Elevating Analysis and Optimization Across Disciplines to New Productivity Levels Using AI/ML

During the launch preparation of our latest set of AI/ML-enabled offerings, I heard a term that sounded vaguely familiar—"multidisciplinary design analysis and optimization," or MDAO. In my role leading the vertical and horizontal solution aspects here at Cadence, I often meet diverse groups of engineering disciplines spanning hardware, software, electronics, mechanics, and many others. It turns out that thanks to the advances in AI/ML and computing capabilities, I am lucky enough to see again significant advances in how these disciplines interact.

In a roundtable from last year on "How Heterogeneous ICs Are Reshaping Design Teams," I mentioned when asked about the effect of the interaction of multiple disciplines that "we used to joke that when you brought together hardware and software people in a company, they would sometimes introduce themselves and trade business cards. There's a whole new dimension to this with people on the mechanical side. There are experts in EMI, too. We always talked about a system architect who knows a little bit about everything but nothing about the full details. That person is becoming the moderator between the hardware, software, and other disciplines. More of that is happening because the tools are now at a higher level of complexity to put this all together."

This effect is still valid when doing interdisciplinary work. I just returned from NI Connect in Austin, where we announced linkages from Cadence tools to data management tools, spanning the disciplined design and test. It can take some time to "translate" between lingos at the beginning of a meeting with multidisciplinary participants.

When I arrived in the US in the late ‘90s, I led—fresh off the boat—the technical marketing for the Felix HW/SW Co-Design Initiative. At the time, led by EDA visionary Alberto Sangiovanni-Vincentelli, we were trying to revolutionize the design flow with what Alberto called "function/architecture" co-design. Describe your function in an abstract system model (think UML++), define your target architecture, and map which functions go into which architecture component. Voila! It's a "simple matter of implementation" from there. In retrospect, we got surprisingly far. At DAC 2002, we demonstrated a flow in which decisions at the functional and architecture level impacted the chip layout, with communication automatically being re-mapped depending on whether functions were in hardware or software. While we were about two decades too early, give or take, Alberto's function/architecture vision has fundamentally impacted many aspects of today's design flows.

It's 2022. Meet AI/ML-driven MDAO.

Many disciplines use the term MDAO. The core principles are the same. It is a methodology that enables the analysis and optimization of a complete system by explicitly considering significant interactions and synergies between disciplines. We assessed aspects of power, area, and performance at the architecture level during the Felix initiative, with different functions and architectures defining the design space. In another area of system design, design teams may look at a communication channel from a signal source through the transmitter's IC package, the PCB board, and the receiver's package to the actual signal recipient. System criteria will include return loss, insertion loss, cross talk, and isolation. Multidisciplinary aspects may consist of channel and system requirements like PVT corners, package types, ODT, equalization, suppliers, jitter, and data rates on the transmitter and receiver sides. In the communication channel itself, geometric variables are critical. These can include line width, spacing, length, stack up, pad size, anti-pad geometry, drill size, stub length, and ground via pitches.

Get my drift? These are many variables for even the best engineer or team to keep track of manually, not considering the additional complexity of various disciplines that do not speak the same language.

One of the many challenges in the Felix initiative was that users had to manually set up the experiments for a sweep of simulations. The same is true for classic MDAO. Defining the limits for all parameters and scanning the design space is possible but results in very computationally expensive simulation setups that examine the entire design space. It's like planning the path from your house to the airport but creating a list of all possible street combinations and virtually driving them.

Enter AI/ML combined with MDAO.

It's exciting to see that with the combination of AI/ML and MDAO, we now have a way to get to the right design much faster with many fewer simulations. In contrast to simply scanning the entire design space, machine learning models enhance classic heuristics that guide a design implementation towards the optimal design. From a setup of the optimization in which the user defines design variable ranges and specific objective functions, AI/ML maps out the appropriate design space and runs the simulations. It constantly analyzes the simulations, updating design variables and calculating objective functions and constraints until a stop criterion is reached, like the stop criterion of -35dB in the example design space exploration below.

We call this process “in-design MDAO” and have seen some customers experience reductions in time spent on optimizing transmission line performance from hours to minutes. Others were able to determine quickly and efficiently the optimum return and insertion loss, as well as TDR waveforms.

There is still a lot more to come here, but the combination of AI/ML and MDAO rings in a bright new future of productivity enhancements for system design.