Embedding Lookup Coordinator for Sparse Table Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current data centre configurations face inefficiencies and high latency when handling high-demand machine learning workloads due to slow and inefficient sparse table lookup operations, which are exacerbated by the random and unpredictable nature of memory access patterns in embedding table lookups.

Innovation Solution

The implementation of an Embedding Lookup Coordinator (ELC) as a segregated software component that manages memory access and schedules lookup operations, centralizing embedding-table lookups and allocating dedicated resources to optimize performance, thereby reducing latency and improving throughput.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If sparse table lookup operations are performed using conventional memory access methods, then the embedding table lookups can be completed, but the random and unpredictable nature of memory access patterns causes slow and inefficient operations with high latency

Engineering Contradiction:
Improvetable lookup speedVSAvoidlatency
Core Design Contradiction:
SpeedVSLoss of time

Solution Approach 1:

The patent segments the embedding table into multiple shards or partitions that can be accessed in parallel. Instead of performing random access across the entire embedding table, the lookup operations are divided into smaller, more manageable segments that can be processed concurrently, reducing the overall access time and improving lookup speed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements preliminary actions by pre-computing and storing intermediate results, such as pre-calculating hash values or organizing the embedding table in a way that facilitates faster access. This allows the system to avoid performing expensive computations during the actual lookup operation, thereby reducing latency.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If embedding table lookups are performed separately from arithmetic computations, then system resources can be used more efficiently, but the system complexity increases due to the need for coordination and scheduling

Engineering Contradiction:
ImprovethroughputVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent introduces an intermediary component, such as a lookup coordinator or buffer management system, that mediates between the embedding table lookup operations and the arithmetic computation units. This intermediary manages the data flow, coordinates access to intermediate results, and schedules operations to maximize resource utilization while maintaining manageable system complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent extracts the embedding table lookup functionality from the main computation pipeline into a separate, dedicated subsystem. This extraction allows the lookup operations to be optimized independently and executed in parallel with arithmetic computations, improving throughput while the modular design keeps the overall system complexity manageable.

Inventive Principle:
Principle #2Taking out (Extraction)

3Reliability

If dedicated resources are allocated to the Embedding Lookup Coordinator, then lookup operations can be performed more efficiently with reduced latency, but the resource allocation overhead increases

Engineering Contradiction:
Improvelookup operation reliabilityVSAvoidresource allocation complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent designs the Embedding Lookup Coordinator to perform multiple functions using the same dedicated resources. The coordinator not only manages lookup operations but also handles scheduling, buffer management, and coordination with computation units. This multi-functionality improves lookup reliability while avoiding the need for separate specialized components for each function, thereby reducing resource allocation complexity.

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

Data Source

PatentUS12007974B2Memory processing optimisation
Publication Date: 2024.06.11 MYRTLE SOFTWARE LTD
  • US12007974B2 patent drawing
  • US12007974B2 patent drawing
  • US12007974B2 patent drawing

AI summary

Table lookup from hardware data storage is managed as part of processing one or more computer programs operating on a data processing system. At least one request is received from the one or more computer programs for a table lookup, at a software component that is operating on the data processing system separately or segregated from the one or more computer programs. The software component retrieves the table lookup data corresponding with the at least one request from the hardware data storage. The retrieved table lookup data is returned to the one or more computer programs.