Fitting many roads on one map
01/01/2008
While laudable for establishing consensus on the current state-of-the-art in semiconductor manufacturing, the International Technology Roadmap for Semiconductors (ITRS) has not been shown to be overly accurate as a predictor of the future. The chapters on Lithography, Interconnects, and “Frontend” transistor formation have all failed at times to predict technology developments even a few years out. To be sure, the inherent contradictions of technology risk and manufacturing conservatisms result in extreme market dynamics. We all want to take the most conservative risk. We all want predictable innovation. We all want carefully managed revolutions.
The philosophy of the ITRS, the latest version of which was just completed, may be expressed by the color coding of the cells in the specification tables. The “yellow” bricks represent specifications for which solutions are known from a basic scientific perspective, but cost/commercial challenges remain in terms of engineering. The “red bricks” represent specification for which no known solutions exist even from a pure science perspective. Because scientists and engineers have been clever and productive in the past, most red bricks have evolved into yellow bricks and then into colorless simple specifications over time. Thus, a sort of casual complacency has developed in the industry whereby people tend to assume that any defined technical problem can and will be solved in a few years if enough people apply themselves to it.
Often this dynamic plays out as we’d like-as with the specification for ultra-shallow junctions being met with novel implanting/annealing technologies-and the industry “stays on the roadmap.” But sometimes this doesn’t play out.
The story of lithography technology evolution still tells of this failed dynamic better than most. Considering the wavelength of illumination and the minimum dimension desired, the ITRS dictated that the stepper wavelength should decrease beyond 193nm to 157nm as late as 2003. Rumors of cumulative industry spending of over $500M to try to develop 157nm seemingly could have been avoided if materials scientists had been consulted about the properties of the calcium fluoride crystals needed to replace simple silica glass for lens elements.
Once it became clear that 157nm would not work, everyone had to scramble to find alternative paths. Double exposures, phase-shifting masks, hard masks, high-NA immersion, and extreme off-axis illumination have been used to push 193nm steppers down to 45nm production. All these litho enhancements were known when 157nm was put on the ITRS, yet all were considered as more expensive than just specifying a tighter wavelength stepper and so were not the first choice.
With hindsight, it’s always easy to see what should have been done, though the inherent complexity in the moment makes decisions tough. The level of complexity in nanometer-era CMOS manufacturing in 2008 is truly staggering by practically any measure:
- A 65nm-node IC can have one billion transistors on a single chip, and the corresponding number of polygons that must be modeled in design is one trillion.
- Fabs can cost more than $1B, and individual pieces of equipment can cost at least $10M.
- Design costs may be more than $10M, and the design may yield zero functional chips. It’s difficult to predict how many “spins” around the design-loop may be needed.
- Mask sets can cost above $1M, and some complex masks may be good for only 100 wafers.
…and all of the numbers above are very conservative.
Technology decisions must be made, yet decision-makers lack all the information needed to assess the options of different paths. We expect the ITRS to provide this information, yet we may be asking too much of it. Despite the best efforts of smart and dedicated individuals from around the industry, the technology world in which we now work is so complex that predictions may just be impossible.
Even the concept of the most basic step along the road-formerly known as a “generation” and now commonly known as a “node”-now means very little in the real world. What is the transistor gate-length for a “45nm” chip? It depends on whether it’s memory or logic, and for logic, whether it’s high-performance or low-power. DRAM fabs now do fractional-node shrinks.
With so much complexity to model and manage, a single roadmap just can’t provide significant details about the future. Different companies will use different combinations of known technologies to optimize within the risk/cost/specification abstract space. The ITRS still provides a valuable snapshot in time of where we’re at. However, it can’t tell us where we’re going because we’re all going in different directions these days. Company and consortia roadmaps are probably the only reliable guides remaining.
 |
Ed Korczynski
Senior Technical Editor