Marvell's 3nm FinFET 212.5Gbps SerDes hits 54.7dB reach and 2.05pJ/bit, the longest-reach and most efficient published at this rate. A shared PLL plus heavy DSP equalization breaks the old rule that reach costs power, giving 1.6T racks structurally more room in cable length and power budget.

Instead of piling on more laser wavelengths, the University of Washington uses microring modulators for coherent QAM, pushing a single wavelength to 320Gb/s at just 0.69pJ/b. RAMZI sidesteps MRM amplitude-phase coupling and time-multiplexed thermal tuning cuts the cost of thermal stability, delivering coherent, low-power and compact on low-cost 45nm silicon photonics.

NTT cut the junction capacitance of a III-V electro-absorption modulator to 26fF with a membrane structure, pushing device bandwidth past 67GHz at just 0.75 pJ/bit system energy. We walk through all 11 figures of this VLSI 2026 paper.

NVIDIA's 32 Gb/s optical receiver reaches -18.5 dBm sensitivity at 0.416 pJ/b with an almost free trick: reusing the rail-to-rail TIA output as a noise-free DFE decision, trading under 1 mA for 2 dB of sensitivity. Why receiver sensitivity is the best investment in the whole optical link — explained figure by figure.

Direct-detection optical receivers normally use only one terminal of the PD. NVIDIA recovers the wasted cathode current to build a differential TIA, improving sensitivity by roughly √2. The TIA doubles, but laser power across the whole link drops. A figure-by-figure walkthrough of VLSI 2026 C20.2.

AI's performance bottleneck is no longer compute but data movement across chips and packages. At VLSI 2026, TSMC answered with its 3DFabric platform: make the package itself the vehicle for AI scaling. We walk through all 19 figures, from 48x compute and 34x HBM bandwidth to power-thermal co-optimization and agentic EDA.