Failure is a word no one ever wants to hear in the advanced packaging industry. When AI chips have made it as far as the packaging stage, the last thing anyone wants is for them to not work after millions of dollars and hundreds of hours have been poured into them.
Fortunately, late-stage failures can often be prevented through a process called predictive analysis, and it’s exactly what it sounds like. It’s the process of analyzing existing data to make predictions about how the chip package will perform in the future. Often, the biggest reason why packages fail so late in the game is because of thermal challenges. Each component in a package generates so much heat that if what’s causing the failure isn’t addressed adequately, it could exponentially reduce the component’s lifespan.
To learn more about how predictive analysis helps prevent failures, I spoke with Andras Vass-Varnai, a thermal management and reliability engineer at Siemens EDA. He explained that as technologies advance, thermal management becomes increasingly complex, and predictive analysis must be completed as early as possible.
Common Reasons for Thermo-Mechanical Failures
Vass-Varnai explained that heat and thermo-mechanical stress go hand-in-hand. While individual components naturally generate heat, the industry’s push to add more transistors in less space isn’t helping thermal issues either. Vertical stacking is another culprit, and thermal challenges are further exacerbated when dies are less than 100µm thick. When the dies are extremely thin, this can limit their lateral heat spreading capabilities, making hot spots even hotter.
Although there are many advanced packaging designs for AI applications, they each have their own thermal downfalls. For instance, through silicon vias (TSVs) create complicated heat distribution patterns, micro-bump arrays hinder local thermal resistance (the heat flow within a specific area of the package), and interposer designs often require collaboration between design and analysis tools. In addition, hybrid bonding introduces new thermal and mechanical considerations that didn’t exist in traditional packaging.
Rising temperatures also create thermo-mechanical stress due to CTE mismatch between different materials. This is when materials expand and contract differently when subjected to temperature changes. When CTE mismatch occurs, it can lead to warpage and reliability challenges.
Selecting the right materials is one potential solution, but unfortunately, there’s only so much that can be done on the materials side. However, designing the package so that the individual components are strategically arranged can help mitigate thermal concerns. In addition, Vass-Varnai emphasized the importance of continually checking to make sure the heat is within an acceptable range during the design process.
“If you fail to do any of this, you’ll always be taking two steps back, wasting money and time problem solving,” he said.
Technically, predictive analysis can be leveraged at any point during the design phase to evaluate the potential performance of the package. According to Vass-Varnai, as long as predictive analysis is done before the tapeout phase, concerns about thermal can still be raised. For context, tapeout is the last step of the design process before the actual production process begins. This is when the final design of the integrated circuit (IC) is sent to the foundry for fabrication.
However, redesigning the package costs time and money, making it even more critical to complete predictive analysis as early as possible.
How Does Siemens Handle Predictive Analysis?
Siemens approaches predictive analysis by using multiple tools simultaneously to create engineering workflows. For example, its Calibre 3DThermal software does thermal property mapping and detailed model creation. Its mPower software can inform thermal loading for both average and peak conditions, and its Solido software (when used with Calibre 3DThermal), can increase SPICE simulation accuracy.
SPICE stands for simulation program with integrated circuit emphasis. It means that designers can simulate circuit performance without needing to build them, and it can demonstrate the impact of temperature and stress on circuit behavior. In other words, it’s a form of predictive analysis.
In addition, Calibre 3DThermal and Innovator3D can be used together for package architecture and assembly planning and design. These examples barely scratch the surface of Siemens’ suite of tools, but Figure 1 provides more context.
While Siemens isn’t directly using AI for its complex simulations, the company is working toward solutions that’ll allow for it eventually. Even though the industry at-large knows AI can one day aid in this area, having enough structured data to train AI models is a major hurdle preventing immediate adoption.
“Everybody’s working on it, but not many companies have the data sets,” said Vass-Varnai.
Lastly, although thermal is a major consideration for package design, it’s important to note that it’s still only one piece of the puzzle. Today’s AI ICs are more complex than ever, and this is changing the nature of how the packages are designed and manufactured across the entire industry.
Learn more about Siemens and its suite of software solutions.
