![](papers/GSEDA_note_EDPRev.gif)
DOWNLOAD PDF
"EDP
2009 – An Honest Look at Parallelism" |
EDP 2009 – An Honest Look at Parallelism
- EDP 2009 once again had an excellent technical program.
Four main themes were presented: Network on Chip (NoC),
the RTL ASIC Hand-off, Automation of Analog Design and
Parallel Computing. Thursday dinner was held at the Monterey
Yacht Club with Paul McClellan giving a thought provoking
talk on EDA’s future. Most of the presentations are available
at http://www.eda-stds.org/edps/
NoC – Will We Miss the Bus?
Will we Miss the Bus was a series of talks, and a panel,
investigating the move from Bus-oriented architectures
to Network on Chip. This was an especially good panel
technically covering a topic that has become very hot
recently. ASP-DAC this year showed the rapid switch from
Bus architectures to NoCs in the Asian market.
RTL ASIC Hand-Off
The good news on the RTL Hand-Off session was that Atrenta
gave an excellent presentation and following discussion.
Unfortunately they were the only EDA vendor presenting.
If you are interested in my presentation on the decline
of the Upper Mainstream companies, it is posted on the
EDP website (www.eda-stds.org/edps). This decline is what
is driving the move to the traditional ASIC hand-off,
which is now being shifted from the Gate Level net-list
to an RTL hand-off with the Silicon Virtual Prototype.
Automation of Analog Design
The fourth session, on Analog automation, was a bit of
a surprise. Cadence, Magma and Ciranova presented, three
of the four main players in the new set of analog tools
that were introduced recently. What impressed me the most
was that they all agreed that analog designers are the
most conservative designers in electronics. Analog automation
seems to still be at the stage digital automation was
when Daisy introduced features such as copy command and
rubber-banding.
All in all not an impressive view of what is becoming
the bottle-neck in SoC design.
An Honest Look at Parallelism
Parallel computing took up one session and the two other
Keynotes. One thing that is notable about the last year
is the openness of the talks on parallel computing. Almost
all of them have started out with the statement, “This
is really, really hard stuff.” That usually was followed
with a synopsis of the dead ends we have encountered in
programming multi-core/processor architectures. That’s
a far cry from what we were hearing just two or three
years ago.
Chris Malachowsky, one of the founders of NVIDIA, gave
a overview of what works (loosely coupled, multi-core
systems such as server farms and cell phones), special
applications that work, Embarrassingly Parallel programs
(video and other data-streaming applications) and what
doesn’t work (everything else so far). He then went on
to discuss CUDA, NVIDIA’s approach to programs that have
a large data-streaming content but do not fall within
the Embarrassingly Parallel problem category. These all
fit into their own Applications Specific category, yet
range from oil & gas exploration, medical imaging,
Financial applications and of course graphics. Another
interesting aspect of CUDA is that it is language agnostic.
It works with C and C++ now but will work with FORTRAN
and OpenCL. Other languages can be added if need be. This
is a recent trend with parallel computing. The actual
language is no longer a large part of the discussion.
On the topic of threads, NVIDIA’s approach seems to be
hard-wiring them as the safest way to handle them.
The “Are Threads Dead” session had advice such as “Threads
– a Necessary Evil”, you really need to know your algorithms,
don’t expect more that a 2x to 4x speed-up and threads
only scale to 4 processors (actually we’ve seen 16 processor
architectures that work). One recurring theme was the
need for parallel programming development tools. Right
now multi-core software developers are flying blind and,
in the heavy weather that is parallel computing, that
is a recipe for disaster.
Patrick Groeneveld, from Magma, gave the Friday Keynote.
It was an honest look at the difficulties of writing parallel
versions of EDA applications, specifically Logic Synthesis.
Patrick presented a translation of Amdahl’s Law in to
an electrical circuit, which really helped out the hardware-centric
audience. Personally it cleared up the emphasis being
placed on the Concurrent Memory by the Super Computer
community. That looks like one of the only ways to attack
the initial resistance in Patrick’s circuit diagram.
All in all EDP 2009 was once again an important workshop
for advancing the state-of-the-art in electronic design.
Gary Smith
|