Commentary: DFM difficult, but no 'deathwatch'
As we enter 2007, design for manufacturability (DFM) is under siege. After the frenzied promotion of the past few years, most DFM start-ups that have been funded since 2000 are struggling for product acceptance by customers and meaningful revenues for their investors. Last October, John Cooley posted a provocative article on his DeepChip web site now known simply as the DFM deathwatch list. For the 20+ DFM start-ups, the situation is indeed difficult, but is not in the dire condition proclaimed by Cooley.
DFM — where are the difficulties?
As semiconductor processes continue Moore's Law advancement, DFM issues will only become more and more severe with each process node at 65 nm, 45 nm, 32 nm and beyond. Design-induced systematic yield loss will become more dominant than random yield loss at advanced processes. So far, most DFM tool development efforts have been focused on injecting process information into the design flow with IC designers as target users.
There are a few major obstacles for adoption of DFM. Given that design closure for advanced technology nodes is already very hard to attain, users will not consider adding time-consuming steps for DFM unless they are absolutely necessary. Today most IC designers have limited interest in, or are reluctant to learn, about IC process and fabrication.
However, they are increasingly requiring their EDA tools to integrate necessary DFM functions, but are also demanding no big changes to the existing design flow. This situation puts DFM start-ups at a disadvantageous position in competing with major EDA venders. It also forces DFM start-ups to spend substantial amounts of resources to build various interfaces for major EDA vendors' design flows. This increased integration effort on top of DFM R&D costs puts additional resource strains on all DFM start-ups.
Semiconductor manufacturers, especially foundries, have thus far not made DFM tool usage mandatory in tapeout signoff for some good reasons. For most advanced technology nodes, stable process and optical proximity correction (OPC) data are usually not available yet to support DFM at early design stage. And then there is the general reluctance on the part of foundries/IDMs in turning over process OPC data to DFM tool developers due to proprietary considerations.
Loading comments... Write a comment