Hybrid Computing Network Topology for Dynamic Bandwidth Routing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing network topologies for computing devices are inefficient in processing machine learning workloads due to imbalanced bandwidth distribution and fixed connections, leading to performance loss and inefficiencies.

Innovation Solution

A hybrid network topology that dynamically adjusts the ratio of direct and indirect connections based on workload demands, allowing flexible bandwidth allocation and repurposing indirect connections as direct connections within computing groups.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If a fat-tree topology is used with full bandwidth switches at base layers, then network bandwidth and routing capability are improved, but network cost and device complexity increase significantly

Engineering Contradiction:
Improvenetwork bandwidthVSAvoidswitch complexity and cost
Core Design Contradiction:
PowerVSDevice complexity

Solution Approach 1:

The patent implements a hybrid topology where the network dynamically switches between direct connections and indirect switch-routed connections based on workload requirements. This dynamic reconfiguration allows the system to achieve full bandwidth capability only when needed, rather than permanently provisioning expensive full-bandwidth switches at all base layers, thus resolving the contradiction between network bandwidth and device complexity/cost.

Inventive Principle:
Principle #15Dynamics

2Device complexity

If topology flattening is applied to reduce switch overhead, then network cost is reduced, but network performance and generality deteriorate

Engineering Contradiction:
Improveswitch overheadVSAvoidnetwork performance
Core Design Contradiction:
Device complexityVSPower

Solution Approach 1:

The hybrid topology design makes network connections universal by allowing indirect connections through switches to serve multiple purposes: they can function as regular routed connections for general traffic, and can be dynamically converted to direct connections for bandwidth-intensive workloads. This multi-functionality enables the network to maintain reduced switch overhead while preserving full performance capability when needed.

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

3Device complexity

If fixed connections are used in predetermined topologies, then network structure simplicity is improved, but adaptability to different workload bandwidth requirements deteriorates

Engineering Contradiction:
Improvenetwork structureVSAvoidworkload adaptability
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The system employs dynamic connection establishment where the network topology is not fixed but adapts based on workload characteristics. The hybrid architecture enables connections to be reconfigured from indirect to direct mode dynamically, allowing the network structure to adapt to different bandwidth requirements of various workloads while maintaining relative structural simplicity through the consistent hybrid design pattern.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentUS20250284529A1Hybrid Computing Network Topologies For Workload Execution
Publication Date: 2025.09.11 GOOGLE LLC
  • US20250284529A1 patent drawing
  • US20250284529A1 patent drawing
  • US20250284529A1 patent drawing

AI summary

Methods, systems, and apparatus, including computer-readable storage media for hybrid network topologies that flexibly support diverting bandwidth for indirect connections to bandwidth for direct connections for computing devices arranged in the hybrid network topology, on a workload-by-workload basis. The hybrid network topology described herein arranges computing devices into groups and connects neighboring computing devices to common switches based on the dimensionality of the computing device groups. Indirect connections connecting devices in different groups through a common switch can be at least partially diverted to form an additional direct connection between the neighboring computing devices in the same group. The network of nodes can divert bandwidth, for example based on executing different workloads with different bandwidth requirements for executing the workload locally, e.g., within a group of nodes, versus globally, e.g., across different groups of nodes.