Data Warehouse Engine Aggregates Memory Access
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radio access network (RAN) systems-on-a-chip (SoCs) face inefficiencies due to non-uniform memory access patterns, leading to CPU and DSP management of memories, causing stalls and overdesigning, which limits scalability and increases power and cost.
Innovation Solution
A data warehouse engine (DWE) with a central processing unit (CPU) core and data organization units (DOUs) that aggregate read and write operations, perform on-the-fly data reorganization, and utilize elastic buffers and write managers to optimize memory I/O utilization, reducing non-uniform access disruptions and enhancing scalability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If memories are treated as dumb storage and managed by CPUs and DSPs, then memory management flexibility is improved, but compute engine stalls due to non-uniform memory access increase
Solution Approach 1:
A data organization unit (DOU) is introduced as an intermediary component between the compute engines and the memory system. The DOU includes a data warehouse that acts as a buffer and reorganization layer, converting non-uniform memory access patterns into uniform access patterns for the actual memory, thereby eliminating compute engine stalls while maintaining management flexibility.
Solution Approach 2:
The memory system is segmented into multiple memory banks, each managed by a separate data organization unit. This segmentation allows parallel processing of different data sets and prevents a single non-uniform access pattern from stalling the entire compute engine, as each DOU can independently handle its assigned memory bank.
2Reliability
If huge multi-ported memories are used to handle worst case random stalls, then memory access reliability is improved, but device complexity and scalability worsen
Solution Approach 1:
The data organization unit serves as an intermediary that absorbs the complexity of handling non-uniform access patterns. Instead of requiring the memory itself to be complex and multi-ported, the DOU handles the reorganization logic, allowing the use of simpler, more scalable memory structures while maintaining reliable access patterns.
Solution Approach 2:
The data organization unit creates organized copies of data in its internal data warehouse before writing to memory. This copying mechanism allows the system to present uniform access patterns to memory while handling non-uniform access requirements from compute engines, reducing memory complexity while maintaining reliability.
3Adaptability or versatility
If CPUs and DSPs directly manage memories, then control flexibility is improved, but power consumption and cost increase
Solution Approach 1:
The data organization unit acts as a specialized intermediary that offloads the power-intensive task of memory management from general-purpose CPUs and DSPs. The DOU's dedicated hardware structure is more energy-efficient for memory reorganization tasks while maintaining the flexibility needed for different access patterns.
Solution Approach 2:
The data organization unit provides self-service functionality by autonomously managing memory reorganization and access patterns without requiring continuous intervention from CPUs or DSPs. This autonomous operation reduces the power consumption associated with having general-purpose processors constantly manage memory operations.
Data Source
AI summary
A data warehouse engine (DWE) includes a central processing unit (CPU) core and a first data organization unit (DOU), where the first DOU is configured to aggregate read operations. The DWE also includes a first command queue coupled between the CPU core and the first DOU, where the first command queue is configured to convey commands from the CPU core to the first DOU.