Improving Design Robustness with PrimeShield: A Discussion with Li Ding

Hi Li. First, thank you for taking the time today. Please outline how you see the variability in today’s design methodologies evolving? What new challenges do you see at advanced nodes?

Li Ding: As technology scales down to 5-nanometer and below, we are now down to 3- and 2-nanometer designs, so the variability with respect to nominal values becomes much bigger. Coping with the variability is becoming a much more important topic, so there is a strong need to deal with the variability in all dimensions, including process. One example is high-sigma requirements for automotive applications. Another example, in the dimension of voltage, is the time impact with respect to IR drop. Finally, the dimension of temperature is particularly important for 3DIC. 

So, if designers do not do anything and stay with their traditional timing margin-based approach for signoff safety, they are really leaving a lot of PPA (power, performance, area) on the table. 

That is indeed challenging! Can you please share how PrimeShield addresses these challenges and what role it plays in building confidence in the design flow?

Li Ding: PrimeShield performs design variation analysis to assess overall design robustness against process variation. So, this new product, empowered by a unique, fast STA Monte Carlo engine, can accurately model the correlation effect when there are multiple timing paths going through common cells. It also answers designers’ questions on which path-level sigma is required to reach the desired design-level sigma target. In addition, it can identify common weak cells to guide ECO fixing to reach design-level sigma targets and improve design Fmax distribution. 

Can you further elaborate on the recent improvements to PrimeShield for handling the ever-increasing performance and productivity requirements? Any machine-learning technologies? 

Li Ding: As most designers are aware, Monte Carlo SPICE simulations are often used to get highly accurate path timing in the presence of large process variation. In automotive applications, it is often needed to close timing at high-sigma levels, for example, 5.2 sigma (which is equivalent to 0.1 PPM defective parts per million) or 6 sigma. Monte Carlo SPICE simulation of a timing path at high sigma requires years of CPU hours, which is not practical for production use. PrimeShield features modern technology that uses an innovative machine-learning engine to significantly speed up simulation runtime and, at the same time, maintain true SPICE accuracy.  Using fast Monte Carlo path simulation in a typical 3-sigma signoff, we have seen over 100x speedup compared to Monte Carlo SPICE simulation. We have also seen over 10,000x speedup in automotive high-sigma signoff, reducing CPU time needed from years to mins-hours range. 

Overall, PrimeShield seems very well positioned to address the challenges. What are the future directions of PrimeShield and something that users can look forward to?

Li Ding: Traditionally, designers are using POCV (parametric on-chip variation) for local process variation and using STA corners to cover global variation. The latter is limiting because, for example, one cannot assume all metal layers are at the Cworst corner at the same time. This is often known as the global skew effect. We will be rolling out native support for such global skew effects including Vt skew, P/N skew, and interconnect skew to reduce pessimism compared to timing margin-based solution. Second, we are also developing solutions to enable designers to rapidly perform what-if analysis with respect to PVT (process-voltage-temperature) for timing robustness analysis and for technology option assessment for best PPA. 

Thank you for the overview, Li. It was an illuminating discussion. Any closing statements for our readers?

Li Ding: Thank you. We are extremely excited about PrimeShield’s existing features and roadmap capabilities. We look forward to helping designers to improve their advanced node design PPA and robustness.