HBM Channel Repartitioning for Multi-Processor Memory Bandwidth
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-processor systems face inefficiencies due to static mapping of processor memory channels to high bandwidth memory (HBM) bus channels, leading to suboptimal performance and increased power consumption, particularly in scenarios where processor cores have varying memory access needs.
Innovation Solution
Implementing a reconfigurable partitioning system that dynamically maps processor memory channels to HBM bus channels using channel and segment multiplexers, allowing flexible allocation of HBM bus segments based on processor core needs, enabling dynamic reconfiguration of memory access bandwidth.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If static mapping of processor memory channels to HBM bus channels is used, then system simplicity is maintained, but system performance deteriorates due to suboptimal memory access
Solution Approach 1:
The patent implements dynamic mapping configuration that allows the system to adaptively allocate HBM bus channels to processor cores based on runtime memory access requirements. The mapping is no longer static but can be reconfigured dynamically through control elements that respond to varying workload demands, thereby improving memory access performance without permanently increasing system complexity.
Solution Approach 2:
The system changes the mapping parameters between processor memory channels and HBM bus channels based on performance requirements. By adjusting which processor cores are mapped to which HBM channels and modifying the number of channels allocated to each core, the system optimizes memory access performance while maintaining a manageable configuration complexity through controlled parameter variation.
2Device complexity
If static mapping of processor memory channels to HBM bus channels is used, then device complexity is reduced, but power consumption increases due to inefficient memory access patterns
Solution Approach 1:
The dynamic mapping configuration allows the system to adjust HBM channel allocation based on actual memory access needs of different processor cores. When certain cores have lower memory demands, their channel allocations can be reduced or reassigned, thereby decreasing power consumption while maintaining sufficient complexity only when optimization is required.
Solution Approach 2:
The system modifies mapping parameters to match actual workload requirements, avoiding the continuous high power consumption associated with static over-provisioning. By changing which channels are active and how they are mapped to cores based on real-time needs, the system reduces energy waste while keeping the configuration management complexity limited and controllable.
3Productivity
If dynamic reconfiguration of HBM bus channels is implemented, then memory access bandwidth is optimized for varying core needs, but device complexity increases
Solution Approach 1:
The HBM bus is segmented into multiple channels that can be independently allocated to different processor cores. This segmentation allows flexible reconfiguration where each channel can be assigned to different cores based on bandwidth requirements, optimizing memory access productivity while managing complexity through modular channel-level control rather than monolithic system reconfiguration.
Solution Approach 2:
The control elements implementing the mapping are designed to be universal and multi-functional, handling both the configuration and reconfiguration of HBM channel allocations. By using versatile control logic that can adapt to different mapping scenarios without requiring separate dedicated circuits for each configuration type, the system achieves high memory bandwidth optimization while limiting the increase in device complexity.
4Adaptability or versatility
If reconfigurable partitioning is implemented, then adaptability to varying memory access needs is improved, but ease of operation deteriorates due to configuration management
Solution Approach 1:
The reconfigurable mapping system is designed to be self-service in that it can automatically adapt its channel allocations based on detected memory access patterns and workload requirements. The control elements monitor system state and perform reconfiguration without requiring manual intervention, thereby maintaining high adaptability to varying needs while preserving ease of operation by eliminating complex manual configuration management.
Solution Approach 2:
The system incorporates feedback mechanisms that monitor memory access requirements of processor cores and use this information to automatically adjust HBM channel mappings. This feedback-driven approach enables the system to adapt flexibly to changing workload demands while simplifying operation, as the configuration management is handled automatically based on real-time system state information rather than requiring manual analysis and adjustment.
Data Source
AI summary
Various example embodiments of a capability for supporting reconfigurable partitioning of high bandwidth memory (HBM) in a multi-processor computing system are presented. The capability for supporting reconfigurable partitioning of HBM in a multi-processor computing system, where the multi-processor computing system includes a multi-processor system-on-chip (SoC) including a set of processors supporting respective sets of processor memory channels and an HBM interconnected via an HBM bus supporting a set of HBM bus channels, may support dynamic mapping of the processor memory channels of the processors of the multi-processor SoC to the HBM bus channels of the HBM, thereby overcoming various limitations of a static mapping of the processor memory channels of the processors of the multi-processor SoC to the HBM bus channels of the HBM. The dynamic processor-to-HBM channel mappings enable mapping of any of the processor memory channels of the processors of the multi-processor SoC to any of the HBM bus channels of the HBM.