The end of the roadmap is near for the advanced chips sector, the CHIPS Act should be forging a new path.

The US Department of Commerce recently delayed several billion dollars of funding for research into advanced chips. Specifically, it axed a third Notice of Funding Opportunity to subsidize commercial chip research and development facilities under the CHIPS and Science Act — despite protests from industry associations and state executives.

At the same time; Congress directed the CHIPS office to inject $3.5 billion into a “secure enclave” providing advanced chips for military and intelligence applications, and drawing from the same $52 billion pot for research, development and manufacture of advanced semiconductor technologies set out by the act.

The CHIPS office is facing “overwhelming demand” for funding, and doesn't have the cash or administrative capacity to do everything. A potential zero-sum decision between funding for commercial R&D, and spending on secure military and intelligence, shows the considerable difficulties of managing a multi-billion dollar industrial policy in the world’s most important technology sector.

This is an edition of Inferences, by Minerva Technology Policy Advisors. If you'd like to read it each week on a Monday, please consider subscribing here.

What’s happening? Opportunity cost. Commerce Secretary Raimondo has said that the requests for CHIPS funding are already running billions of dollars over the total available allocation. The task of doling out huge amounts of money across a range of pressing concerns requires time and attention; competing initiatives will make the tension between short-term spending on current chip technologies, and long-term spending on a new path for the future even more clear.?

Why does it matter? Because solving for present security concerns could slow down potentially riskier bets on R&D for the future. Such bets are needed over the longer-term to keep the semiconductor industry, and by extension, the AI sector, innovating.?

Industry associations are already concerned that the underlying objectives of the CHIPS and Science Act could be jeopardized by penny pinching on research funding, because the sector needs to make breakthroughs rather than merely increasing the volume of traditional computing capability.

Inference: Chip innovation is nearing the end of its current roadmap, based on packing more and smaller transistors into thinner components. Breakthroughs are needed; but funding the process of devising and implementing new fabrication techniques and chip designs is risky and demands long-term funding.

The bigger picture is that the future of the industry looks to be in a number of cutting-edge techniques and designs that can help break the ceiling on traditional means of making chips faster and more efficient. Embracing them is key for the push to enable advanced artificial intelligence, but further research into the commercialization of new technologies is required. Specifically in;

• dedicated accelerators, which are optimized for the types of data used to train artificial intelligence models. They have their memory tailored, and often cached on the chip itself, to the specific data-forms that the model is using for inference, and can be faster and more efficient;

• photonics, which use photons rather than electrons to transfer information around the circuit; experimental studies have shown that photonic components can have higher calculation speed, be more energy efficient and process more information than traditional electrical chips. Photonic interconnecting technologies that could help traditional components communicate faster are also showing promise; and
• multi-die systems, also known as chiplets; which are disaggregated components that break up monolithic chips into multiple pieces which can function together, with reduced power delivery and cooling demands. Imec, the world-leading semiconductor research institute, has projected a new innovation roadmap out to 2036; it is based on thinner chiplets stacked into 3D architectures.

This is exciting, furiously complicated stuff, but the strategy for synthesizing these breakthrough chip designs and unlocking the improvements in performance they could offer is not yet clear. The geopolitical contest unfolding around semiconductor technologies, as the key enabler for more advanced artificial intelligence, could be defined by developing such a strategy.?

There are myriad challenges to overcome. One engineer at a leading chip company told Inferences that “skepticism remains about photonic processing, because the power consumption implications are untested, but with the advent of chiplet designs you could find it’s used for high speed inter-chiplet comms instead of going through a substrate. I wouldn’t expect to see it in the mainstream for ten to twenty years. Moving electrons around is hard enough. If you have to impart a photon and catch it you’re going to be much more power hungry.”

Geopolitically speaking, China may be relying on photonics, accelerators and disintegrated designs as a way of leapfrogging the US and recovering from the innovation debt created by export controls on advanced chips; something Beijing University scientist Yao Yang sees as an opportunity to “change lanes to overtake”. A group of Chinese scientists published a paper earlier this month detailing a large-scale photonic chiplet, called Taichi. It features a hybrid photonic and electronic architecture, and claims improved performance and scalability. MIT Technology Review also reported earlier this year that China’s move to embrace chiplets is an effort to move to the frontier and circumvent US semiconductor technology controls. For the US, the move to allocate more funding to experimental science is risky, when increasing the volume of conventional chips and data centers can be effective in the short to medium term.?

In the market, all the major chip manufacturers, including Intel, Nvidia, TSMC and AMD are working on next-generation designs; but making advanced chips is difficult. Making the machines that make advanced chips is even harder. Organizing them into a fabrication facility is harder still. Training expert engineers to optimize the process might be the biggest challenge facing modern economies with ambitions in cutting-edge technology. It will take significant amounts of R&D spending and a degree of risk appetite to pull off, something that could become harder to muster as trade restrictions on current cutting edge chips pinch revenue that would otherwise turn into profit that could be recycled into R&D.?

The upshot. Governments hoping to see sustained innovation need to support financing for:

• new fabrication facilities; that can do more than the lithography required for conventional chips, and can accommodate manufacturing processes for chiplets, photonic (interconnectors and components), and other emerging chip design concepts;
• research into new and more efficient chip architectures, beyond photonics, chiplets and dedicated accelerators with enough runway to turn them into commercialisable products; and
• energy systems that can cope with the demands of decarbonisation and rapid global warming, as the demand for power is set to increase, even if chips become more efficient.

The danger is that appetite for addressing these knotty long-term commercial and innovation challenges ends up taking second seat to the surer bet of solving for more immediate national security needs, at least in the short term.