[Beowulf] Moores Law is dying

[Lawrence Stewart]

Steve Herborn wrote:
>  
> There are disciplines like EDA which generate large programs today.
> Computer generated, of course. Companies like Intel and AMD use these
> programs to design microprocessors. So yes, their architects are well aware
> of this issue.
>
>   
In particular, compiled simulators, which use very long straight line 
code to evaluate what the gates are doing in the correct order, every 
cycle.  The simple ones evaluate every gate every cycle, the complicated 
ones figure out what blocks of gates can be skipped.  These things 
really stress i-stream bandwidth.

I've not heard of multicore or cluster versions of these, but I'm not 
paying much attention to it.  Most design groups have so many test cases 
to run that many copies of the simulator is just as useful as one faster 
one.

Compiled sims generate very large text segments. At one instruction per 
gate, plus some overhead (loads and stores are "overhead" !) it is 
pretty easy to get to gigabyte text.   The counterpressure is the 
performance - with 4 GB/s istream bandwidth it takes one second per 
simulated cycle to run.

The point is that this class of large program does not have unmanageable 
complexity.

OT - I once was writing programs to generate programs with very large 
basic blocks, maybe 50K instructions.  It is a Big Mistake to try gcc 
-O3 on these, the optimizer goes into brain freeze for 15 minutes or so...

-Larry
Sector IX