( ESNUG 412 Item 12 ) ------------------------------------------- [05/22/03]
Subject: ( SNUG 03 #10 ) User Agrees Behavioral Design Must Make A Comeback
> So in 5 years, according to Gary's market trends report, this market will
> grow almost 15x. I called him on this and he replied:
>
> "The second generation behavioral synthesis tools failed as did the
> first generation. They had no language standard and limited
> control logic generation ability. Synopsys did the best job it
> could for this, but once you get into a datapath heavy design,
> Behavioral Compiler broke.
>
> Behavioral synthesis is going to come back for three reasons:
>
> 1.) We have a standard language in SystemC.
> 2.) We're in the middle of a crisis of design complexity and
> the power users are screaming for higher level tools.
> 3.) Once we get the planning & estimation tools fully debugged,
> we'll be able to check the QoR of the SystemC behavioral
> synthesizers.
>
> I expect this to be real by the end of 2004."
>
> - Gary Smith, Chief EDA Analyst at Dataquest
>
> In short, he thinks that SystemC will be why behavioral synthesis makes
> its comeback. Ask me and I'll tell you that Gary was obviously smoking
> some world class ganja when he wrote this! My gut says that Behavioral
> Compiler will die off because its functionality is already being absorbed
> into DC-Ultra. But hey, I could be wrong on this one.
>
> - from http://www.deepchip.com/items/snug03-10.html
From: John Mark Johnson <jjohnson51=person domain=cfl.rr watt bomb>
John,
Help -- I've been given the Double Kiss of Death! Agreeing with Gary Smith
is rare enough, but agreeing with Gary when the Almighty Cooley suspects him
of smoking weed -- my career is doomed!
Sure, programming in assembly language is fun, but not on somebody else's
schedule. Sure, you can write C programs to parse text, but why bother when
you have Perl? Designing 10M+ gates at the RTL level makes about as much
sense.
If I remember correctly, the only BC features that moved into DC-Ultra were:
1) optimize_registers,
2) transform_csa,
and 3) pipeline_design.
These have nothing to do with behavioral scheduling or resource lifetime
analysis.
Do you remember their classic complex multiplier demo? You simply code
Real <= Xr*Yr - Xi*Yi;
Imag <= Xr*Yi + Xi*Yr;
and let BC figure out how many multipliers/adders to synthesize, how to
schedule them, and how to implement the control FSM's and intermediate
storage. Pure RTL required separate code (and lots of it) for each
possible implementation; behavioral lets BC pick the best implementation
for the given constraints.
That example is small and well-defined. I don't expect behavioral EDA tools
to be mature enough for real 10M+ gate designs by the end of 2004, but if
they're not close, we're in trouble. Labor costs to design ASICs will go
through the roof, and life will suck as we work even more overtime.
Behavioral market growth of 15x by 2007? I doubt it, but I hope it's more.
You may call me crazy, too, but that's my final answer, Regis, and I'm
sticking to it. Yes, you could be wrong on this one, John...
- Mark Johnson
Logicmeister Consulting
|
|
