Dynamic CPU-Accelerator Offload for Power and Throughput Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques face challenges in minimizing power consumption overhead while mitigating performance bottlenecks caused by CPUs during high-load operations, either when processing is solely performed by CPUs or when offloading tasks to accelerators, without modifying application programs.
Innovation Solution
An accelerator offload device that dynamically switches processing between CPUs and accelerators based on data processing amounts, using a data processing amount acquisition part, operation destination determination part, and operation processing offload part to optimize power consumption and performance without altering the application program.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If processing is performed by CPU alone, then power consumption is reduced, but processing performance becomes bottlenecked during high-load operations
Solution Approach 1:
The system dynamically switches between CPU-only mode and accelerator offload mode based on real-time workload conditions. The determination unit monitors processing load and automatically transitions the system from using only the CPU to offloading to the accelerator when high-load conditions are detected, and reverses when load decreases, thereby adaptively optimizing the balance between power consumption and processing performance.
Solution Approach 2:
The system changes the operational parameters of the processing architecture by adjusting the degree of parallelism and the distribution of computational tasks between CPU and accelerator based on workload intensity. During high-load periods, the system increases parallel processing capacity by activating the accelerator; during low-load periods, it reduces parallelism to minimize power consumption, thus dynamically adapting system parameters to match demand.
2Productivity
If processing is offloaded to accelerator, then processing performance is improved, but power consumption overhead increases
Solution Approach 1:
The system employs dynamic workload management by continuously monitoring processing load and adjusting accelerator utilization accordingly. When workload exceeds predetermined thresholds, the system activates the accelerator to handle excess processing demands; when workload decreases below thresholds, the system deactivates or reduces accelerator usage, thereby maintaining high performance during peak loads while minimizing power consumption during low-activity periods.
Solution Approach 2:
The system applies partial offloading to the accelerator only when necessary to handle exceeding workload, rather than continuously utilizing the accelerator. By selectively offloading only the excess processing demands beyond CPU capacity and keeping the accelerator dormant or partially active during low-load periods, the system achieves performance improvement only when needed while avoiding unnecessary power consumption overhead.
3Productivity
If accelerator is always operating, then processing bottleneck is eliminated, but power consumption overhead occurs even when data processing amount is small
Solution Approach 1:
The system implements dynamic accelerator management by continuously monitoring workload intensity and adjusting accelerator operational state in real-time. When processing load is high, the accelerator is activated to eliminate bottlenecks; when processing load drops below predetermined thresholds, the accelerator is deactivated or placed in low-power state. This dynamic adjustment ensures the accelerator operates only when necessary to maintain performance, eliminating continuous power consumption overhead while preserving performance benefits during high-demand periods.
Data Source
AI summary
An accelerator offload device (100) includes: a data processing amount acquisition part (110) that acquires a data processing amount that is a data processing amount at the current time of offload target processing being performed and/or a predicted data processing amount; an operation destination determination part (120) that determines whether to perform an operation by a CPU (11) or offload the operation to an ACC (12) based on the data processing amount of an APL (1), and when a change is necessary, changes an offload destination of an operation processing offload part (130); and an operation processing offload part (130) that receives a request of operation processing from the APL (1), stores data to be processed and a processing instruction in a shared memory, and causes the CPU (11) or the ACC (12) to execute pertinent processing.