( DAC'19 Item 8b ) ------------------------------------------------ [04/14/20]
Subject: Cadence Tempus fast ECOs, sign-off, and MMMC is Best of EDA #8b
YOU'VE COME A LONG WAY BABY: It was 7 years ago when I scooped that Cadence
was taking on the Synopsys PrimeTime monopoly with their Tempus tool.
- from http://www.deepchip.com/items/0524-01.html
Back then in 2013 engineers used Tempus with Innovus for quick PnR/STA
interations -- but they still very religiously used PrimeTime as their
final golden sign-off tool. Then over the years Tempus slowly got to
sign-off quality accuracy.
In this year's report, the user confirm that:
- OLD NEWS: They use Tempus on its own for sign-off, which many
have done for years now ...
- NEW NEWS: the users in this survey think Cadence Tempus + Innovus
has leapfrogged Synopsys PrimeTime + ICC2 with overall integration,
timing ECO integration, and distributed compute MMMC.
Here's the pull quotes citing what the users are specifially saying.
On integration:
"The eye opener here is we've consistently seen the Cadence
flow beat Synopsys DC + ICC/ICC2/Fusion + PrimeTime flow."
On sign-off:
"We've have now had more than 15+ tapeouts, plus silicon
validation from 40 nm to 16 nm with Tempus. So, we are
very comfortable with not using PrimeTime and instead
using Tempus as our STA sign-off tool."
On timing ECOs:
"The two tools [Tempus/Innovus] have a handshake between
them, so for ECOs you don't have to lift a finger."
On MMMC:
"We had large number of corners (50+) because of our wide
voltage range of operation. Tempus' concurrent MMMC was
critical for signoff across these corners."
On runtime:
"Our Tempus/Innovus STA + ECO cycle can be done in 24 hours
or less."
On accuracy:
"We used LVF extensively and correlated SPICE with Tempus.
We got better than 5% mismatch between MC SPICE and Tempus."
TEMPUS OVERALL: "We have used Cadence Tempus for a number of tapeouts. It is
solid for us. It has no major bugs/issues. Cadence has baked it enough now."
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
Cadence Tempus
We stopped using Synopsys PrimeTime for our timing sign-off and instead
we use Tempus during Innovus PnR and for our timing sign-off.
The eye opener here is that we've consistently seen the Cadence flow
beat Synopsys DC + ICC/ICC2/Fusion + Primetime flow.
Tempus + Innovus Timing ECO
What we like this year is Tempus now has tightly integrated ECOs within
Innovus for setup and hold fixing. This is beautiful as you no longer
have to run Tempus separately to do an ECO.
The way it works is that you run sign-off timing from Innovus, and give
it options to recover the timing:
1. Innovus sends the design to Tempus to time it (from
the Innovus cockpit)
2. Innovus brings back the ECOs that Tempus suggests
3. Innovus implements the ECOS and then reroutes the design
4. Innovus sends the revised design to Tempus to time it
to make sure it met the spec.
You can do this multiple times if needed, and never have to run/call
Tempus separately. The two tools have a handshake between them, so
for ECOs you don't have to lift a finger.
This is a big deal for us because of the time savings, and a serious
differentiation from Synopsys ICC2 and PrimeTime.
Tempus now has concurrent MMMC.
- We can have multiple delay constraint files.
- Tempus can run multiple corners through different CPUs
at the same time.
- Tempus tracks all jobs spun off and coming back.
This means we can now get very high throughput.
Tempus is an excellent tool. It also gives us SPICE accuracy and you
can't ask for better than that.
---- ---- ---- ---- ---- ---- ----
Cadence Tempus
I use Cadence Tempus with Innovus APR. We run it on 8-16 CPUs.
Tempus' ECO is very good.
- I feed the APR output files into Tempus to get sign-off timing.
- I try the ECO changes right in the Tempus database, as a fast
way to get an indicator of the timing impact.
- Run the ECO change through Innovus to close the design.
Cadence's Tempus and Innovus command correlation is also good. They
share the same format, syntax, options -- so it's very efficient
Because the same commands can be executed for both tools, it improves
our productivity. (another integration benefit)
My overall conclusion about Tempus' strengths -- is that it's fast,
accurate, and good integration with APR, including ECO.
---- ---- ---- ---- ---- ---- ----
Cadence Tempus
Anirudh's Tempus/Innovus integration makes our ECO flow efficient.
It also handles AOCV and derates recommended by the foundries well
for timing margins.
---- ---- ---- ---- ---- ---- ----
We've been using the Tempus/Innovus ECO flow extensively for a couple
of years now. It continues to get progressively better. What we like
is it's now seamless. When we do a timing ECO, we don't have to do
separate Tempus and Innovus runs now; moving data back and forth.
We also like that Cadence does hierarchical ECOs, so we can:
- Optimize our top-level interface logic. (We could not do
this when it was just a sub-chip.)
- Push the changes into the blocks.
We require a lot of speed for STA and have run Tempus on up to 64 CPUs.
We also use concurrent MMMC and are happy with it and we use statistical
signoff for our latest devices. So far, our final silicon performs as
expected with no timing issues.
We have used Cadence Tempus for a number of tapeouts. It is solid for
us. It has no major bugs/issues. Cadence has baked it enough now.
Our biggest STA headache in general is that it has large netlists and
a lot of corners. To accommodate the big designs, companies go wider;
adding more CPUs to achieve overnight runs. The tradeoff is that we
run so many MMMC modes across the CPUs it becomes a huge compute
investment for us.
---- ---- ---- ---- ---- ---- ----
Cadence Tempus has very good correlation with Innovus. Since Innovus
is our primary PnR this greatly helps us close our timing much faster.
Our Tempus/Innovus STA + ECO cycle can be done in 24 hours or less.
Tempus' has Concurrent MMMC, which works well. We use it in all our
runs. It also has integrated ECO with Innovus (setup/hold fixing),
that we rely on it a lot for closing timing.
CAVEAT: We tried Cadence's "fully distributed" approach (multi-CPU and
multi-threading). For our designs, the performance benefit vs. the
number of licenses did not warrant the need. If we had much larger
and more complex set of views in the STA, then we would definitely
switch over to it.
Tempus has definitely come up as a strong alternative to Synopsys
PrimeTime. Tempus is generally on the conservative side with regards
to to default settings. That gave us a more comfortable feeling on
signoff than when we first used Tempus some years ago.
We've have now had more than 15+ tapeouts, plus silicon validation
from 40 nm to 16 nm with Tempus. So, we are very comfortable with
not using PrimeTime and instead using Tempus as our STA sign-off tool.
---- ---- ---- ---- ---- ---- ----
We now use Cadence Tempus exclusively for signoff. LVF validated and
works - no issue. It's definitely our signoff tool now for our 16nm
and below nodes.
We use it on 16nm, 28nm, and older node designs, for broad market
and automotive applications.
The Tempus single pass ECO now works reliability for timing closure.
(Cadence R&D took some time to fix that!)
Room for improvement: Add the ability to recover area in critical
regions to improve timing ECOs.
Cadence has now also integrated EMIR + STA, but we have not gone
down that path yet.
Cadence has made excellent progress with Tempus ECO. Our designers
no longer find low-hanging fruit with missed ECO optimization as
Tempus seems to find all the obvious solutions.
---- ---- ---- ---- ---- ---- ----
We used Tempus with Innovus for a 7 nm design and the integration
works well.
---- ---- ---- ---- ---- ---- ----
Cadence Tempus
We've used Tempus at 55 nm, 90 nm and 180 nm nodes.
We had large number of corners (50+) because of our wide voltage range
of operation. Tempus' concurrent MMMC was critical for signoff across
these corners.
We used LVF extensively and correlated SPICE with Tempus. We were able
to get better than 5% mismatch between Monte Carlo SPICE and Tempus.
Additionally, the timing tool worked well for us particularly at near
threshold voltages.
We got our silicon back and have not found any problems with it.
---- ---- ---- ---- ---- ---- ----
Related Articles
CDNS Tempus still against SNPS PrimeTime is #8 story at DAC'14
Cadence Tempus vs. Synopsys PrimeTime as #1 hot tool at DAC'13
Leak says Cadence taking on Synopsys PrimeTime monopoly at DAC'13
Join
Index
Next->Item
|
|
