( DAC'19 Item 11 ) ------------------------------------------------ [06/12/20]
Subject: Redhawk/ICC2/Fusion vs. Voltus/Innovus/Tempus is Best of 2019 #11
REDHAWK HAPPY; REDHAWK-SC MEH: 5 years ago John Lee in his Gear start-up
joins Anysys Apache with his new Hadoop-driven IR-drop tool that's supposed
to revitalize Redhawk. Then 3 years ago, even after John Lee renamed his
new Gear IR-drop tool to be "Redhawk-DMP", and then to be "SeaHawk", and
then to finally be "Redhawk-SC" -- it still lost benchmarks to CDNS Voltus:
The Nvidia stealth benchmark of CDNS Voltus vs. ANSS Gear RedHawk
http://www.deepchip.com/items/0560-03.html
A second CDNS Voltus-DP vs. ANSS Gear RedHawk-DMP user benchmark
http://www.deepchip.com/items/0561-01.html
Plus in 2016, a neutral outside observer, Joe Sawicki of Mentor, cited in
the DAC'16 Troublemakers Panel about what is now called "Redhawk-SC":
"I've heard some things in terms of power, in terms of one
solution, where if you do a different number of CPU's you get
a different answer for nodal voltage. (See ESNUG 561 #1)
That's not good. It's worse than inaccurate, inaccurate is
kind of eh, compared to what? But inconsistent sucks."
- Joe Sawicki, DAC'16 Troublemakers Panel (ESNUG 563 #10)
But that was in 2016. If you look at the Redhawk-SC user comments now, it's
getting both "I love it" and "it sucks" comments.
---- ---- ---- ---- ---- ---- ----
AART TO THE APACHE RESCUE: during this same time while Ansys Apache Gear
Redhawk-SC was floundering -- Aart had an even worst problem -- his very
own SNPS PrimeRail was even in worst condition at nodes under 28nm; which
spells DOOM for his ICC/ICC2/Fusion Compiler PnR tools if he doesn't fix
that problem damn fast!
So these two EDA guys with compatible problems get technology-married in
order to fix their respective problems. And their "baby" is Redhawk Fusion;
an integration between John Lee's (non-Gear) Redhawk and Aart's PnR tools;
while Aart agrees to kill off his Synopsys PrimeRail tool and have his SNPS
salesdroids promote only Ansys Apache Redhawk instead.
REDHAWK STILL ROCKS: On the customer side, for Redhawk (not Redhawk-SC) the
users thought that this was brilliant of John Lee to team up with Synopsys
PnR like this. I have to tip my hat to him for pulling this off. It's not
often two former *rivals* turn around to instead work together like this.
"We're using Redhawk Fusion with Synopsys Fusion Compiler at 7nm.
It's amazing how John Lee got two fundamentally different db's
to more or less play nice with each other. The integration is
not seemless, but it's damn good."
"Redhawk is excellent on ICC2 blocks. We don't use Redhawk-SC."
---- ---- ---- ---- ---- ---- ----
VOLTUS/INNOVUS SELLS: Long story short, if the Voltus user comments below,
by and far the main reason why customers are buying Voltus is *because* of
it's super tight hooks with Innovus and Tempus. That is, from what I've
seen, nobody is using Voltus with ICC2/Fusion Compiler instead of Redhawk.
On the technical side it was interesting to see users actively switch from
Redhawk to Voltus for sign-off -- showing Anirudh has some wins here.
"We use Voltus for sign-off and have taped-out an extremely high-
performance block at 7nm."
"For power sign-off we've used Voltus at multiple process nodes.
It does what we need for power."
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
ANSYS APACHE REDHAWK + SYNOPSYS ICC2 READER COMMENTS
Redhawk should be in that list.
---- ---- ---- ---- ---- ---- ----
John Lee was brilliant to get Redhawk tight with ICC2.
---- ---- ---- ---- ---- ---- ----
We're long time Apache Redhawk and Synopsys ICC/ICC2 users.
---- ---- ---- ---- ---- ---- ----
Apache Redhawk does the job.
---- ---- ---- ---- ---- ---- ----
We're using Redhawk Fusion with Synopsys Fusion Compiler at 7nm.
It's amazing how John Lee got two fundamentally different db's
to more or less play nice with each other. The integration is
not seemless, but it's damn good.
---- ---- ---- ---- ---- ---- ----
We're an early Gear SeaScape user. It's now called Redhawk-SC.
John Lee is a smart man. He took beta code and turned it into
the mature industry leading IR-drop analysis tool.
Teaming up with Aart's ICC2 was a masterstroke.
---- ---- ---- ---- ---- ---- ----
We weren't happy with the early Ansys SeaScape tool.
Since then, it's grown by leaps and bounds.
---- ---- ---- ---- ---- ---- ----
Most of our projects are 16nm.
Redhawk + ICC/ICC2 has worked well for us for years.
---- ---- ---- ---- ---- ---- ----
We like that for our blocks in ICC2, analyze_rail calls RedHawk
inside ICC2. We like that power noise is handled early at the
PnR block-level instead as an afterthought in overall PnR.
---- ---- ---- ---- ---- ---- ----
The Cadence sales guys came by touting Voltus inside Innovus.
Don't tell them that the problem is we have too many man-centuries
invested in ICC2/Redhawk to make a major tool switch like that.
Also, ANSS and SNPS work exceptionally well in supporting the
integration of ICC2 with Redhawk. We're not used to two EDA
rivals working together like this. We're not complaining.
---- ---- ---- ---- ---- ---- ----
Looking at Voltus because we feel we leave too much margin on the
table with Redhawk/ICC2. (My mgmt in never happy.)
---- ---- ---- ---- ---- ---- ----
Ansys Redhawk -- great
Ansys Redhawk-SC -- still meh
---- ---- ---- ---- ---- ---- ----
Redhawk-SC is overwhemed at 7nm.
---- ---- ---- ---- ---- ---- ----
We script Redhawk to run block after block after block.
It's not elegant but it does the job.
---- ---- ---- ---- ---- ---- ----
We cheer Anirudh for taking on Redhawk.
Customers always win when there's competition.
---- ---- ---- ---- ---- ---- ----
Machine learning in Redhawk-SC caught our attention this year.
---- ---- ---- ---- ---- ---- ----
1. ICC2
2. Redhawk
3. Calibre
---- ---- ---- ---- ---- ---- ----
When we shifted from ICC over to ICC2, the Synopsys guys were
adamant that we stick with Redhawk Fusion with ICC2.
---- ---- ---- ---- ---- ---- ----
Always used Redhawk for IR-drop analysis.
---- ---- ---- ---- ---- ---- ----
Redhawk is excellent on ICC2 blocks. We don't use Redhawk-SC.
---- ---- ---- ---- ---- ---- ----
We're a SNPS/ANSS house. ICC2/Redhawk.
---- ---- ---- ---- ---- ---- ----
I might be biased. I bought our first Apache Tomahawk-S license
from Andrew Yang and Keith Mueller back in 2003.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
CADENCE VOLTUS/INNOVUS READER COMMENTS
---- ---- ---- ---- ---- ---- ----
Cadence Voltus
We use Cadence Voltus.
Voltus has continued to get better, and we've gotten better at using
it.
In the last few months, running Voltus EMIR inside Innovus and Tempus
has been a more integral part of our methodology.
- We run it to make timing, power grid and placement design
decisions.
- We use Voltus for sign-off and have taped-out an extremely high-
performance block at 7nm.
Measuring power at design time can be challenging.
- Voltus runs SPICE under the hood and takes power grid into account
for better accuracy.
- Voltus' tight integration with Innovus and Tempus makes our design
flow very efficient and removes the possibility of human errors.
I heard from a colleague that Cadence Voltus is doing better on large
designs in total vs. ANSYS Redhawk/Seahawk. (I haven't done a
firsthand comparison myself.)
---- ---- ---- ---- ---- ---- ----
Cadence Voltus
Our team has moved completely over to Cadence Voltus.
Our initial motivation for using Voltus was its integration with
Innovus. Then over time, it was accurate enough and had good enough
features where we no longer needed our prior commercial tool.
For power sign-off we've used Voltus at multiple process nodes. It
does what we need for power. Cadence does killer integration for:
- physical design (Innovus)
- packaging & simulation
The trend we are seeing is package co-design.
- You throw a test case at the entire system and try to simulate
it. E.g. SoC + Package + Board model
- There is a move afoot to integrate the process more, such as
file format consistency and putting it together into a matrix
solver.
10 years ago, SoC simulation was fine, but now current scaling is taxing
system-level design. We need to do a lot more, and earlier, to ensure
power integrity.
The more you can integrate, the more "good" you get.
---- ---- ---- ---- ---- ---- ----
We've used Voltus on 7nm designs that have gone to silicon.
Voltus' #1 benefit for us is its integration with Innovus, Genus
and Xcelium; which eliminates the manual data transfer between the
various design and verification steps.
Additionally, we do power grid ECOs using Voltus and Innovus, so the
integration is a big factor there, also.
Cadence claims a billion-gate capacity, but we haven't tested that,
as our designs are under a million gates.
We have validated Voltus' accuracy based on its strong correlation for
IR-drop when compared with silicon.
---- ---- ---- ---- ---- ---- ----
We've used Cadence Voltus for signoff for more than a year now.
It seems to be getting some traction in the industry these days.
Ansys looks like they are losing market share here to Cadence.
Voltus' integration with other Cadence (Genus/Tempus/Innovus) tools
is pretty solid.
Our hardest problem in design closure right now is IR/EM/reliability
closure. We are starting to look at a flow for proactive IR fixing in
Innovus with Voltus-provided data. There is a large motivation to look
at putting in additional grid-based power in our design style.
---- ---- ---- ---- ---- ---- ----
Cadence Voltus
Cadence has made many improvements to Voltus over the past 6 years.
- Voltus is now a breeze to use.
- We can now do IR or EM analysis on over 2 million gates with
Voltus in only 2.5 to 3 hours.
- Its close integration with Innovus and Tempus.
Cadence Tempus, Voltus and Innovus working together is very intricate
and we use it in a production capability.
- Innovus initiates IR-aware placement with Voltus.
- Cells impacted by IR-drop can have their timing compromised.
- Tempus timing is called from Innovus to time the DB.
- Voltus provides feedback to Innovus and placement gets tuned
to reduce these IR effects
- The new placement database is sent to Tempus to retime and
verify improvement.
- This process is iterative.
Cadence Voltus is not best power tool out there, but it's one of the
good ones. We still use Redhawk for our final check but we have found
very few (if any) violations in Redhawk after finishing the design
with Innovus and Voltus.
---- ---- ---- ---- ---- ---- ----
Cadence Voltus
Cadence Voltus and Innovus command integration is good. If Innovus has
a specific command, I can find either the same command or a similar one
in Voltus.
The Voltus GUI is also easy to use, e.g. the widgets and the layout.
---- ---- ---- ---- ---- ---- ----
We use Cadence Voltus and Innovus.
Voltus is tightly integrated with Innovus. We use it for static power
analysis, but it's not as accurate with dynamic analysis.
We use Redhawk for dynamic power analysis. Unfortunately, dynamic power
analysis is too hard to do during design, as we can't dynamically
analyze and optimize at the same time.
So, we are probably overdesigning power right now, because of a lack of
tight loop between dynamic power analysis and place and route.
---- ---- ---- ---- ---- ---- ----
Related Articles
Costello on EDA ossification, cloud, and RedHawk vs. GreenHawk
Anirudh and Sawicki on iffy Apache IR-drop #'s vs. Voltus/Innovus
A second CDNS Voltus-DP vs. ANSS Gear RedHawk-DMP user benchmark
The Nvidia stealth benchmark of CDNS Voltus vs. ANSS Gear RedHawk
ANSS "Jolly" on why Big Data is a bad fit for EDA and chip design
Industry Gadfly: "A brief history of Apache and its IPO"
Join
Index
Next->Item
|
|
