Dynamic Clock Frequency Management for GPU Power Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High Performance Computing (HPC) server graphics processing units (GPUs) face challenges in dynamically optimizing performance across varying workloads due to static tuning assumptions, which do not account for workload behavior changes over time, and require dynamic allocation of power between I/O and compute subsystems based on runtime telemetry.
Innovation Solution
A system that monitors runtime telemetry to determine if a GPU is I/O bounded or compute bounded, allowing for dynamic allocation of power between the I/O subsystem and the compute subsystem, adjusting clock frequencies of I/O and compute chiplets based on workload characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If static tuning is used to configure the system for each type of workload, then performance is improved under specific workload conditions, but the system cannot adapt when workload behavior changes over time
Solution Approach 1:
The patent implements dynamic frequency adjustment for I/O and compute circuitries based on runtime telemetry data. The system continuously monitors workload characteristics and adjusts clock frequencies accordingly, transitioning from static to dynamic configuration. This allows the system to adapt to changing workload conditions in real-time, resolving the contradiction between optimized performance for specific workloads and adaptability to workload changes.
Solution Approach 2:
The system employs feedback mechanisms by monitoring runtime telemetry to determine whether the GPU is I/O bounded or compute bounded. Based on this feedback, the system dynamically calculates and adjusts power distribution to I/O and compute subsystems, enabling continuous optimization of performance as workload conditions change.
2Productivity
If power is allocated statically to I/O and compute subsystems, then power management is simplified, but performance optimization under varying workloads is limited
Solution Approach 1:
The system uses runtime telemetry feedback to dynamically determine whether the GPU is I/O bounded or compute bounded, automatically adjusting power allocation between subsystems. This feedback-driven approach enables performance optimization under varying workloads while keeping the control logic relatively simple, as the system automatically adapts based on monitored conditions.
Solution Approach 2:
The system performs self-service by autonomously monitoring its own performance characteristics through telemetry and automatically adjusting power distribution without external intervention. The GPU driver calculates the distribution of power to I/O and compute subsystems based on runtime conditions, enabling the system to optimize its own performance dynamically.
3Speed
If clock frequencies are increased for both I/O and compute chiplets, then processing speed is improved, but power consumption increases beyond power limits
Solution Approach 1:
The patent applies local quality by selectively adjusting clock frequencies of specific circuitries (I/O or compute) based on their current performance needs. Instead of uniformly increasing frequencies across all components, the system dynamically allocates higher frequencies only to the subsystem that is currently performance-bounded, while maintaining lower frequencies in other subsystems. This localized frequency adjustment optimizes processing speed while constraining overall power consumption within limits.
Solution Approach 2:
The system dynamically changes operational parameters (clock frequencies) of different circuitries based on runtime conditions. By calculating the distribution of power and adjusting frequencies according to whether the system is I/O bounded or compute bounded, the system optimizes processing speed while maintaining power consumption within acceptable limits.
Data Source
AI summary
A system that includes first circuitries to operate at a first clock frequency, second circuitries to operate at a second clock frequency, and circuitry to adjust the first and second clock frequencies. In some examples, the circuitry is to selectively adjust the first and second clock frequencies provided to the respective first circuitries and the second circuitries according to a target ratio based on temperature and power consumption of the first circuitries and the second circuitries, wherein the target ratio is based on clock frequencies of the first circuitries and the second circuitries, stall time of the first circuitries, and dynamic capacitance of the first circuitries and the second circuitries.