Finite State Machine Tokenization for GPU Acceleration

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The process of tokenizing alphanumeric text, followed by embedding and text classification, is computationally expensive, particularly due to the inefficiency of tokenization on central processing units (CPUs) rather than graphics processing units (GPUs), which hinders the use of more complex and accurate text classifiers.

Innovation Solution

Implementing tokenization algorithms as finite state machines, allowing for the selection of appropriate algorithms based on the identified text classifier, thereby reducing processing time and increasing computational efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional tokenization algorithms are used, then text classification can be performed, but processing time increases significantly due to CPU inefficiency

Engineering Contradiction:

Improvetext classification speedVSAvoidtokenization time

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent replaces conventional CPU-based tokenization algorithms with GPU-accelerated finite state machine implementations. This substitution of the computational platform fundamentally changes how tokenization is performed, leveraging parallel processing capabilities of GPUs to achieve significantly faster tokenization speeds that match the speed of downstream embedding and classification operations.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent changes the implementation parameters of tokenization algorithms by transitioning from sequential CPU execution to parallel GPU execution. This parameter change in the computational platform enables the tokenization process to operate at speeds comparable to subsequent processing stages, thereby resolving the time loss issue.

Inventive Principle:

Principle #35Parameter changes

2Measurement precision

If more complex text classifiers are used, then classification accuracy improves, but computational expense increases

Engineering Contradiction:

Improveclassification accuracyVSAvoidcomputational expense

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent applies preliminary action by optimizing the tokenization stage through GPU acceleration before the data is passed to the classifier. By reducing the time and computational cost of tokenization through finite state machine implementations on GPUs, the system prepares data more efficiently, allowing complex classifiers to operate within acceptable computational budgets.

Inventive Principle:

Principle #10Preliminary action

3Speed

If tokenization is performed on CPU, then implementation is simple, but processing speed is slow

Engineering Contradiction:

Improvetokenization speedVSAvoidsystem complexity

Core Design Contradiction:

SpeedVSDevice complexity

Solution Approach 1:

The patent substitutes the CPU-based tokenization system with a GPU-based implementation. This replacement fundamentally changes the computational architecture, using the parallel processing power of GPUs to achieve high tokenization speeds. The finite state machine algorithms are reimplemented to leverage GPU hardware capabilities, thereby achieving the desired speed improvement.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS20250181833A1Tokenizing alphanumeric text through use of finite state machines

Publication Date: 2025.06.05 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

Described herein are technologies related to tokenizing alphanumeric text through use of a tokenization algorithm that is at least partially implemented as a finite state machine. The tokenization algorithm is configured to output numeric identifiers that represent tokens or sub-tokens in the alphanumeric text.