In mid-January 2024, the semiconductor EDA giant Synopsys announced the acquisition of the
software firm, Ansys, for $35 billion. Ansys produces software for
Computer-Aided Engineering or CAE. They have a fascinating history. But why buy it? There is a trend
going on here and it involves the new sexy technical trend of the day - chiplets. In this rather wide-spanning video, we look at the
intersection of design and computer simulation. ## Beginnings In the 1960s, John Swanson was working at the
Westinghouse Astronuclear Lab in Pittsburgh. He went there to work on nuclear-fueled rockets
with the goal of going to Mars - the NERVA nuclear rockets. His work involved doing
stress analysis on various components. To do this, Swanson had to model how a structure
might function under a particular load. However trying to mathematically model a big
complicated design was challenging. But if you divide that structure
down into smaller pieces, those smaller pieces are easier to
predict. Tie those pieces together and you get a far better idea of
how the larger structure will act. This is what we now call Finite
Element Analysis, or FEA. In FEA, we break something like a car design
down into millions of little shapes and use that to predict how that car
design might react if it hits a wall. In those days, the FEA equations had to be
worked out by hand using desk calculators, which was long, complicated, and error-prone. So Swanson learned how to code and
wrote a few programs to do this FEA. ## Ansys By the late 1960s, it became clear that
the nuclear rocket was not going to Mars. So Swanson left, moved back in with
his parents, and founded the company "Swanson Analysis Systems" to develop
the Ansys simulation software suite. Westinghouse brought Swanson back as a
consultant, which helped Swanson keep food on the table while he coded up Ansys on
a time-shared mainframe owned by US Steel. Ansys - Swanson went with the name
because the lawyers said it didn’t violate anyone’s trademark - was finished in
1970, and Westinghouse was the first user. Over time, Swanson and his small team ported the software to minicomputers like
the VAX and added crucial features. In 1994, Swanson sold majority control of his
company - which had annual revenues of about $30 million - to the private equity
firm TA Associates and stepped down. They took the company public two years
later, changing its name to that of the flagship product. Swanson retired soon after
that and has since focused on philanthropy. ## Design & Simulation Ansys’ FEA tools have since evolved to
embrace a variety of CAE sub-categories. Broadly speaking, CAE is about using
sophisticated algorithms to simulate, validate, and solve problems
in engineering projects. The name is similar to Computer-Aided Design,
or CAD, but they are not the same. CAD is for producing the design while CAE is for testing how
that design performs under certain conditions. This is helpful because engineers can now
diagnose and troubleshoot problems in a design without having to do as much
real-world testing. Since it can cost millions of dollars to build a modern plane or
car prototype - and certain conditions cannot be easily replicated in a wind tunnel
- this saves a lot of time and money. The rise of more powerful computers - including
those within the cloud - has made it more feasible for users to simulate increasingly
complicated designs inside a computer. ## Computational Fluid Dynamics FAE is pretty interesting. And it would be fun to use
it to simulate what happens when an integrated circuit hits a
brick wall at relativistic speeds. But in a semiconductor industrial context,
we want to look at a different major CAE sub-category - Computational
Fluid Dynamics, or CFD. CFD programs help solve the complicated nonlinear
equations that govern the motion of fluids. Brace yourselves for some mind-blowing
stuff but fluids are not like solids. Fluids like air or water react to shear stresses
differently than solids because their shapes deform without losing volume. So these fluid
dynamics equations are extremely intense. Any calculator - human or otherwise -
must simultaneously maintain multiple factors like conservation of mass
and momentum in many directions. CFD applies to all sorts of real world
engineering situations, but is particularly big in heat transfer analysis. Engineers can use
it to track heat sources in a structure, how that heat conducts or radiates throughout, and how we
can carry away that heat using a fluid like air. ## Brief CFD Software History Considering this, we have long sought
ways to model these fluid movements. In 1922, the English scientist and
mathematician Lewis Fry Richardson published the book "Weather Prediction by
Numerical Process". It radically proposed a set of differential equations
for predicting the weather. His book also fantasized about the
possibility of a "forecast factory". In it, you would have many dozens of
"computers" - human computers in this case - calculating these equations
for different parts of a world map. That of course was a fantasy,
but fun nevertheless. By 1940, the available analytical models for modeling
fluid dynamics were fragmented and in disarray. Prior models based on older principles
proposed by the mathematicians Leonard Euler and LaGrange were elegant, but did not
match up with what was happening in reality. This situation changed thanks to the urgency of
war. In the waning years of the Manhattan Project, it became crucial for the Americans to track the
behavior of shock waves and how they propagated through things. Research continued after the
War ended and hydrogen bomb development began. In 1950, Robert Richtmyer and John von
Neumann published a paper introducing the concept of "artificial viscosity". This
simplified the equations for computers while maintaining fidelity to the basic physics. Over the decade after that, a team at Los
Alamos led by Frank Harlow contributed several foundational CFD algorithms
like Particle-in-Cell or Fluid-in-Cell. These works together enable the foundations of CFD as we know it today. Professor John
Chu of Columbia coined the term in 1967 in addition to contributing crucial
differential equations for the discipline. CFD programs have since evolved to reproduce and
visualize all kinds of thermal dynamics. This is helpful for air conditioning systems, vents,
and - interestingly enough, semiconductors. ## 3D-IC In previous videos, I have talked a bit about
the growing trend in advanced packaging. One of the major examples of this is AMD's
flagship AI accelerator chip - the MI300. This chip has been built up to be some kind
of a Nvidia-slayer amongst AMD enthusiasts. But whatever you might think about AMD and Nvidia,
it is a magnificent piece of silicon engineering. I quote Dylan from SemiAnalysis here.
The MI300 is the most incredible form of advanced packaging in the world. Over a
hundred pieces of silicon stuck together in three layers on top of an interposer - which
is kind of like a PCB for silicon chips. These silicon pieces are connected
using Through Silicon Vias, which are holes drilled into the die. We can then
run copper interconnects through these vias. Advanced packaging techniques like these allow
us to produce larger, more powerful pieces of silicon without majorly hurting yields.
Certain compute chiplets can be made using advanced nodes while other less advanced stuff
like I/O can be done in trailing edge fabs. ## Heat Problems One of the big issues however
with these vertically stacked semiconductor packages and others like it is heat. Chips generate heat. In a traditional 2D
monolithic IC, we know how to deal with this heat. All of the silicon pieces are
accessible so heat generated by the chip can be dissipated out either through the silicon
substrate on the bottom or a heat sink on the top. But when you stack more than maybe,
two pieces of silicon together, we can't just throw a big old
heat sink on top and call it a day. That is because we no longer
have access to all of the chips. When you consider that the silicon pieces
do not all have the same heat profiles and that heat cannot travel very far, then
you end up with local hot spots. These higher temperatures hurt the chip's
performance and overall lifetime. I am compelled to inform you of the fable
of the Asian giant hornet. If it ventures too deep into a Japanese beehive, the bees
tightly surround the hornet and beat their wings to high temperatures. The poor hornet -
which only wants to eat - is cooked to death. Likewise for an advanced packaged IC.
Heat from silicon operations or beating bee wings can compromise performance. It
can fatigue the soldered interconnects, warp the silicon substrate, or
even crack the die entirely. ## Taking the Heat Considering the consequences, it is
important that chip designers build structures into their stacked
packages to deal with this. The industry has tried several things
- there is a whole part of the industry dedicated to heat control. We have silicon or
software-level measures to take. For instance, writing code to limit the chip's performance
so that we don't overheat the package. There are also packaging-level things we can do
like immersing the whole chip in a heat dispersing fluid or even creating silicon "fins" sticking
out from the interposers to shuttle away the heat. There is strong focus on adapting existing
technologies to minimize risk of failure. One example is the solder ball interface
between the dies. Changing the filling in between the balls can greatly improve the
heat distribution profile of the stack. ## Heat Simulations Regardless of the methodologies
employed to deal with the heat, we are going to want to simulate
it out to see if it works. A wafer takes about three to five months to
produce and that does not include packaging. So it could take months to get a complete package
to measure, test, and validate its heat profile. But we don’t have that much time to wait.
Advanced chip design takes far too long as it is. So being able to simulate the thermal
profiles of a particular design inside a computer without needing to go and build a prototype
to test in real life is hugely beneficial. That means CFD. For instance,
calculating heat transfers from between two components within a package
requires us to know the local fluid temperature - the temperature of the parts
in contact. But as we established earlier, CFD is extremely difficult to model due to
the complicated interplay of various factors. Such programs need to produce accurate
results that match with real world experimentation while also being intuitive to
use since designers are often time-pressured. For this reason, CFD programs for thermal
simulations of electronic packages are very mature - around for over twenty years.
So semiconductor tool and IP ecosystem providers like TSMC prefer to partner with CFD
makers like Ansys rather than make their own. In 2021, Ansys announced a partnership
with TSMC to deliver a thermal analysis solution for 3DFabric. 3DFabric is TSMC's
umbrella of advanced packaging technologies. Import a package from an EDA software
and you can run electrical, thermal, and structural simulations on it. Getting this
stamp of approval as well as others from the world's biggest foundry is a compelling sign of
Ansys' growing presence in semiconductor design. It does make me wonder if we should allow
Synopsys to actually buy them. On the surface, the businesses seem quite separate. But
Ansys is growing in the industry. And semiconductor design and simulation
are soon to collide - especially as generative AI makes its way into
the design flow down the line. Simulations can be a way to validate those
designs - like how we use a Python runtime to validate ChatGPT's written code.
Just a random thought. Maybe I'm crazy. ## Conclusion One more thing. I do think that Synopsys
being willing to pay $35 billion does say that heterogeneous integration - i.e.
the chiplet approach - is here to stay. Expensive AI chips like the MI300 are
only the beginning, the leading edge of the sword. I get the feeling that
we are going to be seeing the MI300's approach trickle out to almost every part of the
semiconductor space except maybe the smartphone. Considering all the trailing edge
semiconductor fab capacity being built around the world right now,
users will want to mix-and-match process nodes and then just slap it
all together with advanced packaging. This means more complex advanced packages. And
that is going to present a lot of challenges because heat isn't the only major challenge
for packaging designers to overcome. Another one that comes to mind, for instance, is
capacitance - which arguably might be harder. This all in turn incentivizes us to
produce better tools for designing and manufacturing such advanced package designs.
Synopsys is setting out on that journey, and is putting down $35 billion to back it.
