On 5/24/06, [EMAIL PROTECTED] <[EMAIL PROTECTED]> wrote:


By changing inteface to better suit the implementation (as RISC
designers did) you reduce energy spent to develop something new. By
locking inteface you lock out a good part of you total energy.


That was the 80's.  Even now, RISC is somewhat passe in academic
circles, because of its own inherent drawbacks.  For one thing it has
code size issues that people would like to overcome.  But the bigger
problem is this:  When RISC processors were first developed, they
tailored the instruction set to the underlying architecture, which
made it very efficient.  But some of the design decisions that made
sense then (like branch delay slots) became a liability later when
they decided to extract more performance using superscalar and
superpipelining techniques.

Today, if you want to design a fast processor that's special purpose,
the favored approach seems to be VLIW, although that has problems
worse than RISC.  But when it comes to general purpose processors,
we're seeing that the translation from the ugly x86 ISA is becoming a
smaller and smaller part of the logic on the chip.  The more compact
instruction set requires smaller instruction caches for the same
performance, with few of the drawbacks, because the decoder can be
smart, combining and splitting x86 instructions as necessary.

Back in the 60's and 70's, the approach to designing processors was to
develop an instruction set and then design hardware around it.  The
result is things like the VAX which has completely orthogonal
addressing modes, which is a total waste since only 3 or 4 of them
actually see much use.  In the 80's, there was a "revolution" of
thought in that area, where instruction sets were designed around the
hardware.  Today, with superscalar and other sorts of multi-issue
super-pipelined designs, we're starting to return to a more abstract
view of instruction sets.

In my opinion, the holy grail right now would be an instruction set
that is based on profiling of thousands of existing applications,
taking the top few percent of instructions and filling in the rest for
completeness.  (Since the apps would be biased by the architectures
they were compiled for, this would have to be an iterative process.)
Then different underlying architectures could be designed around this
more abstract ISA.  It would certainly be easier to translate than
x86.

Now back to your regularly scheduled OGP discussion.  :)
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to