SoC Workload Allocation Across Accelerators, NICs, and Ports

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Balancing input/output (IO) data traffic across network paths in hardware processing systems is challenging, particularly in distributed computing platforms, leading to inefficiencies, increased latency, and reduced bandwidth due to continuous monitoring and adjustment of load balancing across multiple network ports.

Innovation Solution

A hierarchical load balancing method that divides computer workloads across hardware accelerators, network interface controllers (NICs), and network ports, ensuring even distribution and reducing the need for continuous monitoring, thereby improving system performance and bandwidth.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If continuous monitoring and adjustment of load balancing is performed across multiple network ports, then load distribution can be maintained, but system latency increases and bandwidth decreases

Engineering Contradiction:
Improveload distributionVSAvoidsystem latency
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-calculating and storing optimal load balancing decisions in lookup tables before runtime. The load balancing controller pre-computes distribution strategies based on historical data and system characteristics, then directly retrieves pre-determined routing decisions during operation without real-time calculation, thereby maintaining reliable load distribution while eliminating continuous monitoring overhead and reducing latency

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating simplified models of complex load balancing scenarios and storing them as lookup tables. Instead of performing complex real-time calculations, the system copies pre-analyzed optimal solutions into easily accessible data structures that can be queried instantly, trade-offing some memory usage for dramatically reduced processing time and eliminated continuous monitoring

Inventive Principle:
Principle #26Copying

2Reliability

If continuous monitoring and adjustment of load balancing is performed across multiple network ports, then load distribution can be maintained, but bandwidth is reduced

Engineering Contradiction:
Improveload distributionVSAvoidbandwidth
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies preliminary action by pre-calculating and storing optimal load balancing decisions in lookup tables before runtime. The load balancing controller pre-computes distribution strategies based on historical data and system characteristics, then directly retrieves pre-determined routing decisions during operation without real-time calculation, thereby maintaining reliable load distribution while eliminating continuous monitoring overhead and reducing latency

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses copying by creating simplified models of complex load balancing scenarios and storing them as lookup tables. Instead of performing complex real-time calculations, the system copies pre-analyzed optimal solutions into easily accessible data structures that can be queried instantly, trade-offing some memory usage for dramatically reduced processing time and eliminated continuous monitoring

Inventive Principle:
Principle #26Copying

3Productivity

If workload is distributed across multiple hardware accelerators and NICs, then processing capacity increases, but system complexity increases

Engineering Contradiction:
Improveprocessing capacityVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies segmentation by dividing the complex workload distribution task into distinct hierarchical layers: the load balancing controller segment handles high-level accelerator assignment, while individual NICs segment the network transmission tasks. This hierarchical segmentation allows each component to manage a simplified subset of the overall complexity, enabling scalable processing capacity without linearly increasing system complexity

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses an intermediary approach by introducing a dedicated load balancing controller that acts as a mediator between the workload source and multiple hardware accelerators. This intermediary centralizes the complex decision-making logic for workload distribution, while downstream NICs handle simplified network-specific tasks, thereby enabling high processing capacity through clear separation of concerns and reduced inter-component complexity

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4664287A1Computer workload allocation for hardware processing system
Publication Date: 2025.12.17 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4664287A1 patent drawingFigure 1
  • EP4664287A1 patent drawingFigure 2
  • EP4664287A1 patent drawingFigure 3

AI summary

A method (200) for computer workload allocation at a system-on-chip (SoC) (100) includes, at a load balancing controller (102) of the SoC (100), dividing a computer workload (104) for distributed processing between each of a plurality of hardware accelerators (106) of the SoC (100) as a plurality of accelerator-specific data allocations (108). At a hardware accelerator of the plurality of hardware accelerators (106), after receiving an accelerator-specific data allocation (108) from the load balancing controller (102), a resulting dataset output by the hardware accelerator (106) is divided between a plurality of network interface controllers (NICs) (110) of the SoC (100) as a plurality of NIC-specific data allocations (112). At an NIC of the plurality of NICs (110) an NIC-specific data allocation (112) assigned to the NIC (110) is divided between a plurality of network ports (114) of the NIC (110) for transmission over a computer network (118).