PIM Weight Arrangement Pattern for Convolution Row Skipping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing PIM architectures face challenges in maximizing the row-skipping ratio during convolution operations due to irregular weight distribution and increasing channel sizes, leading to lower probabilities of skipping rows with zero values, which hampers energy efficiency and computation speed.
Innovation Solution
A PIM control device and method that determines a weight arrangement pattern for the kernel considering the PIM array structure and weight sparsity, allowing for optimized mapping of weights to memory cells and skipping of convolution operations in rows with zero weights, thereby enhancing row-skipping ratios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If row-skipping technique is applied to optimize computation speed, then operation speed is improved, but the row-skipping ratio is limited due to irregular weight distribution and increasing channel sizes
Solution Approach 1:
The patent applies preliminary action by reordering weights into the PIM array before convolution operations are executed. Specifically, weights are rearranged according to a predetermined pattern that groups zero-weight rows together, enabling the row-skipping technique to efficiently skip entire rows of operations in subsequent convolution steps, thereby achieving both high operation speed and high row-skipping ratio
Solution Approach 2:
The patent changes the arrangement parameter of weights in the PIM array from conventional random or standard ordering to a specific predetermined pattern. This parameter change reorganizes the weight matrix such that rows with zero weights are concentrated together, allowing the row-skipping technique to maximize the row-skipping ratio while maintaining high computation speed during convolution operations
2Quantity of substance
If channel size of weight is increased to handle more data, then data processing capability is improved, but the probability of row skipping decreases due to more rows being used
Solution Approach 1:
The patent applies preliminary action by reordering weights into the PIM array before convolution operations are executed. Specifically, weights are rearranged according to a predetermined pattern that groups zero-weight rows together, enabling the row-skipping technique to efficiently skip entire rows of operations in subsequent convolution steps, thereby achieving both high operation speed and high row-skipping ratio
Solution Approach 2:
The patent changes the arrangement parameter of weights in the PIM array from conventional random or standard ordering to a specific predetermined pattern. This parameter change reorganizes the weight matrix such that rows with zero weights are concentrated together, allowing the row-skipping technique to maximize the row-skipping ratio while maintaining high computation speed during convolution operations
Data Source
AI summary
There is provided a processing-in-memory (PIM) control device. The device comprises a memory configured to store one or more instructions; and a processor configured to execute the one or more instructions stored in the memory, wherein the instructions, when executed by the processor, cause the processor to determine input data, weights, and information on a size of a kernel, which is a unit for performing convolution operations using the input data and the weights, determine a pattern for arranging each weight in the kernel, arrange first weights corresponding to the determined pattern in the kernel to map each first weight to a plurality of memory cells included in a PIM array, and control the PIM array to perform the convolution operations using first input data corresponding to the size of the kernel and the first weights.