GPU Power Management via Compute Kernel Execution Time Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing power management techniques for GPUs are not effectively extendable to compute workloads, as they rely on graphics-specific metrics like frame rate, which are not applicable to non-graphics tasks, limiting energy-saving opportunities.
Innovation Solution
A framework for predicting the execution time of compute kernels using performance counters and profiling techniques, such as kernel-level and sub-kernel level profiling, to inform power management decisions like voltage and frequency adjustments or power collapse of unused GPU blocks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing power management techniques for GPUs are used, then power consumption can be managed for graphics workloads, but these techniques cannot be effectively extended to compute workloads
Solution Approach 1:
The patent extends power management capabilities from graphics-specific workloads to general-purpose compute workloads by using a unified execution time prediction approach. The system maintains a single power management framework that can handle both graphics rendering and compute kernels, eliminating the need for separate management techniques and enabling energy-saving opportunities across all GPU workload types.
Solution Approach 2:
The patent changes the fundamental parameter used for power management decisions from graphics-specific metrics (frame rate) to compute kernel execution time. By predicting execution time using performance counters and profiling techniques, the system can dynamically adjust power states based on actual compute workload characteristics, making power management effective for both graphics and compute tasks.
2Use of energy by stationary object
If graphics-specific metrics like frame rate are used for power management, then power consumption can be optimized for graphics tasks, but these metrics are not applicable to non-graphics tasks
Solution Approach 1:
The patent creates a universal power management approach that works for both graphics and compute workloads by replacing graphics-specific metrics with compute kernel execution time predictions. This unified metric can be applied across all GPU task types, maintaining energy optimization capability while expanding versatility to non-graphics tasks.
Solution Approach 2:
The patent introduces execution time prediction as an intermediary mechanism that bridges graphics and compute workloads. By using performance counters and profiling data to predict kernel execution time, the system creates a common basis for power management decisions that works regardless of whether the workload is graphics or compute-related.
3Productivity
If the GPU operates at high performance to meet execution deadlines, then compute tasks complete on time, but power consumption increases
Solution Approach 1:
The patent implements dynamic power management by continuously predicting compute kernel execution times and adjusting GPU power states accordingly. Rather than operating at fixed high performance, the system dynamically scales power consumption to match the actual execution requirements of compute tasks, ensuring deadline compliance while minimizing energy usage.
Solution Approach 2:
The patent performs execution time prediction before compute kernels execute, using performance counters and profiling data to forecast completion times. This preliminary action allows the system to proactively adjust power states and frequency settings to ensure deadlines are met without unnecessarily consuming high power throughout the entire execution period.
4Use of energy by moving object
If the GPU frequency is reduced to save power, then energy consumption decreases, but execution time may exceed deadlines
Solution Approach 1:
The patent uses feedback from performance counters and execution time predictions to dynamically adjust GPU frequency and power states. The system continuously monitors execution progress and compares predicted completion times against deadlines, adjusting frequency upward when needed to meet deadlines and downward when power savings are safe, thus maintaining reliability while optimizing energy consumption.
Solution Approach 2:
The patent implements dynamic frequency adjustment based on real-time execution time predictions. Rather than using a static low frequency to save power, the system dynamically scales frequency to match the actual computational requirements and deadline constraints, ensuring that power reduction does not compromise the ability to meet execution deadlines.
Data Source
AI summary
A device for processing data includes a processing unit configured to predict an execution time of a compute kernel on a secondary processing unit and, based on the predicted execution time, make a power management decision for the secondary processing unit.