UMass Amherst’s Sandip Kundu moderated a Thursday afternoon panel at DAC entitled, ‘Designing on Advanced Process Nodes: How many re-spins should you plan for?’
In concert with his four panelists, Broadcom’s Ajat Hukkoo, Intel’s Ashu Bakhle, Samsung’s Hong Hao, and GlobalFoundries’ Luigi Capodieci, Kundu laid out qualitative motivations and quantitative guidelines for predicting how many re-spins can be expected when a design targets next-generation geometries.
Prof. Kundu began with an homage to the costs and challenges: “Chips are expensive to develop, the Spice models are expensive to develop, and the first-pass and second-pass models often are not working.”
It’s within this environment of uncertainty, Kundu said, that designers and their managers are having difficulty predicting how many re-spins will be needed to get things right, and thus budgets and schedules are equally unpredictable.
Ajat Hukkoo agreed: “At Broadcom, every time we migrated from one node to another, the partitioning [of the design] had to be re-evaluated for electrical considerations and costs.
