( SNUG 04 Item 8 ) ----------------------------------------------- [08/11/04]
Subject: Astro, Apollo, Cadence, Magma P&R
SIAMESE TWINS: I laughed when I saw the Dataquest numbers for this market
segment. And it's not because Cadence had 45% marketshare in 2002 or that
Magma had 14%.
Dataquest FY 2002 Digital IC Physical Layout Market (in $ Millions)
Cadence (total) ############################## $176.5 (45%)
PKS, PKS-SE ######## $45.9 (12%)
NanoRoute, Silicon Ensemble ###################### $130.6 (33%)
Synopsys (total) ########################### $156.8 (40%)
PhysOpt ############# $78.9 (20%)
Astro, Apollo ############# $77.9 (20%)
Magma Blast Fusion ######### $56.9 (14%)
others # $3.9 (1%)
What caught my eye was that PhysOpt and Astro/Apollo had a virtual half and
half split of the Synopsys share. It's clear to me from this than even
Synopsys, Inc. themselves doesn't know the exact split because PhysOpt and
Astro/Apollo are so often sold together! They obviously said "Hey! Let's
just split it down the middle and skew $1 million over to PhysOpt because
our own R&D guys made it here. Synopsys pride! Synopsys rules! Yea!"
Another thing that is interesting in the stats is that plain olde vanilla
Cadence (33%), Synopsys (20%), Magma (14%) P&R tools make up 67% of this
market -- and yes, I do count Blast Fusion as a P&R tool despite how the
Magma marketing spin-doctors try to pitch it.
3.) Aart also claimed that Astro 2003.09 had gained 40% in capacity and
runtime sped up 20%. In your experience, is this real or marketing
hype? How does Astro fare vs. Cadence and Magma P&R? Which P&R are
you using now for production design?
Astro is pretty fast and can handle large designs. Our largest is
350 K cells and it is done flat in Astro. But timing closure still
needs some hand-holding. We use Astro for our production flow.
- Sunil Malkani of Broadcom
We've seen Astro got some boosts recently, but I'm not sure it's in
production. It's as much or more than 40% placement speed-up.
- Larry Tesdall of QLogic Corp.
Yep, Astro is getting there vis-a-vis capacity. Don't have any current
in-house comparisons with the other guys. A year ago Cadence couldn't
do clock trees worth a damn, but I hear a lot has changed since then.
- [ An Anon Engineer ]
Today we are using Astro. Other design teams here are using Magma in
our building. About 18 months ago Magma physical placement was ahead
of Astro. Now with the PhysOpt CGPL algorithms in Astro the results
are very close.
We are pushing 700K flat instances in 0.13 u but we are limited more by
runtime than capacity. Our lastest Opterons are enabling very fast run
times with increased gate counts. I would say Astro 2003.09 is faster
than previous versions.
- [ An Anon Engineer ]
Astro is the ultimate physical synthesis and timing closure engine.
It has sped up a lot over the past year. All other tools are
no comparison.
- Haiming Jin of Intel
We use Astro. Since we are reasonably happy with Astro, we haven't done
any real evaluation of Cadence or Magma. We did see some runtime
improvement in the 2003.09 Astro release, but it is hard to say if it
was as much as 20%. My gut feeling would be something like 10-20%.
Can't comment on the Astro capacity question, since we run a strict
hierarcically flow with block sizes well below the Astro capacity limit.
- [ An Anon Engineer ]
We use Astro.
- [ An Anon Engineer ]
Using Astro
- Erica Wickstrom of PMC-Sierra
I don't have experience in comparing (apples to apples) two different
version Astro binaries on the same block using same machines. But I
think, the runtime and capacity of Astro is very good compared to
Apollo. That is a significant improvement. With that, Astro is
comparable to Blast Fusion. I did an eval on a block (130 K placeable
instances) which had difficulty in meeting timing with Apollo. Both
Astro and Blast Fusion chewed it easily and quality of results were
almost equal. We currently use Synopsys tools for production and are
looking at Magma as an alternative due to cost reasons.
- Santhosh Pillai of Parama Networks
Dunno about Astro capacity as it has handled everything we've thrown at
it even before 2003.09, but the speed increase *certainly* is not 20%!!!
Major load of crap. As of 2003.09-SP1-2 clock tree synthesis *finally*
works, but overall I still consider Astro slow and getting good CTS
results is something of a black art which it shouldn't have to be.
They have serious stability issues on the IA64 platform, too.
- [ An Anon Engineer ]
We do not do place and route. We hand off at the netlist level.
- John Blessing of Harris RF
We have done a large design flat with both Cadence Nano Encounter and
Astro. We were ready about 3 months earlier on the Cadence flow
(really). The run times for Astro are immensely longer than for
Nano Encounter and there are a lot less bugs in the tool. We had a
very experienced Astro engineer and a lot of work done by Synopsys as
well to get through this. To be fair, Astro is known to have problems
with larger flat designs, but the myriad bugs and quirks are not really
excusable for such a seasoned tool.
- Louis Morales of Innotech Systems
Unknown capacity/runtime, though Astro 2003.09 addresses various
functional issues as well. No data about Magma. Most of our production
routes are done with Warp, a few in NanoRoute, and only a few in Astro
at the moment. Expect that to swap in the next 6 to 9 months with Warp
getting dropped, majority in Astro, and some in NanoRoute. We see Astro
as the more complete router, but NanoRoute as faster and gaining ground
on completness.
- [ An Anon Engineer ]
I cannot comment on this, since we outsource routing tasks to external
business partners. However they use Astro for those designs.
- Marcello Vena of Xignal Technologies AG
For the record, we're using Astro for P&R, Calibre for DRC/LVS,
Star-RCXT for extraction and PrimeTime for static timing analysis.
- [ An Anon Engineer ]
Astro is getting better. But still slower than Magma.
- Massimo Scipioni of STmicroelectronics
We use Cadence Silicon Ensemble now. We are planning to transition to
a different tool this year, but we haven't decided yet. My gut says
Encounter, but we must evaluate some more.
- [ An Anon Engineer ]
Our P&R guys use Astro.
- Juan Carlos Diaz of Agere
Comparison in the last 2 years of Magma to Astro showed Magma in the
lead. Our recent evidence indicates Astro is catching up. Currently
using Magma.
- [ An Anon Engineer ]
Magma has a reasonable tool but lousy support and even worst attitude
to customers. Life is a wheel. One day you're up, another day down...
- [ An Anon Engineer ]
We use Magma Blast Fusion. We are very happy with it. The best thing
is the single binary/unified data base architecture. It makes scripting
a lot easier. The support has been very good and the AE's are very
knowledgeable.
- Joe Dao of Aeluros, Inc.
Used Magma on last chip. Toshiba also uses Cadence. It was a complete
mess in terms of quality results. Problems with timing closure. Maybe
could be expected. Magma has no full extraction and no real ECO flow
like in Cadence SE. Blocks with wide and deep combinatorial paths that
were routing intense were a significant problem.
Post-Magma power analysis came in 6X compared to what was initially
anticipated. (Our initial chip power estimate was analyzed with
the DC netlist in PrimePower.) Gain-based synthesis/re-optimization
ripped up all the logic structuring done by Synopsys and yielded a very
bloated design with a huge increase in power. Magma even looked at this
themselves in parallel with AEs working from Toshiba. They saw Magma
was eliminating good structures that DC had made, and it was not using
multiple-input gates to its advantage. Eventually we got Toshiba to
P&R our Synopsys DC netlist as-is and this relieved most of the power
bloat issue; although it was still increased as Magma "optimized and
buffered". It still was significantly off at 50% higher power. We
ran out of time to keep debugging Magma's approach further; it was not
clear if the placement was done well.
Magma's response to our using the same floorplan, same constraints, in
a Synopsys DC generated netlist which had significantly better power
than the Magma's netlist was "the customer power expectations were off".
As you can imagine, Magma lost a lot of credibility with all involved.
- [ An Anon Engineer ]
We've use Silicon Ensemble in the past, and that is what we would use
if we were renting or buying. Our design services partner uses First
Encounter and related Cadence tools.
- Brett Warneke of Dust Networks
I think NanoRoute is faster, but we're still using Astro.
- [ An Anon Engineer ]
Our P&R flow is based on Cadence PKS and CCAR (Cadence Chip Assembly
Router), but this is for special memory technologies, not standard ASIC
technology. We may look at Astro in the future, too.
- [ An Anon Engineer ]
We're using Cadence SoC Encounter w/ Amoeba place and Warp route. I
should mention that I've had horrible experiences with all manner of
Cadence clock tree insertion tools and engines. They do INCREDULOUS
things -- CTgen, CT-PKS, and Encounter's CTStch. Though they claim
efficient buffering, on a VERY SMALL 32 leaf net of mine, I've seen
CT-PKS use either a single BUF7 or a 4 level tree of 32+ buffers and
inverters depending on the phase of the moon! Same scripts and design!
CTStch has gone and inverted a single clock phase in a clock group
on me, in addition to overbuffing it!
- Gord Allan of Carleton University (Canada)
|
|
