VLSI 2025 delivered concrete data on the decisions that will define the next decade of semiconductors. SemiAnalysis covered Intel 18A process details and cost comparisons against TSMC, the adoption boundaries of backside power delivery, the 4F2 versus 3D DRAM race, China's FlipFET transistor work, and digital twins operating at every scale from atomic simulation to full fab orchestration. These are not roadmap slides. These are engineering tradeoffs with winners and losers already becoming visible.
The digital twin section alone is worth the full read. Synopsys QuantumATK demonstrated Machine-Learned Force Field simulation using Moment Tensor Potentials achieving near-DFT accuracy in 17 minutes versus 12 days with traditional Density Functional Theory. GPU-accelerated DFT-NEGF on just 4 A100s produced a 9.3x speedup over CPU. These tools are now being used in production contact engineering and gate oxide stack design, the exact layers that determine transistor performance in Gate All Around architectures. The gap between virtual and physical silicon is closing fast, and the cost implications for yield and design cycle time are significant.
The DRAM and backside power sections are where the political and technical arguments get sharp. Not every node will adopt backside power delivery, and the article explains exactly where the economics break and where they do not. On DRAM, the 4F2 versus 3D architecture question determines whether memory scaling continues on the 2D plane or forces a vertical reckoning. The full piece names the likely winners. Read it for the reasoning, not just the conclusion.
[READ ORIGINAL →]