SoC Fabric Peer-to-Peer Communication Ordering via Write Destination Register
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In System on a Chip (SoC) fabrics, ensuring correct peer-to-peer communication ordering is challenging, especially as interconnect architectures move beyond buses and hierarchical trees to more sophisticated structures like Networks-on-Chip, where traditional methods like centralized ordering points or non-posted writes lead to bottlenecks and performance issues.
Innovation Solution
Implementing a low overhead design with a Write Destination Register (WDR) in each network interface to ensure posted-write transaction ordering, using dummy writes to guarantee that writes from the same source agent arrive at their destinations in order, and allowing read requests to be sent directly once ordering is ensured.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If centralized ordering points or non-posted writes are used to ensure peer-to-peer communication ordering, then communication ordering is maintained, but communication delay increases and performance deteriorates due to bottlenecks
Solution Approach 1:
The patent segments the ordering function from a centralized point and distributes it to individual network interfaces. Each network interface maintains its own Write Destination Register (WDR) to track and enforce ordering locally, eliminating the need for centralized ordering control and reducing communication delays associated with centralized bottlenecks.
Solution Approach 2:
Each network interface serves itself by autonomously managing transaction ordering through its local WDR. The network interface independently determines whether to send dummy writes based on its own WDR state, without requiring external ordering control, thereby improving performance while maintaining correctness.
2Reliability
If dummy writes are sent to ensure posted-write transaction ordering, then write ordering is guaranteed, but communication overhead increases
Solution Approach 1:
The WDR is updated in advance with the destination address of the current posted write before the write is sent. This preliminary action enables the network interface to proactively determine whether a dummy write is needed, allowing ordering to be enforced efficiently without requiring post-send coordination or repeated checks.
Solution Approach 2:
The WDR provides feedback about the most recent posted-write destination to the network interface's ordering logic. This feedback mechanism enables intelligent decision-making about when dummy writes are necessary, reducing unnecessary communication overhead while ensuring ordering is maintained when required.
3Device complexity
If traditional interconnect architectures (buses, hierarchical trees) are used, then implementation is simpler, but wire congestion and timing closure issues increase as IP blocks are integrated
Solution Approach 1:
The ordering mechanism is implemented locally at each network interface rather than globally across the entire interconnect fabric. Each interface has its own WDR and ordering logic, allowing independent operation that is compatible with various fabric topologies (mesh, NoC, hybrid) without imposing global structural constraints, thus enabling scalability.
Solution Approach 2:
The patent changes the ordering approach from structural (centralized ordering points in traditional architectures) to state-based (WDR contents at each network interface). This parameter change allows the same ordering functionality to work across different fabric architectures, improving scalability and adaptability to modern SoC designs.
Data Source
AI summary
Methods and apparatus for efficient peer-to-peer communication support in interconnect fabrics. Network interfaces associated with agents are implemented to facilitate peer-to-peer transactions between agents in a manner that ensures data accesses correspond to the most recent update for each agent. This is implemented, in part, via use of non-posted “dummy writes” that are sent from an agent when the destination between write transactions originating from the agent changes. The dummy writes ensure that data corresponding to previous writes reach their destination prior to subsequent write and read transactions, thus ordering the peer-to-peer transactions without requiring the use of a centralized transaction ordering entity.