Big changes come with smaller features
05/01/2003
 |
Momentous changes are ahead for the structure and business models of the semiconductor and allied industries, suggested several speakers and panel discussions at the SPIE Microlithography Symposium recently. A wide array of new technologies are emerging to enable continued shrinks in circuit features that will help keep cost/function steadily declining, thus helping to expand total electronics markets. But economic factors are proving to be an even more potent driver in bringing change to the industry. Let's look at some of the rising pressures.
While foundries are on the rise, they face problems with the exponentially rising costs for mask sets. In a 300mm fab, many runs for specialized system-on-a-chip designs may only involve a few wafers, which makes million-dollar-plus mask sets too costly. While 300mm fabs offer great potential cost savings, they must operate at high capacity utilization rates to cover huge amortization costs. Further, demand is highest for the most advanced processes, which may require tool upgrades before existing process tools have been paid off.
Costs for mask sets appear to be moving not just to the million-dollar level, but on to $6 million and more, according to a panel discussion at SPIE. Some suggest this could doom ASICs, because their runs are too short to pay for such costly masks. New "all purpose" superchips that can be interconnected and programmed for different applications, along the lines of gate arrays of a past era, may be needed to build up the numbers of wafers/mask set.
Pushing existing exposure tools steadily deeper into the subwavelength realm using resolution enhancement techniques is part of the problem. This is leading to steeply rising mask complexity. While the industry struggles to add model-based OPC to the design automation process, circuit designs need prototyping to find out if they will work, a costly proposition. Neal Callan of LSI Logic revealed that because of defects or errors, more than 70% of reticles had to be "re-spun" to correct mask designs. With so many functions being combined on a single chip, maskmakers aren't getting the benefit of the learning curve they had in the past when perhaps a dozen chips might be needed to provide the same functionality. Further, chipmakers like Intel and IBM are using their own mask shops to help improve the communication between designers, maskmakers, and process engineers. This loss of big customers is making it even tougher on the commercial mask industry as it faces rising R&D requirements.
The industry needs to look at Moore's Law in a different way, suggested Chris Mack of KLA-Tencor. It is really a classic learning curve, he said, so rather than plotting feature size versus time, we should look at a plot versus some measure of learning experience, such as total silicon area. "Volume drives progress, not time," Mack said. The key factor, he pointed out, is that cost/cm2 of processed silicon has remained roughly equal throughout the history of the industry. It is essential to remain on that cost curve to enable more applications requiring high volumes of chips, which provide the learning needed to continue moving forward. Die size isn't shrinking any more, and Mack believes the 300mm transition will be the last move to a larger wafer size. That means that to keep lowering costs, it is critical to maintain very high yields, and achieve fast ramps to yield. Meanwhile, innovation will be essential to create the new applications requiring large volumes of chips.
One reason it's getting so much tougher to maintain the momentum is that new lithography systems are being required at a faster rate than in the past, and this is not giving enough time to build the infrastructure needed for them. The 193nm tools were ready before good resists were available, for example, and the problems will be even tougher for 157nm and EUV.
It's a miracle we've gotten this far. We suspect there are more miracles coming, but the industry will need to find ways to realign itself to take advantage of them.
Robert Haavind
Editor in Chief