( ESNUG 448 Item 9 ) -------------------------------------------- [11/11/05]
Subject: ( ESNUG 447 #4 ) Hey! RTL-to-Placed-Gates is Called "Blast Create"
> Though it did seem that RTL2PG was the right thing to do as technologies
> evolved, I see that even Synopsys has moved out of that doctrine by making
> IC Compiler a strictly Gates2PG tool, much like Blast Fusion or PKS.
>
> Does this mean RTL2PG is dead and Synopsys faltered in its effort to
> squeeze the real juice out of RTL2PG? Or were the other players smart to
> visualize that Gates2PG is clearly the way to implement better chips?
>
> - Jay Pragasam
> Open-Silicon, Inc. Sunnyvale, CA
From: Yasuaki Hagiwara <hagi=user domain=infrant spot gone>
Hey John,
I agree with Jay's observation but not his conclusion. In particular,
"...I see that even Synopsys has moved out of that doctrine ..."
I think Jay needs to look around for better solutions that really
implement RTL2PG.
It's been a while since I wrote in ESNUG. I wrote to you about my Magma
RTL-to-Volcano handoff back in Jan 2004. Boy, time flies! I just went back
and reread what I had said then. Quoting my previous post on this subject:
"Blast RTL's gain-based synthesis & gain report give you a good sense
of final timing. However, we still needed several iterations for
timing closure, and we couldn't meet our timing within a reasonable
time. The biggest reason might be because Blast RTL assumes ideal
placement in their fixed-timing methodology. However IO buffer and
Macro placement have certain restrictions in placement, thus nobody
can get the ideal placement. From this experience, we feel Blast RTL
should know some physical information (at least macro placement), and
reflect it into their fixed-timing methodology."
- from http://deepchip.com/items/dac03-26.html
Since then, I have done two more chips, both using Magma's Blast Create and
implemented by NEC America. Unlike my first chip with Blast RTL, in both
recent chips I went through placement in Blast Create and then handed off
the design to NEC. My RTL iterations were based on gain report analysis.
But once I handed off Volcano to NEC, there was no further change. No more
difficulty at NEC in closing timing either. Both chips were larger than
2 M gates, 300 MHz+ and were done with NEC's CB12 technology.
From a personal experience, I'd say RTL-to-Placed gates is alive and well,
and in real use. It's called Blast Create and it helps me achieve timing
closure and avoid surprises; and keeps my interaction with NEC design center
much more pleasant!
- Yasuaki Hagiwara
Infrant Technologies, Inc. Fremont CA
---- ---- ---- ---- ---- ---- ----
From: [ I Wear My Sunglasses At Night ]
John,
I saw Jay Pragasam's post about RTL-to-Placed Gates. I do not think that
RTL2PG is by any means dead, in fact that is the only way to reduce the
long iterations in the backend. Even as recently as 2 years ago, going
from a RTL to netlist was realtively fast, but Gate to PG (the backend
iteration) stretched to months to close timing juggling timing paths
between functional and DFT modes.
We are a small company, and we do not have large engineering teams. We
have a small team of frontend engineers in California and a small team of
physical designers overseas. We can't afford the time for iterations
between our two teams. Conceptually, we use RTL-to-Placed gates flow by
the frontend team that hands off the netlist, constraints and initial
placement to the back-end team. The rest of the implementation is carried
out overseas. Each RTL or constraint change goes through the same handoff.
After a draining experience in timing closure using our Gate2PG flow 15
months ago, we switched and currently use Magma's RTL-to-GDS flow. We
transfer Volcano at "Placed Gates" stage instead of exporting text files.
Our frontend team uses Blast Create and backend team uses Blast Fusion.
Besides the ease of handoff, the benefits are the common debug environment
and that the backend team does not screw up the placement chosen by the
RTL designer. Both teams have exactly the same view of the design and
they can refer to exactly the same paths, hierarchy, constraints, etc.
even though each is looking for problems in their respective domains
(RTL vs layout).
For a small, distributed team like ours, RTL-to-Placed gates is the
right approach.
For every new foundry and technology node, we still need to setup the
flow and do a trial, but we can initiate it as soon as RTL passes
simulation, hence not cutting into the backend time. I can't answer
Jay's exact question about why other vendors haven't delivered on the
RTL-to-Placed Gates promise. We use Magma Blast Create flow and we see
it working well for us.
Please keep me anonymous.
- [ I Wear My Sunglasses At Night ]
---- ---- ---- ---- ---- ---- ----
From: Yatin Trivedi <yatin=user domain=magma-da spot gone>
John,
I'm writing to you in reference to Jay Pragasam's question about the
value of RTL2PG. I think Blast Create users such as Aeluros, Broadcom,
Infrant, InSilica, Legend Silicon, NuCore, Renesas and others would all
attest to the fact that RTL2PG is the right approach.
I suspect the reason Synopsys is pushing a netlist2GDS approach has
nothing to do with their lack of commitment to an RTL2PG flow -- even
though both are driven by economic factors.
Synopsys has increased efforts in the netlist2gds space because its
aging point-tool flow is reportedly losing benchmarks to Blast Fusion.
IC Compiler is Synopsys' effort to convince the market that it can
compete against Magma -- as an old Indian idiom says, "If your cat can't
roar like a tiger then you need to change your cat." By using a new
product name and echoing industry rhetoric, they're trying to avoid
losing back-end business. Unfortunately, it takes a while for a cat to
become a tiger!
A possible reason the RTL2PG message hasn't worked for Synopsys is
because they wanted to avoid cannibalizing Design Compiler. But new
process technology and design complexity demand RTL2PG, so, they
recently introduced DC-XG -- which does in fact cannibalize DC because
it scraps wireload models and estimates global route on a trial
placement. Only thing is, it is an incomplete commitment; you still get
only the gates, but no placement! Better than the old stuff, but nowhere
close to what customers likely need.
For true RTL2PG flow, ask your readers to check out Blast Create.
- Yatin Trivedi
Magma DA Santa Clara, CA
Index
|
|
