GPU Ethernet Switching Interconnect for Server Bandwidth Bottlenecks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The low bandwidth of PCIe buses and direct interconnections between GPUs limits data transmission efficiency and can lead to congestion, hindering the full potential of GPU computing power.

Innovation Solution

Implementing Ethernet switching chips on a mezzanine board connected via a designated connector to establish Ethernet connections between GPUs, enabling high-bandwidth data transmission using standard Ethernet protocols.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If PCIe buses or direct baseboard interconnections are used to connect GPUs, then the structure is simple and easy to implement, but the bandwidth is limited and data transmission efficiency is reduced

Engineering Contradiction:
Improveinterconnection structureVSAvoiddata transmission efficiency
Core Design Contradiction:
Device complexityVSProductivity

Solution Approach 1:

The patent introduces Ethernet switching chips as an intermediary device between GPUs. These switching chips are integrated on a mezzanine board that connects to the baseboard through a designated connector, providing a high-bandwidth Ethernet-based communication path that mediates data transmission between multiple GPUs while overcoming the bandwidth limitations of direct PCIe or baseboard interconnections

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent transitions from traditional PCIe bus interconnection (one-dimensional bus architecture) to Ethernet-based networking (multi-dimensional network architecture). By using Ethernet switching chips and standard Ethernet protocols, the system moves to a networked dimension that provides higher bandwidth and more flexible data transmission paths between GPUs

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Productivity

If Ethernet switching chips are integrated on a mezzanine board, then bandwidth and data transmission efficiency are improved, but the device complexity increases

Engineering Contradiction:
Improvedata transmission efficiencyVSAvoidinterconnection structure
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent uses standard Ethernet switching chips that support universal Ethernet protocols, allowing the same hardware component to serve multiple GPUs and provide flexible data transmission paths. The mezzanine board with Ethernet switching chips can be configured to support different interconnection topologies and protocols, providing multi-functionality that justifies the added complexity

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

Solution Approach 2:

The patent segments the Ethernet switching functionality from the main baseboard by placing it on a separate mezzanine board. This segmentation allows the Ethernet switching subsystem to be independently configured, maintained, and upgraded without affecting the core baseboard, thereby managing complexity through modular separation

Inventive Principle:
Principle #1Segmentation

3Productivity

If Ethernet connections are used between GPUs, then bandwidth is increased and congestion is reduced, but the baseboard area increases

Engineering Contradiction:
ImprovebandwidthVSAvoidbaseboard area
Core Design Contradiction:
ProductivityVSArea of stationary object

Solution Approach 1:

The patent nests the Ethernet switching chips on a mezzanine board that attaches to the existing baseboard structure. The mezzanine board utilizes the vertical space and existing connector infrastructure, effectively nesting the Ethernet switching functionality within the existing server form factor without requiring a complete redesign of the baseboard layout

Inventive Principle:
Principle #7Nested doll (Nesting)

Data Source

PatentEP4672004A1Method and device for interconnecting gpus of servers, a server, and a storage medium
Publication Date: 2025.12.31 NEW H3C AI TECH CO LTD
  • EP4672004A1 patent drawingFigure 1~2
  • EP4672004A1 patent drawingFigure 3~4
  • EP4672004A1 patent drawingFigure 5~6

AI summary

Disclosed are a graphics processing units, GPUs (101-108), interconnection method, device, and a storage medium. The device is applied to a server comprising two or more GPUs (101-108) interconnected with each other by an Ethernet switching chip (201, 202) to establish Ethernet connections, so that in the server, a source GPU (101-108) may determine that data needs to be transmitted to a target GPU (101-108), encapsulates the data to be transmitted based on a standard Ethernet protocol and obtains an Ethernet packet. The source GPU (101-108) sends the Ethernet packet to the Ethernet switching chip (201, 202); then the Ethernet switching chip (201, 202) sends the Ethernet packet to a designated connector. The designated connector sends the Ethernet packet to a destination GPU (101-108).