Difference between revisions of "Hybrid Compute-in-Memory: abstract simplified (17695824)"

From WikiPatents
Jump to navigation Jump to search
(Creating a new page)
(Creating a new page)
 
(One intermediate revision by the same user not shown)
Line 1: Line 1:
A compute-in-memory array is a technology that combines computing and memory functions in a single chip. In this case, the array is used to implement a filter for a specific layer in a neural network.
+
A compute-in-memory array is described in the abstract. This array is designed to implement a filter for a specific layer in a neural network. The purpose of this filter is to process data by multiplying a set of activation bits with a set of filter weight bits. This multiplication process is performed for each channel within a group of channels. The resulting products are then accumulated through a charge accumulation process using multiple capacitors. Finally, the accumulated charge is digitized to produce the output of the filter.
 
 
The filter operates on a set of activation bits and filter weight bits. These bits represent the input data and the weights associated with the filter, respectively. The filter performs a multiplication operation between each activation bit and its corresponding filter weight bit.
 
 
 
The filter is designed to process multiple channels of data simultaneously. Each channel represents a different aspect or feature of the input data. The filter applies the multiplication operation to each channel independently.
 
 
 
To perform the multiplication, the filter utilizes a charge accumulation mechanism. This mechanism involves a set of capacitors that store and accumulate electrical charge. The activation bits and filter weight bits are used to control the charging and discharging of these capacitors.
 
 
 
Once the charge accumulation process is complete, the accumulated charge is digitized. This means that the analog charge values are converted into digital values. The digitized output represents the result of the filter operation for the given input data.
 
 
 
Overall, this compute-in-memory array provides an efficient and integrated solution for implementing a filter in a neural network layer. It leverages capacitors for charge accumulation and digitization to achieve the desired filtering functionality.
 

Latest revision as of 16:20, 1 October 2023

A compute-in-memory array is described in the abstract. This array is designed to implement a filter for a specific layer in a neural network. The purpose of this filter is to process data by multiplying a set of activation bits with a set of filter weight bits. This multiplication process is performed for each channel within a group of channels. The resulting products are then accumulated through a charge accumulation process using multiple capacitors. Finally, the accumulated charge is digitized to produce the output of the filter.