Method for transmitting weight data, many-core system, electronic device, medium
By storing the weight data of sparse neurons in off-chip memory and storing the weight data of non-sparse neurons in the processing kernel, the problem of large storage space occupation in neural networks is solved, achieving storage space saving and transmission cost reduction, and improving the performance of many-core systems.
CN114792128BActive Publication Date: 2026-06-26LYNXI TECH CO LTD
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LYNXI TECH CO LTD
- Filing Date
- 2022-04-29
- Publication Date
- 2026-06-26
AI Technical Summary
Technical Problem
The amount of data connecting the weights in a neural network is large, occupying a lot of storage space in many-core systems, and existing technologies are difficult to manage and transmit efficiently.
Method used
The weight data of sparse neurons is stored in off-chip memory, while the weight data of non-sparse neurons is stored in the processing kernel. Weight data is acquired and transmitted in different ways to reduce storage and transmission costs.
Benefits of technology
It saves storage space within the many-core system, reduces the cost of weighted data transfer, and improves the overall performance of the system.
✦ Generated by Eureka AI based on patent content.
Smart Images
Figure CN114792128B_ABST
Abstract
The present disclosure provides a method for weight data transmission, comprising: in response to a firing of a sparse current neuron, obtaining first weight data of the current neuron from an off-chip memory, and obtaining all non-zero connection weight values between the current neuron and each successor neuron of the current neuron from a target processing core where each successor neuron is located; in response to a firing of a non-sparse current neuron, obtaining second weight data of the current neuron from within the processing core, and obtaining all non-zero connection weight values between the current neuron and each successor neuron of the current neuron from a target processing core where each successor neuron is located; the sparse neuron is a neuron with a firing rate lower than a preset standard, and the non-sparse neuron is a neuron with a firing rate higher than or equal to the preset standard; the weight data of a neuron at least comprises all non-zero connection weight values between the neuron and each successor neuron. The present disclosure also provides a many-core system, an electronic device and a computer readable medium.
Need to check novelty before this filing date? Find Prior Art