|
On 24/10/12 19:32, Gordon Henderson
wrote:
On Wed, 24 Oct 2012, tom wrote:
On 24/10/12 15:49, Gordon Henderson wrote:
On Wed, 24 Oct 2012, tom wrote:
If you want a cool cluster try the
http://www.kickstarter.com/projects/adapteva/parallella-a-supercomputer-for-everyone?ref=home_popular
I want one!
Hm. I don't particularly agree with their pitch... For
example, over 20 years ago I was working with a company
building what we now know as clusters.. It's not a new
concept. Biggest one I helped build at the time:
http://unicorn.drogon.net/cs2.gif That had 256 compute boards
each with a dual processor sparc and 128MB of RAM...
https://projects.drogon.net/a-box-of-200-raspberry-pis/
10 years ago I worked with a company who developed a general
purpose plug-in pci/pcie card - 96 cores on it - under 10
watts a chip (they put 2 on a single plug-in card) see
http://www.clearspeed.com/
So the concept isn't new. Making it affordable and "open"
is...
The problem then and now still is: How do you program the
things? The question we were always asked went along the lines
of: I have this 30 year old FORTRAN program - make it go
faster with more data ...
Want to play with multi-cores on a chip today? Buy one of
these: http://en.wikipedia.org/wiki/Parallax_Propeller
e.g. http://www.xgamestation.com/view_product.php?id=51
Gordon
The parallela is programmable in C/C++ so no need to learn a new
language and last time I looked multicore stiff was pretty easy
on C++ and most of the apps I'd like to run on it are in C/C++
and easy to mod for multiprocessor running with a compile
switch.
Meikos CS2 20+ years ago was programmed in C/C++ as well as
FORTRAN - each node ran solaris... That didn't make it easier to
program. Their CS1 - Transputers and i860's also was programmed in
C/FORTRAN - again, even with the tools, people them reall did not
want to re-write their 30-year old "babys" that they'd nursed from
PhD through all their working lives - that why the likes of CRAY,
etc. got the advantage - compiler technology and processors highly
optimised to recognising and executing the array/vector
arithmetic..
Clearspeed don’t give a price so you have
to ask them - I tend not to bother with that approach - you
spend more on salesman time than you do on product.
You can't buy them anyway - they are now sort of dormant...
Same problem there though - no-one wanted to re-write their code
"make it go faster". So Clearspeed (and others) would write their
own librarys to inteface to their hardware that then provided the
standard stuff like blas libraries and so on.
It's always been the case that if you write natively you'll get
fantastic performance and always been the case (from what I've
seen first-hand) that no-one wants to write natively because it's
generally too hard.
The parallax is pretty slow - about 1/10th
the speed of one core of my cpu on this desktop.
You wanted parallel processing...
What I'm saying is that it's there now - and has been for decades.
It's nothing new at all.
Gordon
I didnt mean to give the impression that parallel processing was new
- the only thing new in computing are the new group of people who
think they've found or invented something new because they havent
bothered to research their field properly.
Tom te tom te tom
|
