( SNUG 04 Item 4 ) ----------------------------------------------- [08/11/04]
Subject: DC 2003.06 vs. DC 2003.12 vs. DC 2004.06 (Nighthawk2)
NO PINOCCHIO HERE: It's one of the oldest tricks in the marketing books.
Take any old product, slap "New!" and "Improved!" on it, and watch the sales
go up. Or at least that's what I thought when Aart made those claims about
DC being faster, better, etc. Turns out he was telling the truth here.
1.) In his SNUG keynote address, Aart claimed Design Compiler 2003.12
runtime had sped up 50%, areas were 5% smaller, and DC gained 20% in
capacity. In your experience as a hands-on user, is Aart lying or
telling the truth here? How does Design Compiler stack up against
its rivals Cadence Get2chip, Ambit, Synplicity ASIC, Magma Blast
Create, Incentia in your eyes? Which synthesis are you using now?
Yes, I have seen significant runtime improvment with DC Expert 2003.12
running on Linux. I don't know for DC Ultra. This is the first time
I've ever see runtime going down with DC. I got a 25% reduction in
certain cases. The improvment shows better on large top-down synthesis
rins. (725 Kgates/ 40 K FFs) The runtime for that same block was
further reduced using 2004.06 (Nighthawk2) by 47%. This was almost
unbelievable. Just to give you an idea about the runtimes:
DC 2003.06 : 11 hours
DC 2003.12 : 8 hours 30 min
DC 2004.06 (Nighthawk2) : 4 hours 30 min
In some other cases (mostly bottom compilation) there was little or no
changes. Regarding the memory footprint, it also improved though I
haven't got a precise value to report. I believe that some of the mem
improvements were rather bug fixes for MCE that did not release the
memory that it no longer needed.
With regards to area, I haven't seen any change neither with DC 2003.12
nor with 2004.06.
- Pierre Ragon of Lucent
I've noticed a speed up in term of runtime, but not as much as Aart
claimed. In my cases, (customized high-speed DSP circuits) DC was
about 20-30% faster. The area improvement figures were about 3-4%.
- Marcello Vena of Xignal Technologies AG
I've been using DC all along. There have been speed gains, but the
increase is probably not to 50%. Not sure on the area, although it
could be better if DC is better at targeting custom-designed mega
cells from the library.
- Haiming Jin of Intel
I think that DC 2003.12 is one of the worst versions I have seen from
Synopsys in a while. DC 2004.06-beta (Nighthawk2) is one of the best
versions I have seen from Synopsys.
DC 2003.12 does give smaller area and runs faster but deteriorates
timing as compared with its predecessor 2003.06.
In the chip we are developing I have 8 blocks that are basically
similar with the exception of minor decoding. Although 2003.06-SP1-3
was meeting timing, there was an area penalty that we could not afford.
We needed area relief, so I turned to 2003.12. I ran 4 blocks
interactively using 2003.12-SP1-1. I was following the optimization
step by step and I saw results all over the map. QoR results were
completely different -- which did not make any sense. TNS, WNS,
failing paths were all over the map!!!!!
By contrast running using Nighthawk2 (2004.06), the optimization steps
were in lock steps and the QoR were very similar with minor area
differences ,which is what I expected. This was run again with a
2004.06-beta and the results are better and tightly correlated as
expected. I got improved area without sacrificing timing and had the
luxury of faster run times.
I applaud Synopsys for the Nighthawk2 release even before it is
officially unveiled, but I give the the thumbs down for their
DC 2003.12 release.
- Sam Bishai of Cisco Systems
DC is getting faster. I think that these claims are valid, if not a
release premature. The newest DC builds are giving us about these
numbers.
- Andrew Bell of PMC-Sierra, Inc.
Aart's improvement claims for DC are major load of B.S.
Of course, like most people, we have to use Synopsys tools due to
inertia. :( We are considering alternatives on our next chip, though.
Bet the farm on it.
- [ An Anon Engineer ]
I am desperately looking forward using DC 2004.06 (Nighthawk2), and
get the area savings. I could not try 2003.12 because the area savings
are meant to happen if you use XG-mode, but this mode is in limited
availability, and we don't have license keys for it in my company.
Presently I am using DC 2003.06-SP2.
- Juan Carlos Diaz of Agere
Now I am using DC 2004.06 (Nighthawk2), and I think the performance is
improved, but I think it's not so much like Aart said.
- [ An Anon Engineer ]
I am seeing large runtime improvements and memory capacity reductions
using a special pre-released version of Design Compiler code named
Nighthawk2.
- Philip Jackson of Conexant (ESNUG 424 #1)
Amazingly we were able to fit our design on a 32 bit Linux box with the
beta allowing us to take advantage of significantly faster processor
speeds. The runtime and area improvements were just awesome!!!
dc_version: DC 2003.12 Nighthawk DC 2004.06(B)
Machine: 64-bit Sun 32-bit Linux
900 Mhz 2.2 Ghz
Memory: 8 G 2.3 G 3.5X smaller
Compile Time: 75 hrs 24.6 hrs 3X faster
Area: 7403058 5683879 23% smaller
TNS: -5264.14 ns -1770.92 ns 3X lesser
I should mention that we did find one issue that was increasing our TNS
by 5X but Synopsys was able to quickly identify and fix the issue and
deliver me a new beta executable.
- Johnny Zhang of Cisco (ESNUG 429 #2)
Well, this is a joke. Areas are 5% smaller compared to DC 2002.05 but
that's because DC 2002.05* series had issues where the area increased
5% to 10% over the DC 2001.08* versions. It took a while for Synopsys
to admit to this and find the issue, but eventually they did. They are
about back to where they were nearly 3 years ago in terms of area.
- [ An Anon Engineer ]
We are using Synopsys DC, but have not used 2003.12 yet, because of
restriction in the Synopsys version specified by ASIC vendor.
I believe the Synopsys's claim for performance improvements are 'very
best effort', but I hope 2003.12 works fine for my designs.
- [ An Anon Engineer ]
Generally speaking, we have seen the various speed and area DC claims to
be true, more or less. Aart, I think, is quoting 4th quartile numbers.
Everybody is happy right now as we deployed a Linux farm, so there's an
automatic 2-4x gain right there. With Opterons now showing up as well,
getting SNPS to sqeeze a little more is largely irrelevant. CPU clock
speeds are going to increase faster than tool efficiency. We are
DC/PhysOpt based, with a little Get2chips use, too.
- [ An Anon Engineer ]
Without making substantial changes to our DC scripts, we were able to
improve timing and speed up synthesis by 8-12X with the Nighthawk2
release compared to the prior version of DC. For the first time, we
are also able synthesize our monster 90 nm chip top-down.
We're eagerly awaiting the official Opteron release of Nighthawk2 to
use its improved performance and capability for sign-off.
- Chandramouli Mahadevan of Texas Instruments (ESNUG 429 #2)
We're using DC and PhysOpt exclusively for synthesis. However, we're
sticking with 2003.6 since we've found numerous bugs and issues in
2003.12. It may be faster, but it's also highly unreliable.
- [ An Anon Engineer ]
Aart is in the truth-ballpark this time. We see dramatic speed-up in
dc_shell 2003.12 with at least SP1 and maybe SP2??.
- [ An Anon Engineer ]
We used Design Compiler 2003.09 on our last ASIC taped out in Q1 04.
- John Blessing of Harris RF
"Aart claimed Design Compiler 2003.12..." Did he mention that the
version is a so buggy even the Verilog out has problems?
- [ An Anon Engineer ]
We are only using DC in our team. I do not believe any area or timing
estimates from wireload based synthesis until the full physical flow
has implemented the design in 0.13 u and 90 nm. We have observed
faster runtimes with DC 2003.12, but results were slightly worse.
- [ An Anon Engineer ]
We only are using Design Compiler at this point. Have seen a lot of
speed improvement compared to say 2 years ago; not sure what the 50%
comment is referenced to. Magma synthesis/optimzation had a serious
problem with deep combo logic functions with large routing needs.
- [ An Anon Engineer ]
I experienced a 35% runtime improvement on various designs using
DC 2003.12. Did not notice much as far as area was concerned, and the
swap space used by 2003.12 was similar to 2003.06 (no XG mode). So
Aart is stretching the truth slightly... Note that a 1/3 improvement
is nothing to sneer at, though.
However, when using DC 2003.12 on a large design, I found out that the
timing violations were worse, so I didn't switch: I'm still using
20003.06-SP1-3 at the moment, have given up on DC 2003.12 for the
foreseeable future. Awaiting 2004.06 (Nighthawk2).
- [ An Anon Engineer ]
|
|
