Experts: Hassan Khan (CMU) and Neil Thompson (MIT)
The integrated circuit provides the quintessential example of what economists call a “general-purpose technology,” where improvements in the underlying technology “spill-over” to other sectors generating broad-based innovation and prosperity (Jorgensen 2001; Jorgensen and Vu 2007). In semiconductors, decades-long improvements in performance and cost-effectiveness, often called “Moore’s Law,” have made countless new products possible, from intercontinental ballistic missiles to global environmental monitoring systems to smart phones to medical implants. But Moore’s Law is reaching its physical limits, which threatens to drastically narrow the beneficiaries of computer progress and slow the computational power supporting advances in artificial intelligences and other areas, and the innovations accruing to support national security, economic growth, and social well-being (Khan et al., 2017; Khan et al., 2018; Leiserson et al., 2020; Thompson and Spanuth, 2021). By one important measure, the benefits from semiconductor progress have already slowed from 52% per year to 3% per year (Hennessy and Patterson, 2019). It is also clear that this slowdown is making AI progress less sustainable (Thompson et al., 2020; Thompson et al., 2022).
However, there may yet be opportunities to rekindle progress by stepping off the slowing “s-curve” of silicon CMOS technologies, and instead moving to a new one based on materials and systems with different physical characteristics. But the path to these new technologies is uncertain and the institutions that historically guided industry progress are greatly weakened. If the U.S. wants to continue leading the world in advanced computing, a position already at risk (Thompson, Evans, Armbrust, Forthcoming), it will require the U.S. government to invest in facilities for prototyping and scaling-up beyond-CMOS devices. Investment in a scale-up facility is particularly important due to the coordination required across the “computing stack” to introduce a new device, including complementary innovations in equipment, computer architecture, programming languages, and software, and the interdependencies across these layers. Some computer architects have claimed that it is hard to get serious about designing novel architectures until at least 10,000 of a particular novel device are being made (Khan et al., 2014; Khan et al., 2018).
Hassan and Thompson propose to help answer several questions that will help guide U.S. investment in beyond-CMOS devices:
- What types of capabilities are needed to scale-up new technologies?
- How large are the potential gains from investing in these new prototypical and scale-up facilities?
- How broad a portfolio of successor technologies should be pursued?
Answering these questions will involve estimating the economic gains, and building on Thompson’s work on the economics of Moore’s Law (Thompson, 2017) as well as his on-going work on quantum computing, specialized accelerators, etc. When answering these questions, it's also important to consider the technical trade-offs of including multiple technologies in a single facility given the potential for cross-facility contamination; equipment under-utilization; and the additional facility and clean room costs required to keep material systems separate.