A 7th Clarity vs. HFSS Benchmark

September 18, 2020

The following article was originally posted by DeepChip.

CLARITY VS. HFSS WINS PILING UP: And now, in this report, we have EIGHT
users validating Clarity is real -- SIX of them with direct data against
HFSS.  These 9 (total) benchmarks now prove out the Clarity speedup with
the same accuracy claims, along with its smaller memory footprint claim.

  1. Clarity vs. HFSS Speed-Up

       "We ran Clarity on up to 24 CPU's -- it was generally 7x faster
        than Ansys HFSS."

       "Cadence Clarity had up to a 15X faster turnaround time versus 
        Ansys HFSS ..."

  2. Clarity vs. HFSS Accuracy

       "The accuracy difference between Clarity and HFSS was negligible;
        no more than 0.5 dB for the full frequency band."

  3. Clarity vs. HFSS Memory Footprint

       "Clarity's memory footprint was only 15 GB.  HFSS's footprint
        was 116 GB -- nearly 8X bigger."

MORE CAPACITY == MORE ACCURACY: Users were also excited about the extra
capacity Clarity has vs. HFSS -- not just cause they could run more -- it's
because they could now run the entire design getting more accuracy.

    "The largest block we ran on both tools was an 80 mm x 100 mm PCB,
     with 22 layers and 44 excitations. ...  We've always had accuracy
     concerns with the "cutting" method needed by HFSS to improve runtime.
     Clarity's higher capacity resolves this."

    "We can now put our entire design into Clarity for analysis vs. breaking
     it up.  When we break it up for HFSS, we have to make approximations
     and sacrifice some accuracy.  Not with Clarity."

    - from 6 direct hands-on CDNS Clarity vs. ANSS HFSS user benchmarks

 

From: [ Quagmire, from Family Guy ]

Hi John,

Please keep me anonymous.

I work at a very large company that builds consumer electronic systems.  Our
products are used in everything from telecom, to cell phones, to RF and IoT.
My team is responsible for design, verification, and signoff of our PCB
boards as well as full system package designs.  We easily have way, way over
100,000 collective man-hours using tools like HFSS, Sigrity, and SiWave.  We
are quite comfortable with them and are familiar with the pros and cons.

Here's our experience comparing ANSS HFSS to the new CDNS Clarity tool.


HOW OUR HFSS FLOW EVOLVED

Overall I would say Ansys HFSS is very reliable and gets the job done.  It's
been the gold standard in our business for decades.  But as our designs
become more complex, there's a few things trending I've noticed:

  1. A significant increase in HFSS analysis run time
  2. Increased HW compute resources needed to complete our HFSS analysis.
  3. A need to trade-off capacity for performance with HFSS, which
     affects accuracy

The workaround for these HFSS capacity limits in our larger PCB designs is
we typically do certain level of "reduction", or "cuts", on the PCB design
database so that HFSS is able to run on the hardware server configuration
we can grab.  There are two ways we can "cut":

  1. Cut "unnecessary" signals and geometry patterns; but the risk is
     we may cut important coupling effects or a signal returning path.

  2. "Physically" cut the entire chip/package/PCB system into multiple
     sub partitions; we cut the system into pieces and then run analysis
     to compare against results when the system isn't cut.

Regardless how we "cut", it heavily depends on a user's experiences from
similar designs previously done -- and the final analysis is likely to lose
some accuracy.

 

     
If a brand-new chip/package/PCB design comes to us the very first time,
we have to do a mess of "trial-&-error" iterations to build our confidence
until a reasonable "cut" is achieved.  (In our practice, we tend to keep
more while cutting less because of accuracy, but performance sometimes is a
bottleneck in delivering a solution on schedule if large partitions or
signals are still in the run.  There always seems to be some "magic" that
you need to know in order for HFSS to figure out how an EM field behaves.)


OUR HFSS VS. CLARITY BENCHMARK

Our immediate first question was to see is if Clarity can solve some of the
more challenging problem designs that HFSS couldn't.

We ran multiple design styles to benchmark accuracy/capacity/performance;
but I am only going to single out three as our experiences with Clarity. 

        ----    ----    ----    ----    ----    ----    ----

Testcase #1: Establish Accuracy and Performance

First thing we needed to do was establish a Clarity accuracy baseline with
our "known good" results.  We have several designs to choose from depending
on the analysis criteria.  For this case we chose a small SERDES block that
we have "known good" HFSS results that correlates well to our production
system results.  This flip-chip BGA core package was 45mm x 45mm with 12
routing layers.

                  ANSS HFSS         CDNS Clarity       delta
                  -------------     ---------------    -----------
    Performance:  22 hours 5 min    4 hours 10 mins    5.3X faster
  Mem footprint:  600 GB            90 GB              85% reduction

For servers, both ran on two 32 core machines with 32 CPUs each.

For accuracy, we confirmed the Clarity results matched (+/- 1%) our expected
HFSS numbers and real life measurement numbers.  

Side note: it was very easy to setup and run Clarity.  We're HFSS users.
We only had minimal training from the CDNS AE, then we were able to figure
out the Clarity tool on our own.  

        ----    ----    ----    ----    ----    ----    ----

Testcase #2: Can Clarity Analyze An HBM Interposer Design?

Many of our products require analysis of advanced packaging like silicon
interposers, 3DIC, etc.  This particular test case is an HBM interposer.
It is challenging due to its large aspect ratio.  This test case was:

   - 0.276mm x 5.288mm (large aspect ratio)
   - 3 RDL layers
   - 48 signal nets
   - 4.6 million mesh elements
   - 200 ports

This design stressed the capacity limits of HFSS.  It runs but just takes
a very long time if we select many signals to analyze.  Often, we find
that we need to massage the database by hand to get HFSS to recognize the
silicon design database -- meaning we're modifying its layout geometry,
merging layers, etc. -- by hand.

While hand-tweaking is not terribly complex, it's a time consuming step
that can introduce unexpected errors in the FEM analysis.  We've sometimes
spent a few days checking and double-checking our assumptions for accuracy
when a new design comes in just to make sure we're not adding errors to
our analysis.


Unlike HFSS, most of the time Clarity *usually* understands all those IC
layers and runs the raw data out-of-the-box without any modification.  

However, after importing this HBM interposer design, Clarity *struggled* to
generate the initial mesh for analysis.  We've seen this in a few cases
with Clarity where generating the initial mesh required some manual help
from additional cmd options.  In this case we had to run a few experiments
to find the following cmds:

  1. "Equivalent Dielectric" and "Via Clustering" in IC layout
  2. "Simplify Polygons" in Edit->Shape
  3. "Model Cleanup" function

(Keep in mind that this in nothing like the hand modifying of the data that
we had to do to get HFSS to work -- this was more about us trying to find
the right instructions to feed Clarity to get it to work the best way.)


Anyway, once we found them, with these settings we had good starting point
and we were able use Clarity's adaptive mesh solver to best effect.

                  ANSS HFSS         CDNS Clarity       delta
                  ---------------   ---------------    -----------
    Performance:  74 hours 31 min   18 hours 11 mins   4.1X faster
  Mem footprint:  480 GB            82 GB              83% reduction

While we spent several hours experimenting different cmd options to find a
good starting point, Clarity overall was still able to produce accurate
results ~4x faster than HFSS, and with 17% of the memory.

WARNING: While both Clarity and HFSS matched for both ends of the frequency
sweep, there was a range in the center where their results differed.  But
because interposer wires are thin silicon, we can't do physical measurements
to see which tool was closer to being right in that divergent range.

        ----    ----    ----    ----    ----    ----    ----

Testcase #3: Can Clarity Analyze Designs That HFSS Can't?

Our last test case is memory module card.  It's a complex system of 8 memory
pkgs on a single PCB.  This is a very challenging design: 

   - PCB + package: 2L flip-chip packages, 4 on top and 4 on bottom
   - 13.6in x 6.5in
   - 40 ports
   - 58 million elements
   - 12 layer memory module

It has a very wide frequency range and must work under several operating
conditions.  The combo of IC layer stack-up and PCB layout forced us to
partition the design data and make approximations to *try* to get it
working in HFSS.  Even after several tries we've never been able to analyze
it in HFSS due to the large amounts of data.

In contrast, since Clarity understands all those IC layers and it has the
computational capacity, it ran this memory module without any approximation
for the package parts as well as the board on:

          - 4 machines, 32 CPU cores each machine
          - 8.0 GB of memory per CPU slot

with the total analysis time of: 27 hours 53 min

Later this Clarity EM analysis was confirmed to be within +/- 1% of the
physical measurements of the design. 

In short, Clarity is the first EM tool we've seen that can analyze/signoff
a massive 40 port/58 million element system design like this.   

        ----    ----    ----    ----    ----    ----    ----

We can confirm the "multi-CPU distributed processing" claims that Anirudh
makes are real from what we've seen first hand.

We're not getting rid of HFSS in our house.  We need it for legacy designs
plus we have customers who use HFSS, so we have to match them.  But for new
designs, we are encouraging our engineers to instead use Clarity because
it gives them accurate EM results 4x to 5x faster than HFSS while using less
compute CPU's and 85% less memory.

And over time, we want to see how Anirudh's R&D handles the meshing large 3D
structures problem as his Clarity code matures.

    - [ Quagmire, from Family Guy ]

        ----    ----    ----    ----    ----    ----    ----
Previous Article
Cadence Wins Gold Stevie® Awards in the 2020 International Business Awards® Program
Cadence Wins Gold Stevie® Awards in the 2020 International Business Awards® Program

Cadence Wins Gold Stevie® Awards in the 2020 International Business Awards® Program Awards Include Most Inn...

Next Article
6 Clarity vs. HFSS Benchmarks
6 Clarity vs. HFSS Benchmarks

A year ago -- while everyone else missed it -- I *scooped* how Anirudh announcing his Clarity 3D solver at ...