PCIe TLP Peer-to-Peer Routing for Reconfigurable Processors

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing reconfigurable processors face challenges in efficiently and efficiently communicating between integrated circuits, particularly across different dies, which hampers the performance of complex computations like machine learning and artificial intelligence workloads.

Innovation Solution

Implementing a peer-to-peer (P2P) communication protocol over the transaction layer packet (TLP) of PCIe, enabling direct data transfer between reconfigurable dataflow units (RDUs) through encapsulating P2P messages in TLP payloads, and utilizing Address Generation and Coalescing Units (AGCUs) for address translation and data routing.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional host-processed communication is used between reconfigurable dataflow units, then system control and coordination are maintained, but dataflow bandwidth and computational efficiency are limited

Engineering Contradiction:

Improvedataflow bandwidthVSAvoidcommunication protocol complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The communication system is segmented into multiple independent components: source RDUs generate transactions, target RDUs receive and process them, and AGCUs handle address translation. This segmentation allows each component to specialize in specific functions, improving overall dataflow bandwidth while maintaining manageable complexity through modular design.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

AGCUs (Address Generation and Coalescing Units) act as intermediaries between source and target RDUs, handling address translation and routing. This intermediary mechanism enables efficient peer-to-peer communication without requiring complex centralized control, thereby improving dataflow bandwidth while keeping the protocol manageable through automated address handling.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Productivity

If peer-to-peer communication is implemented between reconfigurable dataflow units, then dataflow bandwidth is enhanced, but protocol implementation complexity increases

Engineering Contradiction:

Improvecomputational efficiencyVSAvoidprotocol implementation complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The TLP (Transaction Layer Packet) protocol is designed to be universal, handling multiple communication scenarios (local and remote transactions, different data types) through a single unified mechanism. This multi-functionality improves computational efficiency by eliminating the need for multiple specialized protocols while keeping implementation complexity manageable through standardization.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system dynamically adjusts communication parameters based on transaction types and target locations. By changing parameters such as address translation rules and routing information according to the specific communication scenario, the system achieves high computational efficiency without requiring complex hardwired protocol implementations for each case.

Inventive Principle:

Principle #35Parameter changes

3Extent of automation

If direct peer-to-peer data transfer is enabled, then reliance on host processor is reduced, but address translation and routing complexity increases

Engineering Contradiction:

Improveautonomous communication capabilityVSAvoidaddress translation complexity

Core Design Contradiction:

Extent of automationVSDevice complexity

Solution Approach 1:

Source RDUs autonomously generate transactions with appropriate addresses and routing information without requiring host processor intervention. Target RDUs autonomously receive and process transactions at their designated locations. This self-service capability improves autonomous communication capability while managing address translation complexity through distributed decision-making at each RDU.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

AGCUs perform address translation and routing preparation in advance, before data transfer begins. By pre-processing address information and establishing routing paths ahead of time, the system enables autonomous peer-to-peer communication without requiring complex real-time decision-making, thereby improving automation while managing translation complexity through preliminary processing.

Inventive Principle:

Principle #10Preliminary action

4Adaptability or versatility

If reconfigurable processors communicate across multiple dies, then system scalability is improved, but inter-die communication overhead increases

Engineering Contradiction:

Improvesystem scalabilityVSAvoidcommunication overhead

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The communication system operates across multiple spatial dimensions (different dies) using a unified protocol that handles local and remote transactions similarly. By extending the communication model to accommodate multi-die architecture through standardized address translation and routing, the system achieves improved scalability while managing communication overhead through consistent handling of cross-die transactions.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS20250384007A1Transaction type identifier-based payload transmission in a reconfigurable processor

Publication Date: 2025.12.18 SAMBANOVA SYSTEMS INC

AI summary

A reconfigurable processor. The reconfigurable processor comprising an array of configurable units. Configurable units in the array of configurable units configured to transmit payloads between each other based on a transaction type identifier.