Dynamic Decimal Point Control in Fixed-Point Neural Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
General purpose processors are inefficient in processing large loads, leading to delays in machine learning tasks such as training artificial intelligence models, due to their limited operating speed.
Innovation Solution
A computation device and method that includes a controller unit, operation unit, and conversion unit, utilizing a primary processing circuit and secondary processing circuits to perform machine learning computations by converting input data into fixed-point data and executing operations to improve processing speed and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general purpose processors are used to process machine learning computations, then versatility and ease of operation are maintained, but processing speed and productivity deteriorate due to limited operating speed
Solution Approach 1:
The processor is segmented into a controller unit and multiple operation units (first operation unit, second operation unit, etc.). Each operation unit is specialized for specific computational tasks such as floating-point operations, fixed-point operations, or integer operations. This segmentation allows parallel processing across multiple units while maintaining system versatility through the controller's ability to dispatch different task types to appropriate units.
Solution Approach 2:
The system dynamically adapts its processing capabilities by converting data between different formats (floating-point to fixed-point, fixed-point to integer) based on the computational requirements of different neural network layers. The controller unit dynamically assigns tasks to different operation units based on real-time needs, enabling the system to optimize processing speed for each specific computational task while maintaining overall versatility.
2Ease of operation
If general purpose processors handle large computation loads, then ease of operation is maintained, but processing time increases excessively
Solution Approach 1:
Input data is pre-processed and converted into appropriate formats (floating-point, fixed-point, or integer) before being distributed to different operation units. This preliminary conversion eliminates the need for time-consuming format conversions during the actual computational process, significantly reducing total computational time while maintaining ease of operation through automated format management.
Solution Approach 2:
Multiple operation units perform computations continuously and in parallel without idle waiting time. While one unit processes floating-point data, another simultaneously processes fixed-point or integer data. This continuous parallel processing eliminates gaps in computational workflow, dramatically reducing total processing time for large machine learning workloads.
3Productivity
If data is converted to fixed-point format for processing, then processing speed improves, but manufacturing precision and measurement precision may be affected
Solution Approach 1:
Different operation units use different data formats optimized for their specific computational tasks. Floating-point operation units maintain high precision for operations requiring decimal accuracy, while fixed-point and integer units use faster formats for operations where approximate precision is sufficient. This local optimization of data quality matches precision requirements to computational needs, improving overall speed without unnecessarily sacrificing precision.
Solution Approach 2:
The system changes data representation parameters (format, precision, range) dynamically based on the computational requirements of different neural network layers and operations. The controller unit selects appropriate data formats and precision levels for each task, converting between formats as needed. This parameter adaptation allows the system to achieve high processing speeds while maintaining sufficient precision for each specific computational context.
Data Source
AI summary
The present disclosure provides a computation device. The computation device is configured to perform a machine learning computation, and includes an operation unit, a controller unit, and a conversion unit. The storage unit is configured to obtain input data and a computation instruction. The controller unit is configured to extract and parse the computation instruction from the storage unit to obtain one or more operation instructions, and to send the one or more operation instructions and the input data to the operation unit. The operation unit is configured to perform operations on the input data according to one or more operation instructions to obtain a computation result of the computation instruction. In the examples of the present disclosure, the input data involved in machine learning computations is represented by fixed-point data, thereby improving the processing speed and efficiency of training operations.