Rapid advances in AI and high-performance computing (HPC) are driving next-generation systems to demand unprecedented I/O bandwidth alongside stringent energy-efficiency targets. Meanwhile, advanced packaging and increasing routing densities expose ul...
Rapid advances in AI and high-performance computing (HPC) are driving next-generation systems to demand unprecedented I/O bandwidth alongside stringent energy-efficiency targets. Meanwhile, advanced packaging and increasing routing densities expose ultra-short-reach (USR) electrical links to severe interwire coupling and simultaneous switching noise (SSN), making conventional single-ended (SE) signaling difficult to scale without sacrificing robustness or resorting to complex equalization.
This dissertation presents Xtalk-Minimizing Affine Signaling (XMAS), an affine-transformation-based multiwire signaling framework that mitigates far-end crosstalk while maintaining high wire utilization. The operating principles of XMAS are derived from a rigorous analytical model, and its benefits are validated through numerical simulations and silicon measurements. A 28-nm CMOS prototype binary XMAS transceiver achieves a bandwidth density of 3.6 TB/s/mm and an energy efficiency of 0.65 pJ/bit. Compared with SE signaling over the same channel, XMAS reduces crosstalk-induced peak-to-peak jitter by 75% at 10 GBaud/pin and increases the eye width to 0.2 unit interval (UI) at BER < 10−12.
To further increase pin efficiency, this dissertation explores a ternary extension of XMAS that preserves the affine signaling framework while increasing the information conveyed per UI, analogous to three-level pulse-amplitude modulation (PAM-3). Post-layout system-level simulations demonstrate that the proposed ternary XMAS(T-XMAS) maintains signal integrity over strongly coupled on-chip channels.