Intelligent Memory DVFS Using Inter-Frame Bandwidth Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current frequency selection algorithms for graphics subsystems are ad-hoc and fail to properly account for key metrics like bandwidth profile, leading to inefficient frequency selection for 3D and media workloads with dynamic memory bandwidth profiles.
Innovation Solution
An intelligent memory DVFS scheme that exploits frame-to-frame correlation by using inter-frame history and histogram-based decision-making to determine the optimum frequency for system agents like DRAM, considering power limitations and performance needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If ad-hoc frequency selection algorithms are used, then device complexity is reduced, but power efficiency deteriorates due to frequent switching and energy wastage
Solution Approach 1:
The patent implements feedback by monitoring actual memory bandwidth usage and using it to adjust frequency selection decisions. The system continuously observes bandwidth profiles and uses this information to make informed DVFS decisions, avoiding both excessive frequency switching and energy wastage while maintaining performance when needed.
Solution Approach 2:
The patent performs preliminary action by pre-calculating and storing histogram data representing bandwidth profiles for different workload types. This historical bandwidth information is prepared in advance and used to predict future bandwidth needs, enabling the system to make proactive frequency adjustments rather than reactive ones, thereby reducing frequent switching.
2Productivity
If frequency is increased to meet dynamic memory bandwidth demands, then performance is improved, but power consumption increases
Solution Approach 1:
The patent applies dynamics by implementing adaptive frequency scaling that dynamically adjusts memory controller frequency based on actual bandwidth demands. Instead of using fixed or overly aggressive frequency scaling, the system continuously adapts frequency to match the dynamic nature of graphics workloads, ensuring high performance when bandwidth is needed while conserving power when demands are lower.
Solution Approach 2:
The patent changes the parameter of frequency selection from static or simplistic algorithms to a data-driven approach using bandwidth histograms. By analyzing historical bandwidth data and adjusting frequency based on predicted bandwidth needs, the system optimizes the frequency parameter to achieve better performance-power tradeoff across varying workload conditions.
3Loss of energy
If frequency switching is reduced to improve power efficiency, then energy wastage decreases, but responsiveness to dynamic bandwidth needs may deteriorate
Solution Approach 1:
The feedback mechanism monitors actual bandwidth usage patterns and compares them against predicted patterns from histograms. This allows the system to detect when frequency adjustments are actually needed versus when they would be unnecessary switching, maintaining responsiveness to genuine performance needs while avoiding energy-wasting frequent adjustments.
Solution Approach 2:
By pre-computing bandwidth histograms and using them to predict future bandwidth requirements, the system performs preliminary action that enables proactive frequency adjustment. This allows the system to prepare frequency changes in advance based on predicted workload patterns, maintaining responsiveness without reacting to every minor fluctuation that would cause unnecessary switching.
Data Source
AI summary
Methods and apparatus relating to techniques for intelligent memory DVFS (Dynamic Voltage and Frequency Scaling) scheme exploiting graphics inter-frame level correlation are described. In an embodiment, collection logic collects bandwidth usage information by a system agent during performance of one or more operations associated with a first graphics workload. Memory stores the collected bandwidth usage information. The selection logic causes selection of an operating frequency for the system agent to perform a plurality of operations associated with one or more graphics workloads based at least on the stored collected bandwidth usage information. The one or more graphics workloads occur after the first graphics workload. Other embodiments are also disclosed and claimed.