Moore’s law in peril and the future of computing

Episode Summary

Executive Summary: The episode examines Moore’s law as both a historical observation and a fading guide for computing progress. Guests explain its 50-year run, the physical and economic limits now slowing transistor scaling, and the rise of alternatives like optical computing, compound semiconductors, and 3D chip designs. The discussion also connects chip limits to AI growth, energy use, geopolitics, and UK industrial strategy.

Main Topics: What Moore’s law is and why it mattered (Priority: 5/5): Lewis Barson traces Moore’s law from Gordon Moore’s 1965 observation to its role as a benchmark for exponential chip progress and the digital revolution. Why Moore’s law is slowing (Priority: 5/5): Experts explain that transistor miniaturization is approaching atomic scales, bringing quantum effects, thermal noise, and diminishing returns that make further scaling harder. Economic and geopolitical importance of semiconductors (Priority: 5/5): The podcast highlights chips as critical infrastructure for healthcare, automotive, defense, and telecoms, and notes US-China-Europe competition plus supply-chain vulnerability. UK semiconductor strategy and industrial positioning (Priority: 4/5): Lewis Barson discusses the UK’s niche strengths in chip design, packaging, compound semiconductors, and materials research, along with the government’s semiconductor strategy and its limitations. Future computing paradigms beyond silicon (Priority: 5/5): Researchers describe optical computing, 2D materials, compound semiconductors, photonics, spintronics, and quantum computing as possible ways to extend or replace current approaches. AI compute demand, energy constraints, and sustainability (Priority: 5/5): The conversation links Moore’s law to the rising compute needs of generative AI and warns that energy efficiency limits could make training large models expensive and environmentally costly.

Key Arguments: Moore’s law began as an observation, not a prophecy, but became self-reinforcing because the industry treated it as a target to hit. The law appears to be slowing because transistors are now approaching the scale of individual atoms, where quantum and thermal effects become serious constraints. Semiconductors are economically and strategically critical, with supply disruptions affecting industries such as automotive and national security. The UK cannot compete head-on with the US, China, or Europe on scale, so it should focus on design, packaging, compound semiconductors, and next-generation technologies. Alternative technologies such as optical computing may not replace conventional computers entirely, but they could serve as accelerators for specific workloads like neural networks and matrix operations. AI progress has historically tracked compute growth, but current models are becoming limited by hardware cost, power density, and energy consumption. If AI compute demand continues rising rapidly, data centers and training runs may become unsustainably energy-intensive unless new hardware and algorithms improve efficiency. Diminishing returns in R&D mean that sustaining progress in chip design and computing will require more research input, not just incremental scaling. AI may also help solve the bottleneck by augmenting scientific and engineering research, potentially accelerating discovery of new paradigms. Quantum computing is unlikely to replace classical computers soon because useful algorithms, suitable problem domains, and reliable hardware remain difficult to develop.

Data Points: Original doubling period in Moore’s 1965 article: about every year for 10 more years - Moore’s initial prediction based on the prior three years of chip growth Updated Moore’s law pace: doubling every two years - Moore’s 1975 revision of the trend Observed lifespan of the trend: from the early 1970s to the 2020s - Approximate period over which transistor doubling persisted Transistor count growth: from around 5,000 to about 50 billion - Approximate increase in transistors per chip over five decades UK semiconductor strategy funding: £1 billion over 10 years - Government semiconductor strategy described as a starting point US semiconductor support: more than $50 billion - Recent US government commitment to industry and innovation TSMC Arizona investment: $40 billion - Latest plant triples the company’s overall investment in Arizona Worldwide chip market size: roughly half a trillion dollars - Semiconductors described as one of the world’s biggest markets Global trade ranking of chips: 4th most traded product - After crude oil, refined oil, and cars Automotive production impact during shortage: more than a 25% slump - Chip shortage contributed to auto supply-chain issues during the pandemic Transistor scale: sub-10 nanometer; latest process around 3 nanometers - Used to illustrate proximity to atomic scales Data center electricity use: 3% of global electricity - Current consumption cited as a baseline concern Energy use if scaled 10x: 30% of global electricity - Illustrative warning about rising compute demand AI training compute estimate: about 10^25 floating-point operations - Estimated cost of training GPT-4 over the full training process Projected energy-efficiency bound: ~10^16 floating-point operations per joule - Estimated near-term limit for existing microprocessors Hypothetical large training run: 10^40 floating-point operations - Used to discuss future AI scaling and energy requirements Earth-wide annual energy scale: ~10^24 joules - Comparison for the energy needed for very large AI training runs Model training doubling time during deep learning boom: about 6 months - From 2012 to around 2018 Recent training-compute doubling time: about 11 months - Since around 2017, growth has slowed somewhat Google-scale compute trend in history: roughly every 2 years - Historical AI compute growth before deep learning acceleration

Pivotal Quotes: "Moore's law really charts the contribution I would say of physics and material science to the digital technology revolution" — Lewis Barson: Explaining why physicists should care about semiconductors "you should think of optical computing as ... a new technology that will enable accelerator functions" — Thomas Verjera de Lima: Describing the role of optical computing alongside electronic processors "we're entering this new regime from abundance to scarcity" — Thomas Verjera de Lima: Warning that AI compute will become more expensive and energy-constrained as Moore’s law slows

Implications: Listeners should see chips as a strategic and scientific bottleneck shaping AI, energy use, and national competitiveness. The future likely depends on new materials, architectures, and policies—not just smaller silicon transistors.