US 12,327,591 B2Granted

Method and apparatus for performing a MAC operation in a memory array

AssigneeMediaTek Singapore Pte. Ltd.

InventorsChetan Deshpande, Gajanan Sahebrao Jedhe, Cheng-Xin Xue, Zijie Guo

Life of this patent

Estimating the term…

Loading patent…

Abstract

Aspects of the present disclosure are directed to devices and methods for performing MAC operations using a TCAM array as a compute-in-memory (CIM) device that can enable higher computational throughput, higher performance and lower energy consumption compared to computation using a processor outside of a memory array. In some embodiments, weights in a weight matrix may be programmed in SRAMs of a TCAM bit cell array. Each SRAM may operate as a multiplier that performs a multiplication between the stored weight to an input activation value applied at a search line in the TCAM bit cell array. The two SRAMs within a TCAM bit cell may operate independently to receive independently two input activation values on their respective select lines, and to perform a multiplication operation with the stored weight in each respective SRAM.

Claims20

Dependent claims

A compute-in-memory device, comprising:

a memory array comprising a plurality of ternary content-addressable memory (TCAM) bit cells arranged in a plurality of rows, each TCAM bit cell comprising two static random-access memories (SRAMs), each SRAM configured to apply a weight stored therein to an input activation to generate an output value, wherein the plurality of TCAM bit cells of the compute-in-memory device are configured to store weights in a weight matrix of a neural network, each SRAM of the compute-in-memory device operates as a multiplier that performs an in-memory multiplication between the weight stored therein and the input activation, a first input activation that is multiplied by a first weight stored in one of the two SRAMs of said each TCAM bit cell is received from a first search line of a search line pair coupled to said each TCAM bit cell, a second input activation that is multiplied by a second weight stored in another of the two SRAMs of said each TCAM bit cell is received from a second search line of the search line pair coupled to said each TCAM bit cell, and a period in which a first in-memory multiplication between the first input activation and the first weight is performed does not overlap a period in which a second in-memory multiplication between the second input activation and the second weight is performed;
and
a plurality of activation lines, each activation line interconnecting a row of the plurality of rows of TCAM bit cells and configured to receive output values from both SRAMs in each TCAM bit cell within the row.

A method for operating a ternary content-addressable memory TCAM) cell array to perform a multiply-and-accumulate (MAC) operation of a vector of input activations with a weight matrix of weights, wherein TCAM cells in the TCAM cell array are arranged in a plurality of rows and each comprising two static random-access memories (SRAMs), the method comprising:

storing the weights in the weight matrix of a neural network in the SRAMs within the TCAM cell array, wherein the TCAM cell array is used as a compute-in-memory device;
multiplying, with the SRAMs, the vector of input activations to the weights to produce a plurality of output values, wherein each SRAM operates as a multiplier that performs an in-memory multiplication between a weight stored therein and an input activation of the vector of input activations, a first input activation that is multiplied by a first weight stored in one of the two SRAMs of a same TCAM cell is received from a first search line of a search line pair coupled to the same TCAM cell, a second input activation that is multiplied by a second weight stored in another of the two SRAMs of the same TCAM cell is received from a second search line of the search line pair coupled to the same TCAM cell, and a period in which a first in-memory multiplication between the first input activation and the first weight is performed does not overlap a period in which a second in-memory multiplication between the second input activation and the second weight is performed;
collecting output values from both SRAMs in each TCAM cell of a row of the plurality of rows with an activation line;
and
summing the output values from the activation line to generate a MAC value.

2 more elements

Dependencies are inferred from the claim text.

Facts

Application: 17/709,183Filed 30 Mar 2022 · us provisional application
Priority: 9 Apr 2021
Published: US 2022/0328099 A113 Oct 2022
Granted: 10 Jun 2025
Term ends: Not estimatedThe source records don’t support an estimate.

Classification7 codes

G11C 15/04
Digital stores in which information comprising one or more characteristic parts is written into the store and in which information is read-out…
G06F 7/50
Using non-contact-making devices, e.g. tube, solid state device; using unspecified devices: adding; Subtracting
G06F 7/523
Multiplying; Dividing: multiplying only
G11C 11/412
Forming static cells with positive feedback, i.e. cells not needing refreshing or charge regeneration, e.g. bistable multivibrator or Schmitt…