Dual-Tier Optical Switching for Scalable GPU Cluster Bandwidth

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional optical virtual-circuit switching network systems face challenges in scaling beyond small- to medium-sized HPC data centers due to insufficient bandwidth and excessive manufacturing costs of optical switches, which become non-linear with increased GPU server racks.

Innovation Solution

A cluster-based distributed optical virtual-circuit-switching network system with a dual-tier architecture, comprising a tier-1 optical switching network and a tier-2 optical switching network, allowing for flexible scaling and efficient inter-cluster data transmission using wavelength reselection and uniform optical switch designs.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If the number of optical switches is increased to support more GPU server racks in a single-tier network topology, then the bandwidth capacity is improved, but the manufacturing cost increases non-linearly

Engineering Contradiction:
Improvebandwidth capacityVSAvoidmanufacturing cost
Core Design Contradiction:
Quantity of substanceVSEase of manufacture

Solution Approach 1:

The patent divides the single-tier network topology into multiple two-tier network subsystems, each containing a first-type optical switching network subsystem (cluster) and a second-type optical switching network subsystem. This segmentation allows the system to scale bandwidth capacity by adding more subsystems rather than increasing the size of individual switches, thereby avoiding non-linear cost increases.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from a single-tier to a two-tier network architecture, adding a vertical dimension to the network topology. The first-type optical switching network subsystems operate at one tier while the second-type subsystems operate at another tier, enabling scalable bandwidth expansion without proportionally increasing manufacturing costs of individual optical switches.

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

2Productivity

If the number of optical switches is increased to support large-scale GPU communication, then the data transmission capability is improved, but the device complexity increases

Engineering Contradiction:
Improvedata transmission capabilityVSAvoiddevice complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent segments the optical switching functionality into distributed first-type and second-type optical switching network subsystems. Each subsystem handles specific routing functions, distributing the complexity across multiple manageable units rather than concentrating it in fewer larger switches, thus improving data transmission capability while keeping individual device complexity manageable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements dynamic routing capabilities within the two-tier architecture, where optical signals can be flexibly routed through different paths involving first-type and second-type optical switching network subsystems. This dynamic routing optimizes data transmission capability while maintaining manageable device complexity through adaptive path selection.

Inventive Principle:
Principle #15Dynamics

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

The system supports large-scale inter-cluster data transmission with reduced production costs and enhanced flexibility, maintaining high bandwidth and low latency while optimizing routing options.

Implementation Method 1

Between the tier-1 and tier-2 optical switching network, optical-electrical-optical (O/E/O) conversion is carried out to enable wavelength reselection, thereby facilitating signal transmission across different clusters.

Methodology Applied
Scientific EffectOptical-electrical-optical conversion:

Data Source

PatentUS20260075343A1Cluster-based distributed optical virtual-circuit-switching network system
Publication Date: 2026.03.12 GENOPSYS TECH INC
  • US20260075343A1 patent drawing
  • US20260075343A1 patent drawing
  • US20260075343A1 patent drawing

AI summary

A cluster-based distributed optical virtual-circuit-switching network system comprises a tier-1 optical switching network and a tier-2 optical switching network. The tier-1 optical switching network includes a plurality of first-type optical switching network subsystems, each defining a cluster. The tier-2 optical switching network includes at least one second-type optical switching network subsystem, which comprises a plurality of tier-2 optical switches interconnected with one another. Each of the tier-2 optical switches correspondingly connects to one of the first-type optical switching network subsystems. When optical signals are transmitted between the clusters, the signals are transmitted from one of the first-type optical switching network subsystems to another through the tier-2 optical switches of the at least one second-type optical switching network subsystem.