other
than graphics.
CUDA was originally released February, 2007. It apparently only works with
Nvidia GPUs. However, your tinyurl article was about an AMD graphics card
release that was double the previous fastest AMD cards' processing power.
Nvidia is mentioned in passing in the article as a major AMD rival.

Please point to the reasoning for your saying "So PCs just jumped in
effective power by about a factor of 50."

Don't you think that your apology to S3 Graphics Co., Ltd., is in order?
Other companies eg AMD are following.
Not too big on periods or commas are you? Does your ISP limit you on
punctuation marks?

BTW, Please point to the reasoning for your concluding statement of "So PCs
just jumped in effective power by about a factor of 50."

(I use BTW because my ISP limits my whies and "b the wa" has two.)

FACE
Standard PC CPU offers around 10 GFLOPS of user programmable processing
Well that is wonderful, but please point to the reasoning for your original
concluding statement about this graphic card release that was: "So PCs just
jumped in effective power by about a factor of 50."

BTW, what is a gigaflop of user programmable processing? I always
understood kilo- mega- giga- tera- peta- flops to be expressions of
floating point operations per second, but not processing per se since a
"process" can be composed of many thousands of flops -- actually it is
layers deeper than that and the "flop" becomes a low-level basic building
block leading to a thing called "command"......... Things like RISC was
meant to reduce the flops necessary for processing supposed to simplify
this, OS's like Windows bloats it.
Just shows that you have no understanding of flops.
You are just another usenet fool.

Here is my original reply to dirky boy's idiotic claim. YOU answer it for
him, since he never could..........

~~~~

From Dirk Bruere at NeoPax <dirk.bruere@xxxxxxxxx>, in uk.politics.misc on
Fri, 27 Jun 2008 14:39:25 +0100 :

http://tinyurl.com/4y9ktt

AMD is shipping a graphics chip for use in PCs that can run at 1TFLOP ie
one million million floating point operations per second. And their
rival nVidia has released software to make this generally available as a
PC resource on their own chips. So PCs just jumped in effective power by
about a factor of 50.
From article lead:

"AMD Ships Teraflops Line Of Graphics Cards

AMD's new line of graphics cards is capable of a trillion calculations per
second, double the computing power of the company's previous generation of
high-end cards. "

In any event, this is a graphics card, not a processor. In fact, the
performance of the graphics card will be dependent on relayed information
emanating from central processor unit input.....

So where did your comment "So PCs just jumped in effective power by
about a factor of 50." come from?

(Or just put a check mark here __ to agree that you overstated all of
this........)

FACE

~~~~

Go for it einstein. Show anything I said to be wrong. But fucking ANSWER
it!

OK with pleasure. Start reading this article.
http://www.ddj.com/hpc-high-performance-computing/207200659

Exploiting GPU is a holy grail for those with high number crunching demands.
Obviously the GPU is very specialised for graphics purpose, the floating
points for the quoted flop rates are of are of low precision, but that's
sufficient for graphics. Real number crunching often demands 64 bit precision,
but that can be done with some reduction of the flop rate, but nevertheless
still very high performance to make it a tempting proposition.

The problems for solution must also have a high degree of dense parallelism,
just like the rendering jobs that are the usual GPU task, but that's often
a feature of many problems in physics anyway. It is not for the novice to
start programming this, so it's unlikely to have bearing on MS Office
users, not for quite some time anyway. Lack of portability has always been
an achilles heel of highly parallel systems; you have to be very determined
to get involved.
*** - now you're going to have to explain what "orders-of-magnitude"
means...

LOL. It goes from something like 10x for matrix problems and up to 415x for
MRI processing. Your quote of 50x was not out of order. Numbers may be in the
eye of the beholder, but still I'm sure hat there's something to be gained.
But as said, the programming is not for the faint hearted. Part of my PhD many
years ago was to use a shallow (1 bit only) massively parallel processor
for linear algebra problems. That was very laborious to get anything out of it,
and once programmed, it tended to be inflexible for changes. That's probably
also why you see the list of specialised GPU applications, each solving a
narrow range of problems. Nevertheless interesting if you have the time and
energy.

Not that old ICL thing was it? IIRC they came up with some 1 bit array
that was designed for database searching.

Close call. It was an AMT DAP, the AMT was a spin-off company from the ICL,
developing the ICL DAP. The computers were bigger (4096 processors) and got
faster clock, and the 1 bit processors were augmented by 8 bit floating
co-processors, the precision could be any multiple of 8 bit. Although
the speed was pedestrian compared to modern CPUs nowadays, the massively
parallel principles were transferable and highly relevant. Similar things
have been implemented on PC cards for signal processing.

Anyway, the thing I might be interested in is neural simulations. On
that new PFLOP machine I understand they are already running a full sim
of the Human visual cortex. Then there's grip processing. Imagine that
protein at home folding prog suddenly running 10 or 100 times faster. Do
you know whether grid programming could access the GPU that way at
present? Oh... and SETI

Not into grid processing myself, but I know a man who knows. Grid processing
must be a type of coarse grain processing where the parts doesn't need
communicating too often. The AMT DAP was the opposite end of the scale,
the communication was very slick; shifting from one row of processors to
the next took just one clock cycle (hence the 1 bit architecture). For
that reason it was also called context-addressable, like a CPU turned
around 90-degree with direct bit-access into the memory. Some people
speculated that this was closer to how our brains work, i.e. you recall
things by associations. The machines were used at the time by Reuters
news bureau for fast scanning of databases.


--
Dirk

http://www.transcendence.me.uk/ - Transcendence UK
Remote Viewing classes in London
.