!!! "It's not a BUG,
/o o\ / it's a FEATURE!" (508) 429-4357
( > )
\ - / INDUSTRY GADFLY: "Two stories Aart forgot to complete"
_] [_
by John Cooley
Holliston Poor Farm, P.O. Box 6222, Holliston, MA 01746-6222
Although I couldn't attend this year's San Jose SNUG'11 conference, I was
a little suprised by two big tool announcements I had heard Aart de Geus,
CEO of Synopsys, had made in his keynote this year.
What was interesting wasn't specifically what Aart said, but what Synopsys
Corporate Marketing had Aart "forget" to say...
DC EXPLORER:
Aart's first announcement in his SNUG keynote was a new RTL estimator he
called "DC Explorer". It's basically a fast RTL synthesis tool so you can
do timing/area constraints tradeoffs early on in your RTL. In addition,
DC Explorer interfaces to your downstream DC Ultra and ICC tools.
The SNPS press release had an ST bigwig director testify that he had "seen
at least a 4X faster runtime and 10% correlation to DC Ultra."
Go to Synopsys.com and you'll see SNPS Corporate Marketing data sheets with
data points bragging up to 12X faster, 1% timing and 3% area correlations.
Hell, they even had a fancy 3 min video with SNPS GM Antun Domic chatting up
how super sweeeeeeeeeet DC Explorer was.
This was a BIG marketing push by Synopsys Corporate Marketing! Big!
THE REST OF THE STORY:
Wow. Oasys RealTime has scared Synopsys Corporate Marketing; or at least
those in SNPS marketing whose job is to promote & protect Design Compiler.
Put it all together. Design Compiler is 25 years old; DC does world class
RTL synthesis. You name it; DC has seen it before and can synthesize it.
Seriously, there are parts of Design Compiler that date back to 1986. It's
2011 now. That's 25 years. That's 25 year old internal data structures
and 25 year old core algorithms at the heart of DC.
Then in 2010 in comes Oasys RealTime in ESNUG 484 #2:
"We were initially struggling to get a few of our big blocks synthesized.
Synopsys DC choked due to their large size and it was either crashing or
was taking more than 2 to 3 days to synthesize.
So we tried 3 big blocks of 5.8M, 2.7M, 1.6M instances in Oasys RealTime.
RealTime does not use wire load models during optimization. Instead it
partitions the design and does a congestion-aware global placement. It
then uses the placement information to calculate the approximate wire
delays and optimize your design based on this loading information.
Due to the comparitively slower runtimes & lower capacity of Synopsys DC,
we had to do "bottom up" compiles for our bigger blocks in DC. With
Oasys we didn't have to partition our big blocks in to smaller sub-blocks.
We've found this can result in a better optimized design overall as more
optimization is done closer to flat.
Our biggest 5.8M block finished in 70 min with peak mem usage under 2G."
And in 2011 in ESNUG 487 #1:
Oasys RealTime Synopsys DC-Graphical/Topo
-------------- --------------------------
1.2 M gate block runtime 1.2 hrs 9 hrs
3.6 M gate block runtime 2 hrs 3+ days
It's unfair. Oasys RealTime uses algorithms and data structures that were
not even around back in 1986 -- that's why it's 30X to 60X faster.
While DC is amazing in all of the crazy corner cases it can synthesize, it
is simply too big and too complex and too old to speed up.
That's why Synopsys Corporate Marketing had to come up with DC Explorer;
it's a 4X speed-up pre-processor for the generic DC donkey cart runs to
help overcome those 60X Ferrari runs Oasys RealTime intrinically has.
Or look at this another way -- in its entire 25 year history, SNPS has never
once bothered with an RTL estimator tool for DC -- until now. What changed?
"We now have a 4X turbocharger on our Design Compiler donkey cart! Yay!"
Ouch.
---- ---- ---- ---- ---- ---- ----
As far as RTL estimator tools go, I see some overlap & some key differences
between Synopsys DC Explorer and Atrenta Spyglass Physical.
Synopsys DC Explorer
- fast RTL synthesis for area and timing, but no power nor
congestion considerations
- constraints tradeoffs for area and timing, but not power
- it's mostly an RTL tool, not a physical routing tool
- interfaces to Synopsys DC and IC Compiler only
Atrenta Spyglass Physical
- fast RTL synthesis for area, timing, power, and congestion
- constraints tradeoffs for area, timing, and power
- takes physical routing into consideration for the RTL folks
- uses a rules-driven approach
- interfaces to Synopsys DC and ICC, Cadence Encounter,
Magma Talus, Mentor Olympus SoC, Atoptech Aprisia
Here's an ST user's first look at Atrenta Spyglass Physical in ESNUG 486 #4.
"We used Synopsys DC/DC-Topo for synthesis. We used Cadence SOC Encounter
or Mentor Olympus for place and route, as well as Synopsys IC Compiler
for some macros. The maximum size of our defined Physical Units (block
groupings) was at ~2 M instances to allow for reasonable P&R runtimes.
The Spyglass Physical benchmark data below is based on our design:
Size: 7 M placed instances, 209 M transistors, 500 signal pads
Clocks: 230 total clocks (80 created, 150 generated)
Blocks: 53 RTL IPs, 40 RTL blocks/glues and 160 Hard IPs,
8 different IO libraries, 477 memory cuts
Our Spyglass Physical runtimes were as follows:
- Data prep (tech pre-compilation done once): 4 hours
- RTL floorplanning & prototyping: 4 hours per run
- SOC Encounter netlist floorplanning & prototyping: 8 hours per run
- SOC Encounter or Olympus P&R to timing: 1 week per Physical Unit
With Spyglass Physical we are able to move our physical partitioning
decisions up to the RTL/Pre-RTL level instead of the netlist level.
We've 2 tapeouts with Spyglass Physical, including one in 32 nm. We are
now deploying it in other groups of our division."
- Thierry Sejourne of STmicroelectronics in ESNUG 486 #4.
And at least this way you can have a true best-in-class-tools-at-every-stage
chip design flow instead of being trapped in a Synopsys-only flow.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
I laughed when I read the EE Times write-up of Aart's SNUG cloud computing
announcement. Especially the story's closing:
"In the EDA industry we are still in a learning mode to see if it
will work or not."
- Hasmukh Ranjan, VP at Synopsys (EEtimes 3/29/11)
Did I read that right? Is Synopsys Marketing honestly trying to pretend
that this is *really* first time Synopsys, Inc. had tried to promote cloud
computing and/or EDA SaaS for its customers?
None of the company mktg bloggers on Synopsys.com mentioned DesignSphere;
all they're saying is how neat the idea is.
None of the so-called "independent" EDA bloggers mentioned DesignSphere;
they're too busy brown-nosing Aart and SNPS for being so brilliant.
I searched for DesignSphere on Synopsys.com. Yup, all the press releases
from 2000 touting DesignSphere were "magically" gone. The only thing left
was an 10-K Annual Report from 2000 where SNPS mentioned DesignSphere once.
"Hey! You can't edit a 10-K Annual Report! It'll get us in big trouble with
the SEC! Let's just hope no one remembers DesignSphere..."
It's funny how that works.
THE REST OF THE STORY:
Anyway, despite Synopsys Marketing (and its friends) "forgetting" history,
back in 2000 right before the Dot Com Bubble was about to burst, Synopsys
and Avanti shocked the EDA world when they teamed up at DAC'00 with TSMC to
sell a complete RTL-to-GDSII flow on a web site called DesignSphere.com.
(Remember SNPS didn't buy AVNT until 2002. In 2000, the EDA World War then
was CDN vs. SNPS plus AVNT.) The neat thing about DesignSphere.com was that
it had SNPS Design Compiler, VCS, Primetime, and AVNT Jupiter, Saturn, Mars,
Apollo, Hercules, Star-RC and TSMC Cybershuttle in it.
"There are very few announcements that build a complete value chain,
and there are even fewer that build a value chan with best-in-class
solutions. What you have here is tic tac toe-three in a row."
- Aart de Geus, CEO of Synopsys (EEtimes 6/1/00)
SNPS expected to charge $500 K to $1.5 M per DesignSphere.com seat.
"Tool licensing is not much different with DesignSphere than it is
with our current term license environment. Where we normally offer
one-, two-, or three-year licenses, DesignSphere will be a little
more flexible in terms of granularity, but the pricing is
essentially the same."
- Dave Burow, VP at Synopsys (EEtimes 6/1/00)
Synopsys' DesignSphere.com was a direct response to the two other EDA cloud
computing web sites at the time: ToolWire.com and SliceX.com. Synopsys had
40 employees working purely on DesignSphere.com back in June 2000.
In October 2000, Aart was showboating DesignSphere.com at Boston SNUG'00.
By 2002, all three EDA cloud computing SaaS web sites had flopped.
"Where we went wrong, John, was in not understanding the emotional need
customers had to be near their computers and data. Most just could not
get comfortable with the idea that their computers and data were in a
remote building operated by a 3rd party."
- Dave Burow, SVP at Synopsys (Industry Gadfly 7/26/01)
Interestingly, four months later in November 2002 Dave Burow leaves SNPS
to become CEO of Arithmatica.
Time passes...
By 2008, Cadence had gone through two CEOs (Ray Bingham and Mike Fister) and
lost a good part of its company memory from two purges of Cadence upper and
middle managers. The new guys under Lip-Bu Tan naturally think Web 2.0 is
the hot new thing because "everyone knows Web 2.0 is the hot new thing."
September 2008, Cadence debuts its Hosted Design Solutions, which is fully
cloud computing, SaaS marketing buzzword compliant. They offer "custom IC
design, logic design, physical design, low power, functional verification,
and digital implementation." By 2011, Cadence Marketing reports they have
35 customers using it and all four reference accounts are tiny full custom
start-ups using a one-size-fits-all Cadence Virtuoso flow:
Tagent, Inc. Virtuoso 5 engineers in Mountain View
U-Blox AG Virtuoso 12 engineers in Europe
HKSTP Corp Virtuoso 7 engineers in Hong Kong
Adept IC Solutions Virtuoso 3 engineers in San Jose
"One of the reasons why DesignSphere failed back then was customers were not
comfortable having their designs on someone else's machines," said Tom Anderson
of Cadence. "Given the choice everyone would want to have their own server
farm because it gives you complete control. But for small design houses it's
not cost effective to own all those machines. These types of customers have
been our sweet spot once we earned their trust."
It's a WIN-WIN for CDNS and certain small customers. For small users:
- A one-size-fits-all Cadence Virtuoso flow is OK if you're not doing
anything too challenging on your custom chip. It'll do.
- Small users don't have to deal with pesky and troublesome SysAdmin,
machine, and licensing issues. It's all taken care of!
- It's an easy entry point for small start-ups who can't afford big
EDA budgets. Billing by the hour makes sense and is manageable.
For Cadence:
- Cadence Sales doesn't have to deal with those pesky rivals and their
troublesome Virtuoso add-ons like Apache Redhawk, Mentor Calibre,
Synopsys Star-RC & Tarus, Magma FineSim & Titan, Prolific ProGenesis,
Silicon Frontline F3D, Solido VD, Berkeley AFS, and of course that
Ciranova Helix from those evil PyCell guys. They're all locked out!
- Like early cell phone schemes, Cadence Sales gets to hook users with
cheap deals at first; then nickle-and-dime them later with roaming fees,
texting fees, etc. It's a little baby cash cow for Cadence.
It's only when a design house gets bigger is when they realize the savings
in buying whole EDA licenses and they fully understand all the competitive
advantages of having a true best-in-class-at-every-stage design flow. "If
your wife is happy wearing Levi jeans and one-size-fits-all T shirts from
Walmart, it's a really bad idea to give her gift cards for Nordstrom's."
Time passes...
Again, interestingly, in December 2010 Dave Burow becomes the CEO of a
small Redwood City, CA company in stealth mode called CloudOpt, Inc.
Three months later at SNUG 2011, Dave Burow's old boss, Aart de Geus is up
on stage talking about cloud computing like it's something new. This time
the differences are instead of selling a total SNPS flow like Aart tried
back in 2000 on DesignSphere.com, Aart wants to go slower and just sell VCS
and HSPICE runs on a 100% Synopsys-controlled part of Amazon Web Services.
It's a clever strategy. I'm sure the internal SNPS marketing plan looks
something like this:
Stage 1: Most projects go through a last stage where they need a burst of
simulation licenses to validate that one final design. Synopsys
offering hundreds of VCS or HSPICE licenses in a cloud at a
pennies-per-CPU-per-hour can be mighty helpful at these times.
Stage 2: Once SNPS can get your company's bigwigs and lawyers acclimated
to having your designs OUTSIDE of your company's firewall,
Synopsys will naturally offer more and more SNPS-only tools
inside their 100% Synopsys-controlled cloud.
Stage 3: Like SNUG meetings, customers will get used to a SNPS-only cloud.
Stage 4: As gatekeeper for all EDA tools allowed on the SNPS cloud, say
goodbye to Magma Talus & Titan & FineSim, Cadence Encounter &
Conformal & NC-sim, Mentor Calibre & Olympus & ModelSim, Atoptech
Aprisa, Extreme DA, Apache RedHawk and any other tool or company
that Synopsys Marketing sees as a threat.
Stage 5: Over time, SNPS competitors go in decline or are marginalized
because they can't get on the SNPS cloud.
Stage 6: With this new Apple iStore business model finally in place,
Synopsys can use its near-monopoly power to raise EDA prices.
The problem is that pesky Stage 2 above.
Right now your company's bigwigs and lawyers fear cloud computing because of
the outages, hacker attacks, data losses and data thefts in the news at big
names like Amazon Web Services EC2 and the Sony Playstation Network.
After a server outage last week that took down several web sites such
as Reddit, Foursquare, HootSuite and Quora, Amazon Web Services revealed
in a post that it will not be able to recover about 0.07 percent of the
volumes in its US-East Region that were affected by the blackout.
- Washington Post (4/27/11)
Sony announced recently that all 77 million people who use their
PlayStations to play online video games may have had their names,
and maybe their credit card numbers, stolen by hackers. The hackers
also obtained users' addresses, birthdays and billing history.
- Credit.com (4/27/11)
But over the long term, your company's bigwigs and lawyers wisely fear any
form of cloud computing that's controlled by one of your tool suppliers:
Richard Stallman, founder of the Free Software Foundation and creator
of the computer operating system GNU, said that cloud computing was
simply a trap aimed at forcing more people to buy into locked,
proprietary systems that would cost them more and more over time.
"It's stupidity. It's worse than stupidity: it's a marketing hype
campaign," Stallman told The Guardian. "Somebody is saying this is
inevitable; and whenever you hear somebody saying that, it's very
likely to be a set of businesses campaigning to make it true."
The 55-year-old New Yorker said that computer users should be keen to
keep their information in their own hands, rather than hand it over
to a third party.
- Richard Stallman (The Guardian 9/29/08)
Or, as I'm sure someone at Synopsys Marketing said: "Damn those pesky user
bigwigs and their lawyers! They're ruining our great little Web 2.0 money
scheme we cooked up for Aart this year. Damn it!"
- John Cooley
DeepChip.com Holliston, MA
P.S. Keep in mind I'm NOT saying cloud computing is a bad idea; after you
read this UC Berkeley paper you might want to try your own cloud for
your company -- you just DON'T want SNPS or CDNS running it for you!
-----
John Cooley runs DeepChip.com, is a contract SoC designer, and loves
hearing from engineers at or (508) 429-4357.
|
|