System-Level Optimization of Switched-Capacitor VRM and Core for Sub/Near- Computing

This brief proposes to jointly optimize a switched capacitor voltage regulator module (SC-VRM) combined with a compute core to minimize system energy per instruction. Past work seeking to optimize system energy efficiency has focused on separately maximizing SC-VRM efficiency or operating the compute core at its minimum energy operating point (MEOP). We first propose and verify a core-aware SC-VRM energy model, which explicitly accounts for throughput constraints. Second, we perform joint optimization considering throughput unconstrained applications (TUA) and throughput constrained applications (TCAs). We show that, for TUA, the system MEOP (S-MEOP) voltage is different from both the core MEOP (C-MEOP) and VRM maximum efficiency point (V-MEP) voltages, and operating at S-MEOP achieves 12.3% and 21.8% energy savings compared with C-MEOP and V-MEP, respectively. For TCA, S-MEOP is achieved at the same voltage as C-MEOP but different from the voltage for V-MEP, and 38.9% energy savings can be obtained by operating at S-MEOP.