2017 conference paper
A case for standard-cell based RAMs in highly-ported superscalar processor structures
Proceedings of the eighteenth international symposium on quality electronic design (isqed), 131–137.
Highly-ported memories are pervasive within superscalar processors. Accordingly, they have been targets for full-custom design using multi-ported versions of the 6T SRAM bitcell. Unfortunately, full-custom design of highly-ported memories is becoming exceedingly difficult in deep sub-micron technologies. This paper makes the case for implementing highly-ported memories with standard cells (flip-flops, muxes, clock buffers). In lieu of exotic peripheral circuits for each port, standard-cell SRAMs use muxes. Consequently, area differences between full-custom and standard-cell designs are greatly reduced at a high number of ports. To also compete with full-custom memories in terms of timing and power, we introduce a standard-cell memory compiler with three key features: (i) per-row clock gating, (ii) a new tri-state based mux standard cell, and (iii) a modular layout strategy, which is the centerpiece of the memory compiler. For a 16-read/8-write 128-entry register file, our modular standard-cell memory consumes 13% more area and 4% more power, and is 35% faster, than the custom memory produced by FabMem. The automatic (built-in) robustness of standard cell designs further weigh in their favor, contrasted with exquisite transistor sizing/tuning of intertwined sub-circuits in a full-custom design.