Neural Network Controller Input Data Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current artificial intelligence processors face inefficiencies due to not all operation units participating effectively in computations, leading to low operation performance and processing efficiency, as input data dimensions often require padding with zeros to match the count of operation units, resulting in unnecessary operations.
Innovation Solution
An input data processing method that determines the first and second data dimensions to align with the count of operation units, allowing for dimensionality reduction and completing input data accordingly, ensuring that all operation units participate in operations without unnecessary padding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If input data is padded with zeros to match the count of operation units, then all operation units can participate in operations, but unnecessary operations are performed reducing processing efficiency
Solution Approach 1:
The patent applies preliminary action by determining the actual data dimensions before processing and calculating which operation units should be activated in advance. This allows the system to prepare the exact set of operation units needed for the current input data, avoiding the need to pad data with zeros and subsequently perform unnecessary operations. The dimension determination and operation unit selection are performed beforehand, ensuring that only required operation units are activated and participate in computations.
2Productivity
If input data dimensions are increased to match operation unit count, then operation units are fully utilized, but data processing complexity increases
Solution Approach 1:
The patent applies segmentation by dividing the set of operation units into multiple groups based on the actual data dimensions. Instead of treating all operation units as a single block that must be fully activated, the system segments them according to the actual input data size, activating only the necessary segments. This segmentation approach maintains high operation unit utilization while avoiding the complexity of padding data to match a fixed operation unit count.
Solution Approach 2:
The patent implements dynamics by making the activation state of operation units variable rather than fixed. The system dynamically determines which operation units should be activated based on the actual dimensions of the input data. This dynamic approach allows the processing complexity to adapt to the actual data size, avoiding the static complexity introduced by zero-padding to match a predetermined operation unit count.
3Productivity
If data dimension is reduced to match operation units, then unnecessary operations are eliminated, but data alignment becomes more complex
Solution Approach 1:
The patent applies preliminary action by performing data alignment and operation unit selection in advance, before the actual computation begins. The system determines the actual data dimensions and calculates the corresponding operation unit configuration beforehand, creating a mapping between data elements and operation units. This preliminary preparation eliminates the need for complex runtime alignment operations while ensuring that only necessary operation units are activated.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to an operation method, an apparatus, and related products. The products include a controller unit. The controller unit includes an instruction caching unit, an instruction processing unit, and a storage queue unit. The instruction caching unit is used to store computation instructions associated with an artificial neural network operation; the instruction processing unit is used to parse the computation instructions to obtain a plurality of operation instructions; and the storage queue unit is used to store an instruction queue, where the instruction queue includes: a plurality of operation instructions or computation instructions to be executed in an order of the queue. By adopting the operation method, the present disclosure can improve the operation efficiency of related products when performing an operation of the neural network model.