AXES: Approximation Manager for Emerging Memory Architectures

Abstract

AXES is the first self-optimizing runtime manager designed to coordinate configurable approximation knobs across all levels of the memory hierarchy, overcoming the limitations of rigid, design-time policies. It continuously learns and updates its approximation strategy to minimize power consumption while strictly adhering to application-defined quality-of-service (QoS) thresholds. Demonstrated on a RISC-V Linux platform, AXES successfully saved up to 37% energy in the memory subsystem and reduced QoS violations by 75% with negligible overhead.

Report

AXES: Approximation Manager for Emerging Memory Architectures

Key Highlights

• Runtime Self-Optimization: AXES is the first approximation manager capable of learning and optimizing approximation policies at runtime, making it adaptive to unknown workloads and variable conditions.
• Full Hierarchy Management: It coordinates configurable approximation knobs across the entire memory hierarchy, including both on-chip cache and main memory segments.
• Power/Quality Trade-off: The primary objective is to minimize power consumption (energy savings) while ensuring the application's required quality-of-service (QoS) threshold is not compromised.
• Significant Energy Savings: The system achieved up to 37% energy savings within the memory subsystem without requiring any design-time overhead.
• QoS Improvement: AXES reduced QoS violations by 75% when operating with less than 5% additional energy expenditure.

Technical Details

• Architecture Scope: Manages interdependent approximation knobs across memory subsystems (cache and main memory).
• Methodology: Utilizes a continuous runtime policy updating mechanism to achieve self-optimization.
• Target Metric: Optimization focuses on power consumption reduction.
• Constraint Metric: Optimization is bounded by application-developer specified quality constraints.
• Demonstration Platform: The system was successfully demonstrated on a RISC-V Linux platform utilizing approximate memory segments.
• Key Functions: (1) Runtime policy learning for variable QoS constraints, (2) automatic metric optimization, and (3) coordinated decision-making for interdependent hardware knobs.

Implications

• Advancing RISC-V Power Efficiency: By providing a systemic, runtime mechanism for memory approximation, AXES significantly enhances the energy efficiency of RISC-V architectures, which is crucial for edge computing, IoT, and high-performance embedded systems.
• Enabling Approximate Computing in Standard Systems: The demonstration on a RISC-V Linux platform proves that complex approximate computing techniques can be managed dynamically by the OS/runtime layer, moving them out of the realm of specialized design-time solutions.
• Support for Emerging Memory: The manager is specifically tailored for "Emerging Memory Architectures," suggesting that it is vital for efficiently integrating new non-volatile or inherently approximate memory technologies into the RISC-V ecosystem while maintaining application reliability.
• Flexible Hardware Utilization: AXES allows RISC-V hardware designers to integrate approximation features (e.g., approximate cache segments) without needing to hardcode policies, relying instead on the runtime manager to handle optimal usage based on the current workload.