Remote FIFO Management in Reconfigurable Dataflow Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing reconfigurable computing systems face challenges in efficiently managing inter-die communication between reconfigurable dataflow units, leading to inefficiencies in data transfer and processing.
Innovation Solution
Implementing a peer-to-peer (P2P) communication protocol over the transaction layer packet (TLP) of PCIe to facilitate direct data transfer between reconfigurable dataflow units, utilizing a packet-switched mesh network with credit management for efficient data flow and address translation mechanisms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If direct peer-to-peer communication is implemented between reconfigurable dataflow units, then data transfer efficiency is improved, but device complexity increases due to additional communication protocols and network infrastructure
Solution Approach 1:
The patent introduces a packet-switched mesh network as an intermediary communication infrastructure between reconfigurable dataflow units. This mesh network acts as a mediator that enables direct peer-to-peer communication without requiring complex point-to-point wiring between every pair of units. The network layer provides standardized protocols and routing mechanisms that simplify the communication complexity while maintaining high data transfer efficiency.
Solution Approach 2:
The communication system is segmented into distinct layers: the physical mesh network infrastructure, the packet switching layer, and the application layer. This segmentation allows each layer to be optimized independently - the mesh network handles physical connectivity, the packet layer manages data flow and routing, and the application layer focuses on data processing. This reduces overall system complexity by distributing communication functions across multiple specialized components.
2Device complexity
If centralized host processor management is used for inter-die communication, then device complexity is reduced, but data transfer efficiency and processing speed deteriorate
Solution Approach 1:
The patent extracts the communication management function from the centralized host processor and relocates it to the mesh network infrastructure itself. The mesh network includes dedicated routing logic, packet switching capabilities, and credit management mechanisms that operate autonomously without requiring host processor intervention. This extraction enables dataflow units to communicate directly through the network, eliminating the performance bottleneck of centralized management while the network layer provides the necessary coordination functions.
Solution Approach 2:
The reconfigurable dataflow units are equipped with autonomous communication capabilities including packet generation, addressing, and protocol handling. Each unit can independently initiate and manage communications with other units through the mesh network without requiring host processor mediation. The network itself provides self-management through routing algorithms and flow control mechanisms, enabling the system to serve its own communication needs without external control.
Data Source
AI summary
A reconfigurable processing unit includes a first and second internal network, an interface to an external network, a first configurable unit coupled to the first internal network, a second configurable unit coupled to both internal networks, and a third configurable unit coupled to both the second internal network and the interface to the external network. The third configurable unit is configured to receive a payload containing a transaction type identifier and an identifier of the second configurable unit through the interface to the external network, and send a first packet including the transaction type identifier to the second configurable unit over the second internal network. The second configurable unit is configured to increment a counter in response to a particular transaction type identifier, and send a token to the first configurable unit over the first internal network while the counter is non-zero and the first configurable unit is executing.