作者
Michael Zelikson,Kosta Luria,Lior Gil,Yuval Brown,Vadim Goldenbeg,Dor Kasif,Elias Hlees,Alex Vinichuk
摘要
Modern CPUs feature multiple power domains, most of which are grouped and share a common power supply per group. A combination of a common supply with per-domain DVFS implies a need for local voltage regulation, which in some SOCs is based on LDOs. Operation conditions feature dynamically variable$\mathrm{V}_{\text {IN }, V_{\text {OUT }}, \mathrm{I}_{\text {LOAD }}}$transients, and a variety of decoupling solutions. Such usage conditions usually cause$\mathrm{V}_{\text{MI}\mathrm{N}}$to depend on$\mathrm{V}_{\text{IN}}$, which translates into cost (test-time increase) and/or power/performance$(\mathrm{V}_{\text{MI}\mathrm{N}}$guard bands) penalties. In this paper, we report on a CPU-grade digital LDO (DLDO) that supports adequate (to CPU performance requirements) power densities and mitigates$\mathrm{V}_{\text{MI}\mathrm{N}}$dependency on the input voltage. The DLDO was implemented in a 10nm process node, as part of a multi-domain SOC. High-level power-plane partitioning is shown in Fig. 14.3.1 (top). Figure 14.3.1 (bottom) illustrates the high-level DLDO block scheme. The chosen architecture comprises a digital power stage and a digital controller for the following reasons: (a) digital design features weaker process dependency and better design convergence; (b) potentially higher$\mathrm{I}_{\text{LOA}\mathrm{D}}$and regulation BW (iso-area); and (c) better performance in low-DO$\left(\text{dropout }=\mathrm{V}_{\text {IN }}-\mathrm{V}_{\text {OUT}_{-}\text{TARGET }}\right)$conditions and easier transition to a bypass (BP) mode to enable lowest possible$\mathrm{V}_{\text{IN}}$(e.g.,$\Delta \mathrm{V}_{\text{IN}}$of 20mV leads to$\Delta \mathrm{P}_{\text{IN}}$~1 W for a 45W product). Some digital DLDO designs [1–3] feature a dual-loop architecture: a relatively-Iow-BW linear control and a high-BW, event-driven fast control (i.e. Non-Linear Control (NLC)). From an SOC standpoint, the disadvantage of the NLC is that its charge injection is DO dependent. The NLC efficiency drops with$\mathrm{V}_{\text{IN}}$and depends on an SOC power scenario. It was decided to base the DLDO architecture on a linear-only controller, fast enough to cope with$\mathrm{I}_{\text{LOA}\mathrm{D}}$transients and minimizing the dependency of$\mathrm{V}_{\text{OUT}}(\mathrm{t})$on$\mathrm{V}_{\text{IN}}$. In real CPUs,$\mathrm{I}_{\text{LOAD}}$reaches >30A, which corresponds to$> 100\mathrm{A}/\text{mm}^{2}$(DLDO area$\mathrm{R}_{\text{DS}_{-}\text{ON}}$wise) single-, double-, and triple-stack PMOS structures (Fig. 14.3.1, bottom). At a WC stress (high-DO/low-current conditions), only triple-stack structures conduct, reducing the power stress on a single PMOS by$9\times, \mathrm{w}$r.t. single-stack structure. When$\text{I}_\text{L O A D} \approx I_{M A X}$, single-stack structures are active, reducing the PS effective$\mathrm{R}_{D S_{-}O N}$. A CPU domain performance workloads feature$\left(\mathrm{I}_{\text {IDLE }}(<\sim 0.5 \mathrm{A}) \rightarrow\right. \mathrm{I}_{\text{MAX}})/\sim 3\text{ns}$current transients into$\mathrm{C}_{\text{out}}\geq\sim 500\text{nF}$, and corresponding$\mathrm{V}_{OUT}$droop is typically within ~8 to 10ns.$\mathrm{V}_{\text{IN}}$and$\mathrm{V}_{OUT}$variation ranges are around 1V (e.g., 0.6 to 1.6V and 0.55 to 1.3V, respectively). A stable and efficient operation of an LDO under such conditions requires at least a single-zero/single-pole (1 Z1P) compensator with a loop delay$\leq 3\text{ns}$and total computational delay of <2ns.