PCIe-Based CPU Cluster Communication System

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

High-speed information processing systems face limitations due to protocol overheads when using Ethernet for communication between CPUs in tightly-coupled clusters, which restricts processing speed.

Innovation Solution

The implementation of a PCIe-based information processing system that allows continuous data transmission without waiting for response from the destination, using PCIe interfaces and switches configured as root complexes, enabling direct address designation and transmission between semiconductor integrated circuits.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If Ethernet is used for communication between CPUs in tightly-coupled clusters, then cost is reduced and compatibility is improved, but processing speed is limited due to protocol overheads

Engineering Contradiction:
Improveprocessing speedVSAvoidprotocol overhead
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent extracts and removes Ethernet protocol overhead from the communication path by implementing direct PCIe-based communication between CPUs. The PCIe interface eliminates intermediate protocol processing layers, allowing raw data transmission without Ethernet framing, error checking, or flow control mechanisms, thereby achieving higher processing speeds while maintaining system compatibility through standard PCIe hardware.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent replaces the Ethernet protocol mechanism with a direct PCIe communication mechanism. Instead of using Ethernet's layered protocol stack (physical layer, data link layer, network layer), the system directly utilizes PCIe's simplified transaction protocol, substituting a complex mechanical protocol system with a more efficient direct hardware communication path that maintains compatibility through standard interfaces.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Productivity

If high-speed networks are used to enable fast communication between CPUs, then processing speed is improved, but system cost increases

Engineering Contradiction:
Improveprocessing speedVSAvoidsystem cost
Core Design Contradiction:
ProductivityVSEase of manufacture

Solution Approach 1:

The patent employs standard PCIe interfaces and switches that are widely available, cost-effective components rather than expensive specialized high-speed network hardware. By using commodity PCIe technology that can be rapidly manufactured and deployed, the system achieves high-speed communication without the prohibitive costs of specialized network equipment, making the solution economically viable for widespread implementation.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Solution Approach 2:

The patent leverages the universal PCIe interface standard that serves multiple functions: it provides high-speed communication between CPUs, enables memory access across different processors, and supports various device types. This multi-functional approach eliminates the need for separate specialized high-speed network hardware, reducing overall system cost while maintaining high performance through a single standardized interface.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If response waiting is implemented in Ethernet protocol, then data transmission accuracy is improved, but continuous data transmission speed is reduced

Engineering Contradiction:
Improvedata transmission accuracyVSAvoidtransmission waiting time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent implements preliminary error checking and data validation at the source before transmission begins, using PCIe's built-in error detection capabilities. This preliminary action ensures data accuracy is verified upfront, eliminating the need for intermediate response waiting during transmission. The system performs necessary validation in advance, allowing continuous high-speed transmission without pausing for accuracy verification.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent enables continuous data transmission through PCIe by eliminating Ethernet's stop-and-wait protocol mechanism. Data flows continuously through the PCIe interface without interruption for response acknowledgment, maintaining uninterrupted transmission streams. The system achieves continuous useful action by using PCIe's asynchronous transfer capability that doesn't require receiver acknowledgment between data packets, thereby eliminating waiting time while maintaining transmission reliability.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3550439B1Information processing system, semiconductor integrated circuit, and information processing method
Publication Date: 2022.03.30 SOCIONEXT INC
  • EP3550439B1 patent drawingFigure 1
  • EP3550439B1 patent drawingFigure 2
  • EP3550439B1 patent drawingFigure 3

AI summary

Processing by an information processing system is speeded up. A semiconductor integrated circuit (11) designates an address (a1) of a memory (15) connected to a semiconductor integrated circuit (12) that is a data transmission destination, based on memory map information (11m) in which addresses of memories (14 and 15) respectively used by semiconductor integrated circuits (11 and 12) are defined, converts the address (a1) to an address (a3) of the memory (15) defined in memory map information (12m) referred to by the data transmission destination, and outputs the address (a3) and transmission data by using a PCIe interface (11d). A switch (13) transfers the address (a3) and the transmission data to the data transmission destination by using PCIe interfaces (13a and 13b). The data transmission destination receives the address (a3) and the transmission data by using a PCIe interface (12d) and writes the transmission data into the reception buffer region of the memory (15) corresponding to the address (a3).