Machine Learning Compute Architecture Using Fixed-Point Parallel Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
General purpose processors are inefficient in processing large loads, leading to delays in tasks such as machine learning, particularly in 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 efficiently, with a DMA unit and storage unit for data management.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general purpose processors are used to process information, then versatility is maintained, but processing speed and efficiency deteriorate under large loads
Solution Approach 1:
The processor is segmented into a control unit and multiple operation units with different data processing capabilities. The control unit parses instructions and routes them to appropriate operation units, while operation units specialize in specific data types (floating-point, fixed-point, integer). This segmentation allows the system to maintain versatility through instruction routing while achieving high efficiency through specialized parallel processing paths.
2Productivity
If general purpose processors operate at higher speeds, then processing efficiency improves, but power consumption and heat generation worsen
Solution Approach 1:
Different operation units are designed with local quality optimizations suited to their specific functions. Floating-point operation units use architectures optimized for decimal calculations, fixed-point units use binary optimization, and integer units use combinatorial logic. This local quality approach allows each unit to achieve high efficiency at lower power consumption by using the most energy-efficient implementation for its specific data type rather than a universal high-power design.
3Productivity
If more processing circuits are added to increase parallel processing capability, then productivity improves, but device complexity worsens
Solution Approach 1:
The control unit serves as a universal coordinator that handles instruction parsing, operation unit selection, and result aggregation for all operation units. This multi-functional control unit reduces the need for separate control logic in each operation unit, thereby reducing overall system complexity while enabling parallel processing across multiple specialized units.
4Speed
If data conversion to fixed-point format is performed, then processing speed improves, but manufacturing precision of data representation worsens
Solution Approach 1:
The system dynamically selects between floating-point and fixed-point operation units based on the instruction type and data characteristics. The control unit parses incoming instructions to determine the appropriate operation unit, allowing the system to adapt its precision and speed characteristics to the specific computational requirements. This dynamic selection enables the system to achieve high speed when using fixed-point units while maintaining precision when floating-point units are required.
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.