Industry Pressure Grows For Simulating Systems Of Systems

Most complex systems are designed in a top-down manner, but as the amount of electronic content in those systems increases, so does the pressure on the chip industry to provide high-level models and simulation capabilities. Those models either do not exist today, or they exist in isolation.

No matter how capable a model or simulator, there never will be one that can do it all. In some cases, that is because of scale or abstraction, or even the underlying physics. While the semiconductor industry has dealt with mixed-signal designs, whose analysis may require both discrete time and continuous time domains, until recently it has not had to deal with the added complexities that are required for chiplets and advanced packaging. Now, thermal and mechanical analysis cannot be avoided, and they are tightly coupled to functionality, power, airflow, and other many other effects.

This is nothing new for the systems, aeronautical and military industries, where the semiconductor content is just a small piece of what must be analyzed. More recently, the complexity of automotive systems has put increasing burdens on them, and by connection the semiconductor industry, to provide more complete models and simulators that can be integrated into much larger environments.

The military has been talking about the needs for federated simulation for the past 20 years. A paper titled, Federated Simulations For Systems of Systems Integration, states that “federated simulations must be as agile, robust, and interactive as the operational environment they support. Large single models that enable analysis of systems of systems are not effective because no single modeling effort can account for all of the complexities. Instead, modeling and analysis must be done in a distributed and parallel fashion, mirroring the development of training, acquisition, and fielding initiatives. This leads to a number of subsystem models that effectively analyze different domains.”

Pressure is building on the semiconductor industry, and Accellera has started to react to that need. In August, the standards group announced the formation of a Federated Simulation Standard Proposed Working Group (FSS PWG) to focus on the creation of a distributed and orchestrated multi-domain simulation framework. Earlier this month, it held its first meeting in Toulouse, France.

“A group of Accellera members have been part of a larger exploratory team looking at cross-industry collaboration to exchange knowledge and best practices,” according to Lu Dai, chair of Accellera. “The team has been investigating the coordination of efforts in standards development and the integration of simulation technologies. The objective of the PWG is to identify industry interest and consolidate the requirements to drive standardization and development of an open API and federated simulation ecosystem.”

EDA companies are working on this problem. “There are companies that have developed solutions within an industry segment for a specific use case, which might focus on a subsystem but not the whole system, and definitely not on the systems of systems,” says Martin Barnasconi, chair for the Accellera PWG. “As an industry, we have been able to address the local domain-specific issues. The systems companies are looking for a solution in terms of how to integrate, model, or simulate that, while leveraging models or technology from neighboring industries. They want to leverage stuff from the semiconductor domain, or for us to learn more from the aerospace domain. Standardization bodies happen to be in all those individual industry segments or domains.”

Part of the problem is terminology. “I was talking to a bunch of people from the aeronautical industry, and they wanted a model of an M5 or M7 Arm processor,” says Mark Burton, vice chair of the PWG. “Apparently, it was causing a lot of problems because they hadn’t got one. That surprised me because I have hundreds of those, and asked what flavor and color they would like? It transpires that they were all thinking in terms of an interface that’s called SMP2, which has no relation at all to TLM2, apart from the fact it’s absolutely the same.”

Fig. 1: Potential architecture of federated simulation. Source: Accellera

“The big problem is that we’ve all built our silos around how we connect things together in automotive or how we connect things together in aerospace and defense, or in mobile,” adds Burton. “There are lots of standards to connect things together, but those standards don’t talk to each other, and people who are using those standards don’t talk to each other. Some of those standards are good, and some of those standards are bad.”

Others agree. “That’s been a challenge for us as we started to address layers that are higher on the food chain, instead of just looking at the component-level physics and subsystems,” says Shawn Carpenter, program director at Ansys. “It’s been an education for us, and we’re constantly having to retrain our field engineers to make sure they’re very clear on the terminology they use, especially as we start to address more systems-level organizations, which think very differently.”

That is a non-trivial problem. “The challenges always have been trying to get the right expertise, and how to get them to work together,” says Dave Rich, verification methodologist at Siemens EDA. “It’s hard to get people in a room to talk the same language. With the Accellera effort, we are in the learning phase to determine what we want to accomplish. There is the data collection of all these different models, and there’s the simulation and execution aspects, and integrating all this. And then there’s the post-execution, like functional coverage and requirements tracking that need to be collected.”

There are many ways to look at the problem. “They should look at it from an application standpoint, not from a framework standpoint,” says Deepak Shankar, vice president of technology at Mirabilis Design. “If I am doing a space mission, or a car, or a semiconductor chip, take everything that’s there and figure out the underlying technologies required and understand how they’re all connected. Within semiconductors, you have a separate power model, a separate functional model, and a separate performance model. At the system level, you need to integrate them. Below that, they are separate. Is that part of federated simulation?”

It takes a village. “For something complex like autonomous driving, there’s a myriad of different disciplines and systems that are involved,” says Chris Mueth, business development, marketing, and technical specialist at Keysight. “To do a good job, you have to have engineers with specialization in each of those areas. You might have a person who’s good at structural dynamics, another person is an expert at thermal, another one’s good at digital electronics, and other one is good at RF electronics. You have all these disciplines, and they all have to work together.”

The industry loves to use the V model to explain all of this. “If you’re really going to make this work, models have to be universally consumable,” adds Mueth. “Today, you have a lot of different manufacturers, EDA players, who can create different types of models, but they’re not interoperable with other tools. There’s no EDA vendor who owns the roost. There are tools that specialize in things that aren’t covered by the big players, and within that workflow you have to interchange models. That’s something that is not easily done today.”

Progress within EDA
In recent years, the EDA industry has spread beyond its semiconductor roots, as an increasing number of factors play into the long-term reliability of chips. Companies, such as Ansys and Keysight have been expanding their roles into multi-physics problems, and Siemens purchased Mentor Graphics, bringing together a system of systems company with a semiconductor company. Cadence, meanwhile, has been investing in computational fluid dynamics. All of the major EDA companies are aware of the problem and opportunity.

Ansys calls this digital mission engineering. “Think about a collection of 5G base stations,” says Carpenter. “I have a very accurate representation of a city that includes material property characteristics of buildings and streets and things that I populate along the street. And I might have several mobile carriers, mobile phones, radio systems embedded into cars. I may have a quadcopter that I want to fly at very low altitudes through the city canyon with obstacle detection and collision avoidance radar systems on board. Will my channel drop at some point because I go into a shadowed area where I don’t get coverage? It’s a simulation where time is a first-class citizen, where you are orchestrating the motion. It involves kinetic physics, the motion of many platforms that either are fixed, or have some motion associated with them.”

Siemens currently is putting its emulator at the center of this capability. “Veloce System Interconnect is our proprietary attempt at bringing together these disparate systems,” says Neil Hand, director of strategy for design verification technology at Siemens Digital Industries Software. “It’s a tool that is built on open standards to link different simulations together, but it has to understand the simulation interfaces. It has DPI, FLIR, streams interface, and FMI. The need is definitely there in the systems domain. It is being driven by aerospace and automotive guys, because they are used to doing system-level modeling and they need to bring digital into those domains.”

High-level semiconductor models are required. “SystemC has models defined as loosely timed, cycle approximate, cycle accurate,” says Marc Serughetti, senior director of product line management at Synopsys. “What do those mean? How do I go about validating and verifying all the designs in the system knowing that they work together? All those systems are becoming software-driven. The key part we are focusing on is the electronics part of the system of systems. Ultimately, the important aspect is how you go about validating and verifying those models? What we’re seeing in the industry is that this means that all those different models need to be integrated more in the development process.”

Connecting the digital twin
Federated simulation and the digital twin need to come together. “The digital twin is where you have a virtual model — possibly a very accurate model — rooted in physics of a platform or a system, and then you set it in the context of a real operational environment,” says Ansys’ Carpenter.

This need is bringing together top-down design and the extend-right of lifecycle management. “If you talk about system of systems, software-driven type of products, the lifecycle of the model is increasing,” says Synopsys’ Serughetti. “Initially, the desire was to do something before silicon is available. Now, we are looking at those models as they continue their life, post-silicon, post-device, because you want to have the ability to accelerate the validation of that software, and the validation of the system, even when the system is in the field. And this is the other area of digital twins. How do I have a representation of my system that exists in a real context, and the ability to reproduce the problems that are happening in the real device, such that I can go back and figure out what happened. The system is so complex, and I need to figure out where this problem came from and how I go about fixing it. The lifecycle of the models, the lifecycle of these digital twins, has increased. And the quality of the model, the validation of those models, has become a very critical issue that needs to be addressed in that process.”

Simulation in the cloud
Federated simulation requires that either simulators and/or models be distributed. “The cloud is an enabler where different parties can bring their bits and pieces, either open or proprietary, into the federation,” says Accellera’s Barnasconi. “We need to see how we can connect to those types of movements in the cloud. For other application areas, you might not want to have a cloud, or you want to run things locally on a local processor. In terms of scalability, this will be a challenge.”

This is where the cloud fits in. “We have to make it simulation agnostic in a sense, because some customers use Python to model some subsystems, they use MATLAB, they use products from Keysight, they use stuff from National Instruments. They have all of these things floating around that model their different subsystems,” says Carpenter. “They want to connect those into a full-stack mission simulation. It’s a brave new world right now, which is using distributed computing and spreading things out in such a way that everything stays on track. You need to ensure coherency in the data that comes back, where it gets time-locked in the right location.”

One of the big advantages of the cloud is the ability to aggregate and sort the massive amounts of data in systems of systems. “The cloud piece is absolutely critical in this,” says Serughetti. “You can imagine a use case where you have somebody building a system of systems, and that could be a multi-die SoC or it could be a car. Somebody is integrating it, and you have to expose your models in the cloud because this is the place where you’re going to put your subsystem so we can simulate it. And this has to bring multiple people or multiple companies together. So the cloud and the collaboration that the cloud brings is critical.”

Still, making that work with the necessary speed will be difficult. “People think they can just throw a bunch of GPU cores at things and it will just keep speeding up indefinitely,” says Mueth. “That’s not true. Any process that you’re parallelizing is subject to Amdahl’s Law, which basically says that when you set up a compute process, there’s always a little bit of overhead associated with that.”

Even if the simulation overhead is zero, there are still communications latencies. “If you are running your model on your laptop, but the other models are distributed on the network, then you use the edge to the cloud to connect all of them,” says Mirabilis’ Shankar. “Now, the simulation latencies are going to be on a level of seconds, not trying to finish it in nanoseconds anymore.”

The problem statement has to change to make this possible. “Somebody has to be the master,” says Carpenter. “There’s got to be one control node, control thread, because something has to take care of making sure that everybody does their bookkeeping and aligns the data to the right place, and everyone gets tasked with the right data. Can I farm pieces of the simulation off to a series of cloud nodes and have all of this data chugging away in parallel, in their respective containers, and bring the data back? How do I accomplish that?”

The next step
Accellera held its kick-off meeting this week, encouraging attendees to present use cases, requirements, and expectations of the standard. “What are the use cases? What are the applications from the different industries? And what type of modeling approach should be developed?” asks Barnasconi. “It is always tricky with a bunch of engineers in the room who want to start thinking about nice technologies and languages and standards. But first, we need to understand what we would like to model or simulate or analyze, and why are we doing this in the first place.”

Editor’s Note
A future story will look at how model abstractions within a semiconductor flow are changing, with a desire to be able to generate models on demand.
Related Reading
Large-Scale Integration’s Future Depends On Modeling
The progeny of VLSI is 3D-IC and a range of innovative packaging, but all of it has to be modeled to be useful.
Shift Left, Extend Right, Stretch Sideways
Development flows are evolving as an increasing number of optimization factors become interlinked. Shift left is just one piece of it.