Inferensys

Glossary

State Retention Power Gating (SRPG)

State Retention Power Gating (SRPG) is an advanced power gating technique that shuts off power to idle circuit blocks while preserving critical state data in isolated retention registers.
Data scientist building training data pipeline on laptop, data preprocessing visible, technical workspace.
POWER AND THERMAL MANAGEMENT

What is State Retention Power Gating (SRPG)?

State Retention Power Gating (SRPG) is an advanced circuit-level power management technique that eliminates leakage power in idle logic blocks while preserving their internal state data.

State Retention Power Gating (SRPG) is a fine-grained power gating methodology where a header or footer switch disconnects the main power supply (VDD) from a circuit block, eliminating both dynamic and static (leakage) power. A small, isolated retention power domain, supplied by a separate, always-on rail (VDD_ret), powers a minimal set of retention flip-flops or shadow latches that store the block's critical architectural state. This allows the core logic to be completely powered off, achieving near-zero leakage, while enabling a rapid, stateful wake-up by restoring power and reloading the preserved context.

SRPG implementation requires careful power domain isolation using isolation cells to prevent corrupt signals from the powered-down block from propagating. Upon wake-up, a controlled power-up sequence restores the main supply, deactivates isolation, and transfers state from the retention elements. This technique is fundamental in low-power System-on-Chip (SoC) design for mobile and embedded Neural Processing Units (NPUs), enabling aggressive power savings during idle periods without the performance penalty of a full reboot from external memory. It is specified using standards like the Unified Power Format (UPF).

ARCHITECTURAL ELEMENTS

Key Components of an SRPG Implementation

State Retention Power Gating (SRPG) is a sophisticated hardware power management technique. Its implementation requires the coordinated design of several specialized circuit and system-level components to safely remove power while preserving critical data.

01

Retention Flip-Flops

Retention flip-flops are specialized memory cells at the core of SRPG. Unlike standard flip-flops that lose state when power is removed, these cells feature a small, dedicated retention latch powered by a separate, always-on Vret supply rail. During normal operation, they function identically to standard cells. When the main power domain (Vdd) is gated off, the retention latch holds the stored bit value using minimal leakage current. Upon power restoration, the saved state is restored to the main flip-flop, allowing the logic block to resume operation exactly where it left off. These cells are strategically placed to retain only the minimal critical state (e.g., architectural register values, finite state machine states) necessary for correct functionality.

02

Power Switches

Power switches (or power gates) are the physical transistors that connect or disconnect a circuit block from its main power supply (Vdd). Implemented as large arrays of header switches (PMOS) between Vdd and the virtual Vdd rail, or footer switches (NMOS) between the virtual ground and ground, they act as electrical on/off switches for the power domain.

  • Design Considerations: They must have very low on-resistance (Ron) to minimize performance and voltage drop (IR) penalty when active, yet be small enough to limit area and leakage overhead.
  • Control Sequencing: Their turn-on and turn-off is carefully sequenced by the Power Management Unit (PMU) to prevent large inrush currents and ensure stable operation.
03

Isolation Cells

Isolation cells are special logic gates inserted at the outputs of a power-gated block that is being shut down. Their purpose is to prevent floating or corrupted signals from propagating to blocks that remain powered on, which could cause metastability, excessive leakage, or functional failure.

  • Function: When the source domain's power is removed, the isolation cell clamps its output to a known, safe logic level (e.g., logic 0 or 1) as defined by the isolation strategy.
  • Placement: They are placed in the always-on power domain, receiving a control signal from the PMU that activates isolation before power is removed and deactivates it after power is fully restored and stable.
04

Power Management Unit (PMU) & Controller

The Power Management Unit (PMU) is the finite state machine or microcontroller that orchestrates the entire SRPG sequence. It implements the precise timing and control protocol for entering and exiting the retention state.

Standard SRPG Sequence:

  1. Quiesce: Halt new transactions and allow pending operations in the block to complete.
  2. Save: Assert the save signal, triggering retention flip-flops to copy their state to the retention latch.
  3. Isolate: Assert the isolate signal to clamp outputs.
  4. Power Down: De-assert the power switch enable, turning off the main Vdd to the block.
  5. (Sleep Period): The block resides in its low-leakage retention state.
  6. Power Up: Re-assert the power switch enable to restore Vdd.
  7. Restore: Assert the restore signal, loading the saved state from the retention latch back to the main flip-flops.
  8. De-isolate: De-assert the isolate signal, releasing the output clamps.
  9. Resume: Release the block to resume normal operation.
05

Power Domain & Rail Design

SRPG necessitates a power domain architecture. A power domain is a contiguous region of logic that shares a common, switchable power supply (virtual Vdd/Vss).

  • Domain Partitioning: The chip is divided into multiple power domains (e.g., always-on, CPU core, GPU, DSP). The SRPG block resides within its own switchable domain.
  • Power Rail Strategy: This requires a dedicated always-on power rail (Vret) for the retention latches, and a switchable main power rail (Vdd) for the rest of the logic. The power delivery network must be designed to provide clean, stable power to Vret even when Vdd is gated, and to manage the inrush current when Vdd is switched back on.
  • Unified Power Format (UPF): This power intent is formally specified using the IEEE UPF standard throughout the Electronic Design Automation (EDA) flow for implementation and verification.
06

State Save/Restore Controller & Sequencing

This is the logic, often part of the PMU or a local controller, that generates the precise save and restore control signals for the retention flip-flops. Its timing is critical.

  • Save Timing: The save signal must be asserted while the main Vdd is still stable and valid, but after the block is quiesced.
  • Restore Timing: The restore signal must be asserted only after the main Vdd has been restored and stabilized to a valid operating voltage. Incorrect timing can corrupt the retained state.
  • Clock Gating Integration: The controller typically works in concert with clock gating. The clock to the domain is gated off before the save sequence begins and is only restored after the de-isolation step, ensuring no spurious transitions occur during the power transition.
COMPARISON

SRPG vs. Other Power Management Techniques

A feature comparison of State Retention Power Gating against other primary power management techniques used in modern System-on-Chip (SoC) and NPU design.

Feature / MetricState Retention Power Gating (SRPG)Power GatingClock GatingDynamic Voltage & Frequency Scaling (DVFS)

Primary Power Savings Mechanism

Eliminates leakage power in idle blocks

Eliminates leakage power in idle blocks

Reduces dynamic power by preventing clock toggling

Reduces dynamic power via voltage/frequency scaling

State Retention During Low-Power Mode

Wake-Up Latency

10s to 100s of microseconds

100s of microseconds to milliseconds

< 1 microsecond

1 to 10 microseconds

Static (Leakage) Power Reduction

95%

95%

0%

~30-50% (via voltage scaling)

Dynamic Power Reduction

~100% (block inactive)

~100% (block inactive)

~10-50% (block active but idle)

Proportional to V²f scaling

Typical Granularity of Control

Circuit block or power domain

Circuit block or power domain

Register, module, or block level

Core, cluster, or chip level

Hardware Overhead

Retention flip-flops, isolation cells, second power rail

Header/footer power switches, isolation cells

Clock gating integrated into synthesis

Voltage regulators, frequency PLLs, sensors

Software/OS Involvement

Firmware or driver-controlled state save/restore

Firmware or driver-controlled

Typically automatic (compiler/synthesis)

OS-directed (P-states via ACPI)

Use Case Primary Target

Long idle periods where state must be preserved (e.g., CPU core sleep)

Long idle periods where state can be lost/reloaded (e.g., peripheral shutdown)

Fine-grained reduction during short idle bursts within active operation

Active workload adaptation to match performance demand

POWER AND THERMAL MANAGEMENT

Primary Applications of SRPG

State Retention Power Gating (SRPG) is a critical technique for managing power and heat in modern processors. Its primary applications focus on maximizing energy efficiency while preserving critical operational state, enabling advanced power management in complex systems.

01

Mobile & Embedded Processors

SRPG is foundational in mobile SoCs (System-on-Chip) and embedded systems where battery life is paramount. It allows individual cores or functional blocks (e.g., GPU, DSP, ISP) to be completely powered down during idle periods without losing their internal register state.

  • Key Benefit: Drastically reduces leakage power, the dominant source of power drain in advanced nanometer processes when a device is in sleep or low-activity mode.
  • Example: A smartphone application processor uses SRPG to power-gate an idle CPU cluster while a retention flip-flop bank maintains the architectural state, enabling a near-instantaneous wake-up when the user interacts with the screen.
02

High-Performance Computing (HPC) & Servers

In data centers, SRPG manages power and thermal density at a granular level. It enables dark silicon strategies, where non-critical portions of a large multicore CPU or accelerator can be turned off to stay within a strict Thermal Design Power (TDP) or power budget.

  • Application: Selectively gating idle cores or execution units within a core during low-load periods, while retaining the core's context for rapid resumption.
  • Impact: Allows for higher peak core counts on a die, as not all transistors need to be active simultaneously, improving overall performance per watt.
03

AI/ML Accelerators (NPUs/GPUs)

Neural Processing Units and AI accelerators have highly variable workloads. SRPG is used to power-gate entire tensor cores, systolic arrays, or memory buffers between inference/training batches.

  • Mechanism: Between processing layers or data batches, computational units are power-gated. A small retention memory holds configuration data (e.g., weight pointers, layer parameters), eliminating the need to reload from DRAM on wake-up.
  • Advantage: Essential for meeting the stringent power budgets of edge AI devices and reducing total cost of ownership in cloud AI infrastructure.
04

Always-On/Context-Aware Subsystems

Modern devices feature low-power, always-on domains for tasks like sensor hub processing, voice wake-word detection, or ambient display. SRPG is used within these domains to achieve ultra-low quiescent current.

  • Implementation: A tiny microcontroller or dedicated hardware block remains powered by a separate, always-on rail. Its internal logic is aggressively power-gated using SRPG, with only the minimal state for the current context (e.g., a voice model identifier) held in retention registers.
  • Result: Enables features like "raise to wake" or "Hey Google" detection while adding minimal impact to overall device standby time.
05

Functional Safety & Automotive Systems

In automotive and industrial systems, SRPG supports functional safety (ISO 26262) and low-power modes without compromising deterministic recovery. It allows complex Electronic Control Units (ECUs) to enter deep sleep while preserving critical fault logs, diagnostic states, and configuration data.

  • Critical Use: Ensuring a fail-operational or fail-safe state is maintained across power cycles of non-critical subsystems. Retention registers hold error codes and system health data.
  • Requirement: The isolation and separate power supply for retention cells must be designed to be highly reliable, often with redundancy, to prevent data corruption.
06

Power Sequencing & State Transitions

SRPG is a key enabler for complex, multi-state power management protocols like ACPI and PSCI. It facilitates rapid transitions between performance states (P-states) and idle sleep states (C-states).

  • Process: When an OS requests a core to enter a deep C-state (e.g., C6), the core's power domain is gated off. The Power Management Unit (PMU) ensures the retention voltage is stable before the main supply is removed.
  • Standardization: Power intent for SRPG is specified using standards like Unified Power Format (UPF), ensuring correct implementation through the EDA toolchain from RTL to layout.
STATE RETENTION POWER GATING

Frequently Asked Questions

State Retention Power Gating (SRPG) is a critical technique for managing power and thermal constraints in modern NPUs and SoCs. These questions address its core mechanisms, applications, and trade-offs.

State Retention Power Gating (SRPG) is an advanced power management technique that shuts off power to a large portion of a digital circuit block to eliminate leakage power, while a small, isolated region powered by a separate retention power supply preserves the block's critical architectural state (e.g., in retention flip-flops or retention registers). This allows the block to be powered down completely during idle periods without losing its operational context, enabling a fast, low-energy wake-up by restoring power and reloading the saved state, rather than re-initializing from scratch. SRPG is a key enabler for achieving ultra-low-power sleep modes in modern System-on-Chip (SoC) designs for mobile, IoT, and edge AI applications.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.