GPU Workload Metering for Power Budget Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current GPU power management techniques, such as dynamic voltage frequency scaling and pulse skipping, are ineffective in reducing power consumption below the minimum voltage frequency (Vmin), leading to unsatisfactory performance in power-constrained devices, and often require larger, less powerful GPUs to operate within limited power budgets.

Innovation Solution

Implementing a Workload Metering (WLM) process that calculates a duty cycle to transition the graphics engine between active and idle states, using race-to-sleep techniques and power gating to reduce power consumption, thereby managing GPU workload and meeting predetermined power budgets.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Use of energy by moving object

If dynamic voltage frequency scaling and pulse skipping techniques are used to reduce power consumption, then power consumption is reduced above Vmin, but power and performance scale poorly below Vmin due to coarse-grained clock reduction

Engineering Contradiction:
Improvepower consumptionVSAvoidperformance
Core Design Contradiction:
Use of energy by moving objectVSProductivity

Solution Approach 1:

The patent implements periodic action by introducing forced idle periods between active periods for the graphics engine. The duty cycle controller periodically transitions the graphics engine between active and idle states, creating a rhythmic on-off pattern that reduces average power consumption while maintaining performance during active periods. This periodic operation allows the system to meet power budgets without permanently degrading the engine's computational capability.

Inventive Principle:
Principle #19Periodic action

Solution Approach 2:

The patent applies dynamics by making the graphics engine's operational state variable and adaptive. Instead of fixed voltage/frequency settings, the system dynamically adjusts the duty cycle based on real-time power consumption measurements and workload characteristics. The graphics engine transitions between different operational modes (active, idle, low-power states) based on current conditions, enabling flexible power management that maintains performance when needed while reducing power consumption during less demanding periods.

Inventive Principle:
Principle #15Dynamics

2Use of energy by moving object

If the graphics engine is forced into idle state periodically to reduce power consumption, then overall power draw is reduced to meet power budget, but performance may be degraded

Engineering Contradiction:
Improvepower drawVSAvoidperformance
Core Design Contradiction:
Use of energy by moving objectVSProductivity

Solution Approach 1:

The patent implements feedback by continuously monitoring the actual power consumption of the graphics engine and using this information to adjust the duty cycle. The duty cycle controller receives real-time power consumption data and workload information, then dynamically modifies the ratio of active to idle periods to maintain power consumption within the budget while maximizing performance. This closed-loop feedback mechanism ensures that performance is not unnecessarily degraded when power consumption is already within acceptable limits.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent applies partial action by implementing duty cycle throttling that allows the graphics engine to operate at full performance during active periods rather than continuously reducing its operational capacity. Instead of applying a constant performance penalty, the system provides full computational power when the engine is active, accepting that it will be idle during forced idle periods. This approach maintains peak performance capability while reducing average power consumption through temporal partial operation.

Inventive Principle:
Principle #16Partial or excessive action

3Use of energy by moving object

If microcontroller-based software controllers modulate p-state to adjust power consumption, then power consumption is reduced, but the approach is not suitable for highly power constrained devices operating below Vmin

Engineering Contradiction:
Improvepower consumptionVSAvoidoperating range
Core Design Contradiction:
Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The patent applies universality by creating a duty cycle control mechanism that functions effectively across the entire operating range of the graphics engine, including regions below Vmin where traditional techniques fail. The duty cycle approach is universally applicable regardless of the specific voltage/frequency operating point, making it suitable for both high-performance and highly power-constrained devices. This multi-functional approach allows the same control mechanism to serve devices with different power budgets and performance requirements.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent implements parameter changes by fundamentally altering the control parameter from voltage/frequency modulation to duty cycle modulation. Instead of changing voltage and frequency within continuous ranges, the system changes the temporal parameter of operation by adjusting the ratio of active to idle time. This parameter transformation enables effective power management below Vmin where traditional voltage/frequency scaling becomes ineffective due to coarse-grained clock reduction and discrete operating points.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentUS11137815B2Metering GPU workload with real time feedback to maintain power consumption below a predetermined power budget
Publication Date: 2021.10.05 NVIDIA CORP
  • US11137815B2 patent drawing
  • US11137815B2 patent drawing
  • US11137815B2 patent drawing

AI summary

Embodiments of the present invention provide methods and apparatus for metering GPU workload in real time. Metering of the GPU workload is performed by a Workload Metering (WLM) algorithm implemented in software or firmware that calculates a duty cycle for the graphics engine. The duty cycle forces the graphics engine to transition from a busy state to an idle state periodically based on measured power consumption, and engages race-to-sleep techniques to place the engine or engines in a low power state during the forced idle times, thereby reducing the overall power draw of the GPU to meet a predetermined power budget. According to some embodiments, the WLM algorithm is deployed on a microcontroller of a power management unit (PMU).