Parallel Computation Pipeline Configuration for FPGA Throughput
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
FPGAs are more expensive and less efficient than ASICs, despite offering reconfigurability, and existing methods for distributing computations in parallel processing environments often result in excessive synchronization and communication dependencies among computation units.
Innovation Solution
A method for configuring a system of interconnected computation units by identifying instruction subsets, determining dependencies, and associating them to form a pipeline, which reduces cyclic communication dependencies and enables efficient data communication through an interconnection network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If computations are distributed across multiple computation units in a parallel processing environment, then productivity is improved, but device complexity increases due to synchronization requirements
Solution Approach 1:
The computation is divided into distinct instruction subsets that are assigned to different computation units. Each computation unit processes a specific segment of the computation independently, reducing the need for synchronization between units while maintaining overall productivity.
Solution Approach 2:
Instruction subsets are pre-analyzed and assigned to computation units before execution begins. Dependency relationships are identified in advance, allowing the system to establish a predetermined execution plan that minimizes runtime synchronization requirements.
2Productivity
If instruction subsets are assigned to different computation units to form a pipeline, then productivity is improved, but difficulty of detecting and measuring increases due to communication dependencies
Solution Approach 1:
The system performs preliminary analysis of instruction dependencies and communication requirements before assigning instruction subsets to computation units. This pre-characterization of communication patterns simplifies the detection and measurement of dependencies during actual execution.
Solution Approach 2:
The patent introduces an intermediary layer that manages communication between computation units. This intermediary handles the complexity of communication dependencies, making it easier to detect and measure actual data flows without directly analyzing complex inter-unit communications.
3Adaptability or versatility
If existing methods are used to distribute computations in parallel processing environments, then adaptability is improved, but loss of time increases due to excessive synchronization
Solution Approach 1:
By segmenting computations into independent instruction subsets that can be assigned to different computation units, the system reduces synchronization points while maintaining the ability to reconfigure computation distributions. This segmentation allows adaptable parallel execution with minimal time loss.
Solution Approach 2:
The system dynamically assigns instruction subsets to computation units based on identified dependencies and available resources. This dynamic assignment optimizes the balance between adaptability and synchronization overhead, allowing the system to reconfigure computations efficiently without excessive time loss.
Data Source
AI summary
A system comprises a plurality of computation units interconnected by an interconnection network. A method for configuring the system comprises accepting a set of instructions corresponding to a portion of a program that performs a computation repeatedly; identifying subsets of the instructions; and associating each subset with a different one of the computation units to form a specification of the set of instructions such that execution according to the specification forms a pipeline among at least some of the computation units.