( ESNUG 470 Item 1 ) -------------------------------------------- [10/31/07]
Subject: ( ESNUG 469 #1 ) Synopsys or Magma won Broadcom -- NOT Atoptech!
> I've heard tell of a big place-and-route benchmark that one of the larger
> Left Coast companies ran. I heard it turned out:
>
> 1. Atoptech
> 2. Cadence
> 3. Mentor Sierra
> 4. Magma
> 5. Synopsys
>
> I imagine these results would cause much gnashing of teeth at both Magma
> and Synopsys. Who are these Atoptech guys? Is this just startup luck, or
> are they for real? What was the design? What was the criteria for eval?
>
> - Steve Golson
> Trilobyte Systems Carlisle, MA
From: Jonathan Bahl <jonathan.bahl=user domain=cotconsulting not calm>
Hi John,
You have misleading benchmark data there. Those are the results only if you
look at RUNTIME! (BTW, I heard it was a block benchmark.) Why Synopsys
is so dog slow is that it doesn't support multi-threading during opto, while
Atoptech does. Heard Cadence was fast because it spit out a low QOR result.
Magma is a slow tool, too, once you bog it down with OCV and SI.
What I heard was that if you looked the TIMING (QOR) the results were:
1. Mentor Sierra
2. Atoptech
3. Magma
4. Synopsys
5. Cadence
I also heard the margins of victory were not much between the top 4 tools;
it was only Cadence that lagged on timing.
However, for 65/45 nm mainstream designs, runtime and timing QoR should NOT
be your biggest concern about a P&R flow.
For a LOW POWER design, on the frontend Cadence is definitely ahead because
all their tools agree on CPF; I don't think anyone can hold a candle to them
on that. Problem is this CPF unity hasn't yet been proven on their backend
flow for any flat multi-voltage designs that I've heard of. For low power
P&R, I'd guess:
1. Synopsys
2. Magma
3. Cadence
4. Mentor Sierra
5. Atoptech
I chose Synopsys #1 here because they've done flat multi-voltage designs for
years with at least 30-40 solid tape-outs. I've seen Magma come out ahead
of Cadence for a number of years on lots of benchmarks, so my gut says they
should be at #2 here. The tapeouts I've seen for Cadence, their approach
to low power was stiched together the top level (i.e. no flat, no multi-
voltage support) so they should be #3. I easily put Mentor and Atoptech at
the bottom on power because they're just too green; tools lack maturity.
For a TEST INTEGRATION benchmark, I'd say that the ranking would most likely
break out to:
1. Synopsys
2. Cadence
3. Mentor Sierra
4. Magma
5. Atoptech
Synopsys would be #1 because they own DFT MAX, which is integrated into
DC-Topo. The only test operation done in PD is reordering. If you insert
test logic on a placed design, it will increase your area 1% to 2% (which
sounds like nothing, but its jambed all in one location preterbing your
design big time!) The DC-Topo hook with ICC bypasses this mess. That's
why a PD guy only has to reorder using ICC. Cadence is #2 because they do
support a good re-odering flow plus they own the DEF format. Owning DEF
means they can use DFT MAX or TestKompress easily. From a PD view, Mentor
owns TestKompress but Sierra is still a raw tool, hence they're #3. Magma
has no compression and I don't know if they can even support DFT MAX nor
TestKompress. Atoptech a newbie here.
For AREA, I'd be afriad to do a ranking, because I know ICC has a really
good placer for timing and congestion, but (as I said before) it's dog slow
and will choose bigger cells to fix timing. It's tough to figure out area
ranking because it's a forced conflict with timing; better timing, worst
area; better area, worst timing. I haven't heard much about area for Mentor
nor Atoptech (because they're new tools), but I have heard that Cadence PKS
does well in area. I know of one customer who used to run PhysOpt for
placement/timing/congestion, and then he's later his design through Cadence
PKS to reduce its area.
For HIERACHICAL DESIGN, the breakout I'd bet on would be:
1. Cadence
2. Magma
3. Synopsys
4. Mentor Sierra
5. Atoptech
Cadence easily owns this because of their 5 years with First Encounter.
Hierarchical design is all about floorplanning, modeling, and budgeting.
Cadence is really good at this. I hear all the time about people using
First Enounter for their hierarchical and then using Synopsys or Magma
flows for their P&R.
Magma is also good at hierarchical, but you're stuck in their flow 100%.
Synopsys also has good hierarchical, but they're stuck porting Jupiter-XT
over to ICC. Jupiter-XT is still Avanti Scheme script and Tcl is an add-on
for it. The majority of Jupiter-XT commands are still "form based", which
means you have to type in 20 lines for one real command. Every extra switch
requires a new line of code. Replay scripts are nightmare in Jupiter-XT
compared to any true Tcl-based tool. If it wasn't for this porting problem,
I'd give Synopsys #2 here. ICC was made for flat designs, not hierarchical.
I did a design where Jupiter-XT placed 500 macros and it worked nicely.
Again, Mentor Sierra and Atoptech have problems in hierarchical because
they're so new.
If you look purely at ROUTING, IC Compiler would probably win now. It'd
win on wire-lengths, via counts, redundant via counts, and SI/crosstalk.
(Three years ago, Magma was giving Synopsys headaches, but at 65/45 nm
Synopsys caught up. Magma's sweet spot was 130 nm.) My bet:
1. Synopsys
2. Cadence
3. Mentor Sierra
4. Magma
5. Atoptech
Where Cadence will win is with its shape-based fill; Synopsys is not good at
that. Mentor (Sierra) won't do well because its too new and Calibre DFM is
all done in GDSII -- you can't measure timing impact in GDSII because for a
sea-of-gates design the extraction will take forever. (Literally, it might
take 10 years!) For routing, I've also heard that Magma can sometimes
be heavy on jogging. But overall, I'd say Magma and Sierra are tied, cause
the Magma router is more mature; yet Mentor brings Calibre in with Sierra.
I don't think Atoptech has anything for DFM/SI, so they're last.
At 65/45 nm, what's the most important in routing are the DFM/SI issues plus
the real life pain spent fixing on-chip variation. This is where Atoptech
falls apart. They don't even have tape-outs yet. Quick P&R of fast blocks
that you can't cheaply manufacture, you can't test, and which consume a lot
of power is a complete waste of everyone's engineering time and dollars. In
a point tool flow Atoptech might find a niche, but for small COT houses with
limited tool budget, Atoptech and Mentor/Sierra won't be major players until
they can bring a real, full production proven flow to the party.
- Jonathan Bahl
COT Consulting, Inc. Toronto, Canada
---- ---- ---- ---- ---- ---- ----
From: John Cooley <jcooley=user domain=zeroskew not calm>
Hi, All,
I just wanted to add that on Oct. 1, Raj Seth of Cowen & Co. put out a note
that referenced this Broadcom P&R benchmark specifically:
Following Benchmarks -- BRCM rolling out Talus (LAVA next gen product)
... We've seen some references (DeepChip.com) to a major west coast
benchmark result that suggest LAVA finished behind several start-ups
and CDNS. This same report implied that SNPS finished last. We don't
the exact order of finish in the benchmark (or the exact criteria) but
we're confident that LAVA finished ahead of larger incumbents. ...
... Very often a relatively narrow point solution can show well but not
scale or not have other required funtionality. What we are confident
about is that LAVA withstood attack by all competitors and will retain
dominant share at BRCM. In fact, BRCM is now in the process of rolling
out LAVA's new flagship implementation tool, Talus. ...
... For some time, CDNS had the best detail router on the market with
NanoSim. According to LAVA they feel confident in Blast Fusion's
ability to benchmark against larger competitors, but they need Talus
to beat the capability in some newer solutions. In any event, LAVA
remains the one to beat at BRCM.
So basically, both guys are saying that Atoptech really isn't ready for
prime time yet -- but if you believe the engineer, Synopsys pretty much won
the Broadcom P&R benchmark -- and if you believe the Wall Street analyst,
it was Magma that actually won the Broadcom P&R benchmark. You decide.
- John Cooley
ESNUG/DeepChip.com Holliston, MA
Index
Next->Item
|
|
