Granular Power Gating Override for Compute Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing power gating methods for processing components, such as inference engines, incur significant latency due to collective power down and up operations of compute blocks, particularly microprocessors and memory devices, which fail to meet entry/exit latency timing requirements for lower power states.
Innovation Solution
Implementing a granular power gating override system that switches individual compute blocks from a first voltage rail to a second voltage rail during power gating, allowing them to remain powered on during low power states, thus avoiding context save/restore operations and reducing latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If collective power gating is applied to all compute blocks, then power consumption is reduced, but entry/exit latency increases significantly
Solution Approach 1:
The system segments the compute blocks into two categories: granular power-gatable blocks and non-granular power-gatable blocks. This segmentation allows selective power gating of only the granular blocks while keeping non-granular blocks powered on, thereby reducing overall power consumption without incurring the full latency penalty of gating all blocks.
Solution Approach 2:
Different power management strategies are applied to different compute blocks based on their characteristics. Granular compute blocks receive fine-grained power gating control while non-granular blocks maintain continuous power or use coarser power management, optimizing the balance between power savings and latency for each block type.
2Loss of time
If individual compute blocks are power gated independently, then entry/exit latency is reduced, but power management complexity increases
Solution Approach 1:
The system merges the power management of granular and non-granular compute blocks under a unified power management circuit that applies different power gating strategies to different block types. This consolidation manages complexity centrally while enabling fine-grained control where needed, avoiding the need for completely independent power management for each block.
Solution Approach 2:
The power management system dynamically adjusts its strategy based on which compute blocks need to be active. The power management circuit can switch between granular power gating for some blocks and coarser power management for others, adapting the level of granularity to the operational requirements of the system.
Data Source
AI summary
The disclosed device includes a processing component having various compute blocks, and a control circuit that switches at least one of the compute blocks from a normal voltage rail for the processing component to a second voltage rail in response to power gating a normal voltage rail. Various other methods, systems, and computer-readable media are also disclosed.