Optimizing for Thermoelastic in Package and PCB Design from the Outset

Signoff-quality design tools are mandatory for signoff, though not so good for early design exploration. Signoff demands high accuracy in design description and high precision in analysis. The first requirement can’t be met early in design where architecture decisions are still in flux, and the second depends on very detailed numerical solvers, too slow for fast what-if iteration. These constraints are particularly evident in multiphysics analysis and are a critical factor for AI systems both in chiplet-based packages and in PCB design. Here, today’s state of the art tools analyze physics behaviors such as heat diffusion and resulting mechanical stresses through differential equations evaluated over a mesh. The accuracy needed at this stage demands long run-times together with detailed implementation data and for those reasons is commonly run only in late-stage implementation.

A recent announcement from Vinci is especially interesting in this context. Vinci uses an AI-native computational architecture built around a Foundation Model for Physics to evaluate thermo-mechanical behavior. It can do this orders of magnitude faster (they claim) than existing signoff tools. In fairness, established platforms also offer AI acceleration, though I understand those methods to be wrappers around established signoff tools, exploring across a limited range of parameter options in a reinforcement learning flow. The Vinci approach is fundamentally different in at least a couple of ways.

How might this work?

I talked to Hardik Kabaria (CEO of Vinci) about the product. He is understandably reluctant to share too much detail, but he did offer a few clues. First this is a transformer-based technology, pre-trained sufficiently that it needs no fine-tuning in deployment. Second, inference is constrained by the underlying laws of physics and must respect fundamental physical invariants, for example conservation of energy. Third, no manual meshing is required; Vinci automates the geometry preparation and meshing that traditional simulation workflows require engineers to design and refine. And fourth, there aren’t separate models for thermal and mechanical; one model covers both domains.

I can’t write effectively about a technology without constructing a plausible mental model, so with the caveat that this is my own view and not an authorized Vinci description, here is my take. Pretraining with no need for fine-tuning makes sense because there is a limited range of materials used in semiconductor/multi-die and PCB design and a limited range of assembly options. Also within normal operating ranges for thermal and mechanical, states should vary relatively smoothly without wild swings and local instabilities. Both factors imply a bounded training set is possible, derived (I assume) from signoff quality data.

In inference, hallucination can be minimized by energy conservation constraints. I would imagine that boundary conditions also need to be specified in some manner to reflect energy flux out of the system through (unspecified) cooling mechanisms. Heating sources are an input to an inference and can be determined and injected externally or added by the user.

Back to input from Hardik; because Vinci can evaluate physical behavior so quickly, engineers can explore large design spaces by evaluating many variations of the same design. Vary heat sources reflecting different loads, component placements, cooling constraints, material choices, and so on. He asserts that Vinci can evaluate engineering problems exceeding a billion degrees of freedom in a couple of minutes, where a conventional multiphysics analysis (even AI-assisted) would take hours, days, first to build a mesh, then to numerically solve over that mesh, then to iterate for reinforcement learning.

Takeaways

This is a huge difference in runtime, making it practical to use Vinci for early architecture planning, for multi-die chiplet partitioning and placement on an interposer and for component placement on a PCB. Also to plan cooling options – forced air perhaps, more likely these days through liquid cooling. Allowing for as many iterations as needed to optimize thermal and mechanical design up-front rather than having to scramble to fix the design or needing to document limitations in the hardware errata.

An important advantage per Hardik is that Vinci’s physics analysis is visible. A designer can see exactly what it predicts and how this varies with parameter changes, building confidence in the reliability of predictions (see figure above).

John Bruggerman (CMO, Vinci) added that they already have multiple engagements underway, supported by paid licenses. Clearly customers see important complementary value to existing platforms. John doesn’t see this technology displacing those established technologies in the near term, as those will remain important for signoff. Overtime, as Vinci training becomes more refined, who knows?

You can learn more about Vinci HERE.

Register for the upcoming Vinci webinar HERE.

Also Read:

CEO Interview with Dr. Hardik Kabaria of Vinci