The Transistor Cliff

Asterisk Magazine, Issue 03, June 2023. By Sarah Constantin.

Key Takeaways

• Two independent walls arrive around 2030: thermodynamic (transistors at 4–5nm cannot switch reliably against heat fluctuation) and photolithographic (light wavelength limits feature size).
• Cost already broke before density did. “The cost per transistor stopped decreasing in 2011” and has since risen — which violates Moore’s original price criterion even while density keeps climbing.
• Memory is the live bottleneck. DRAM bandwidth improved ~30x in twenty years while compute rose ~90,000x; in a six-month training run roughly 2.5 months is data movement rather than computation.
• Alternatives are modest or unproven: 3D stacking, specialized accelerators, and experimental paradigms (optical, memristors).
• The consequence for AI: progress shifts from make the model bigger to algorithmic innovation and task-specific work.

Connections

• Ingested for Chips For America - Research and it is the physical-limits floor under that whole argument: industrial policy can move where chips are made, not how small they get.
• Pairs against Bubble Architecture — if scaling stops paying, the subsidized-inference economics get worse, not better.