I heard about Cuda and GPU acceleration for HPC applications before
but this time I feel (like the author Kevin Morris) that this solution
is getting traction. I know some guys that three years ago tried to
use a Nvidia GPU to do FFTs and had to work at the OpenGL level. Not
very friendly. Now with Cuda getting more mature (and Telsa getting 64
bits floating point), it looks like it's becoming a nice alternative
to FPGAs.
Patrick
and also this
http://www.embedded.com/products/softwaretools/208700454
["SAN JOSE, Calif. — Apple Inc. has submitted the Open Compute Language
(OpenCL) to an ad hoc industry group that aims to define a programming
environment for applications running across both x86 and graphics chips.
The move is one of a growing number of efforts to extend the ubiquitous
C language for increasing parallelism in multicore processors.]
Next, will we see boards with GPU and FPGA ?
-jg
Speaking of FPGA alternatives, this recently caught my eye. Don't know
much about it, but it sure looks cool:
http://www.tilera.com/products/processors.php
-Jeff
Jeff,
Where have I seen that before?
Ah yes, http://en.wikipedia.org/wiki/Transputer
Syms.
>Where have I seen that before?
>Ah yes, http://en.wikipedia.org/wiki/Transputer
And now, for bonus points, put an [X] against any
of the following reasons if you think it helps to
explain why the Transputer was a market flop...
[ ] it was at least two decades ahead of its time
[ ] it was about one decade ahead of the technology
needed to make it powerful enough to be competitive
[ ] it was based on robust, sound academic theory
[ ] it had a clean, effective, readable programming
language that was not C
[ ] it was a British design
[ ] the software community was even more clueless about
how to make use of multiple scalable compute resources
in the late 1970s than it is today
[ ] when presented with a model that permits parallel
processing to be spectacularly efficient, the
software community retreats into its standard cosy
set of assumptions that serial execution is sure
to be faster and more efficient, and therefore
Concurrent Is Bad, mainly because it isn't C
For maximum credit, write a 10,000 word dissertation
explaining why the social dynamics of the software community
will ensure the early death of anything that looks like
a general-purpose, tightly-coupled network of compute nodes.
Not that I'm bitter, or anything like that :-)
--
Jonathan Bromley, Consultant
DOULOS - Developing Design Know-how
VHDL * Verilog * SystemC * e * Perl * Tcl/Tk * Project Services
Doulos Ltd., 22 Market Place, Ringwood, BH24 1AW, UK
jonathan...@MYCOMPANY.com
http://www.MYCOMPANY.com
The contents of this message may contain personal views which
are not the views of Doulos Ltd., unless specifically stated.
>On Thu, 19 Jun 2008 10:37:14 +0100, "Symon" wrote:
>
>>Where have I seen that before?
>>Ah yes, http://en.wikipedia.org/wiki/Transputer
>
>And now, for bonus points, put an [X] against any
>of the following reasons if you think it helps to
>explain why the Transputer was a market flop...
>
>[ ] it was at least two decades ahead of its time
>[ ] it was about one decade ahead of the technology
> needed to make it powerful enough to be competitive
>[X] it was based on robust, sound academic theory
>[X] it had a clean, effective, readable programming
> language that was not C
>[X] it was a British design
worse than that: it was a Government funded design under a government
that didn't fund technology...
>[ ] the software community was even more clueless about
> how to make use of multiple scalable compute resources
> in the late 1970s than it is today
Definitely not this one.
as well as O***m, they came up with Ada, where you can:
create a task type;
create an array of tasks of that type;
loop over the array starting the tasks;
loop over the array collecting the results;
in pretty much no more code than the above.
Nowadays, C++ has given the SW community ONE positive thing:
the old argument that Ada is too large and complex to be usable no
longer holds water...
But count the hand-wringing woe-is-me stories about the difficulty of
parallelism and the massive new collaborations to solve the problem on
one hand, and the mentions of Ada (or O***m) on the other...
>[X] when presented with a model that permits parallel
> processing to be spectacularly efficient, the
> software community retreats into its standard cosy
> set of assumptions that serial execution is sure
> to be faster and more efficient, and therefore
> Concurrent Is Bad, mainly because it isn't C
>
>For maximum credit, write a 10,000 word dissertation
>explaining why the social dynamics of the software community
>will ensure the early death of anything that looks like
>a general-purpose, tightly-coupled network of compute nodes.
>
>Not that I'm bitter, or anything like that :-)
The funny thing about the transputer (from the VERY little I did on it
at the time; paper only since I couldn't afford one as a hobbyist) is
that there was a time window when it appeared to be the fastest single
processor available, (ignoring 3-chip solutions) even without its hooks
for parallelism... but that got overshadowed by the parallelism.
If it had had an MMU, it could have made a nice Unix workstation
(or single-chip Lilith, or...)
- Brian
What is missing is an embedded solution. To my knowledge, there is no
compact embedded system using the latest Nvidia GPUs. We can't fit a
PC (even a smallish one) inside our instrument. Even if we could, heat
dissipation would be a problem. So I guess that FPGAs will still reign
for a while in high performance embedded applications.
Patrick
Availability? Price?
Dave
Nvidia. Google.
Yes. Cheap.
MikeWhat?
Really? I got _no_ hits on google for the middle speed-range part
number. Where did you find it available from? Price?
Dave
The C870 is $1300, Nvidia direct or second sourced. The C1060 is slated for
fall release.
Try te...@nsc-nvidia.com for more directed info. Would you like the phone
number also? Here:
For Information on How to Purchase Call 408.392.4120
Second cousin.
(Full initials. Also a pronunciation aid. You get my age and you'll find the
dimunitive inappropriate also.)
Ahh. The old switcheroo. Jeff wrote about the Tilera TILE64 processor
and I asked about its availability and price. You decided to give
availability and price (sort of) on Nvidia's Tesla GPU card.
"Dave" <da...@comteck.com> wrote in message
news:f008$485ae84d$40b83d5e$25...@COMTECK.COM...
I thought your second cousin might explain what Nvidia had to do with
Tilera. I see another branch of the thread has cleared that up! :-)
Cheers, Syms.
Explains why I don't mind top-posting as much as others. :) Didn't notice
the change in topic.
I also didn't find info on the Tilera. Seriously thinking about ordering the
C870 here, though.
The Nvidia store lists the C870 as "Available for Backorder" so you may
have a wait to get one. The D870 is in stock and is only $5K. ;-)
Dave
One of the things that strikes me about the Transputer is that it
was parallel hardware designed hand-in-hand with parallel software.
Not C/C++
It seems odd that there is so much convergence happening between a very
complex highly sequential CPUs and functionally simple, highly parallel
FPGA type devices, with lots of innovative hardware flying about.
All this convergence is happening in hardware, but for it to really work
don't the software environments need to do the same...
---
cds
The new XMOS devices (out of the same stable as the transputer) are
intended to bridge the gap between FPGAs and processors:
Leon