Industrial Technology Research Institute

Tsinghua University

Jian-Wei Su

Chih-Sheng Lin

Fu-Cheng Tsai

Tuo-Hung Hou

Meng-Fan Chang

Sih-Han Li

Yu-Hui Lin

Ping-Chun Wu

Jin-Sheng Ren

Yen-Chi Chou

Shih-Chieh Chang

Shyh-Shyuan Sheu

Heng-Yuan Lee

Kuo-Hua Tseng

Chih-Ming Lai

Li-Yang Hong

Chuan-Jia Jhang

Je-Min Hung

Ru-Hui Liu

Disclosed is a memory cell including a first transistor having a first terminal coupled to a bit line; a second transistor having a first terminal coupled to a bit line bar; a weight storage circuit coupled between a gate terminal of the first transistor and a gate terminal of the second transistor, storing a weight value, and determining to turn on the first transistor or the second transistor according to the weight value; and a driving circuit coupled to a second terminal of the first transistor, a second terminal of the second transistor, and at least one word line, receiving at least one threshold voltage and at least one input data from the word line, and determining whether to generate an operation current on a path of the turned-on first transistor or the turned-on second transistor according to the threshold voltage and the input data.

Memory cell

An apparatus and a method for neural network computation are provided. The apparatus for neural network computation includes a first neuron circuit and a second neuron circuit. The first neuron circuit is configured to execute a neural network computation of at least one computing layer with a fixed feature pattern in a neural network algorithm. The second neuron circuit is configured to execute the neural network computation of at least one computing layer with an unfixed feature pattern in the neural network algorithm. The performance of the first neuron circuit is greater than that of the second neuron circuit.

Apparatus and method for neural network computation

According to an exemplary embodiments, the disclosure is directed to a memory circuit which includes not limited to a first half sense amplifier circuit connected to a first plurality of memory cells through a first bit line and configured to receive a unit of analog electrical signal from each of the first plurality of memory cells and to generate a first half sense amplifier output signal corresponding to the first bit line based on a first gain of the half sense amplifier and an accumulation of the units of analog signals, a locking code register circuit configured to receive a locking data and to generate a digital locking sequence, and a source selector circuit configured to receive the digital locking sequence and to generate a first adjustment signal to adjust the first half sense amplifier output signal corresponding to the first bit line by adjusting the first gain.

Memory circuit with sense amplifier calibration mechanism

A memory unit includes at least one memory cell and a computational cell. The at least one memory cell stores a weight. The at least one memory cell is controlled by a first word line and includes a local bit line transmitting the weight. The computational cell is connected to the at least one memory cell and receiving the weight via the local bit line. Each of an input bit line and an input bit line bar transmits a multi-bit input value. The computational cell is controlled by a second word line and an enable signal to generate a multi-bit output value on each of an output bit line and an output bit line bar according to the multi-bit input value multiplied by the weight. The computational cell is controlled by a first switching signal and a second switching signal for charge sharing.

Memory device and memory array structure using charge sharing for multi-bit convolutional neural network based computing-in-memory applications, and computing method thereof

A memory unit with time domain edge delay accumulation for computing-in-memory applications is controlled by a first word line and a second word line. The memory unit includes at least one memory cell, at least one edge-delay cell multiplexor and at least one edge-delay cell. The at least one edge-delay cell includes a weight reader and a driver. The weight reader is configured to receive a weight and a multi-bit analog input voltage and generate a multi-bit voltage according to the weight and the multi-bit analog input voltage. The driver is connected to the weight reader and configured to receive an edge-input signal. The driver is configured to generate an edge-output signal having a delay time according to the edge-input signal and the multi-bit voltage. The delay time of the edge-output signal is positively correlated with the multi-bit analog input voltage multiplied by the weight.

Memory unit with time domain edge delay accumulation for computing-in-memory applications and computing method thereof

A computing in memory (CIM) cell includes a memory cell circuit, a first semiconductor element, a second semiconductor element, a third semiconductor element, and a fourth semiconductor element. A first terminal of the first semiconductor element receives a bias voltage. A control terminal of the first semiconductor element is coupled to a computing word-line. A control terminal of the second semiconductor element is coupled to a first data node in the memory cell circuit. A second terminal of the third semiconductor element is adapted to receive a reference voltage. A control terminal of the third semiconductor element receives an inverted signal of the computing word-line. A first terminal of the fourth semiconductor element is coupled to a first computing bit-line. A second terminal of the fourth semiconductor element is coupled to a second computing bit-line.

Computing in memory cell

A computing in memory (CIM) cell includes a memory cell circuit, a first semiconductor element, a second semiconductor element, and a third semiconductor element. A first terminal of the first semiconductor element is coupled to a first computing bit-line. A control terminal of the first semiconductor element is coupled to a computing word-line. A control terminal of the second semiconductor element is coupled to the memory cell circuit. A first terminal of the second semiconductor element is coupled to a second terminal of the first semiconductor element. A first terminal of the third semiconductor element is coupled to a second terminal of the second semiconductor element. A second terminal of the third semiconductor element is coupled to a second computing bit-line. A control terminal of the third semiconductor element receives a bias voltage.

A memory device for in-memory computation includes data channels, a memory cell array, a maximum accumulated weight generating array, a minimum accumulated weight generating array, a reference generator and a comparator. The data channels are selectively enabled according to data input. The memory cell array generates an accumulated data weight value according to the quantity of enabled data channels, a first resistance and a second resistance. The maximum accumulated weight generating array generates a maximum accumulated weight value according to the quantity of enabled data channels and the first resistance. The minimum accumulated weight generating array generates a minimum accumulated weight value according to the quantity of enabled data channels and the second resistance. The reference generator generates reference value(s) according to the maximum and minimum accumulated weight values. The comparator compares the accumulated data weight value with the reference value(s) to generate a data weight state.

Memory device and data weight state determining method for in-memory computation

A memory structure with input-aware maximum multiply-and-accumulate value zone prediction for computing-in-memory applications includes a memory array, an input-aware zone prediction circuit and an analog-to-digital converter. An input-aware maximum partial multiply-and-accumulate value voltage generator is configured to generate a maximum partial multiply-and-accumulate value according to at least one input value. A prediction-aware global reference voltage generator is configured to generate a plurality of global reference voltages, a maximum reference voltage and a selected minimum reference voltage. A maximum partial multiply-and-accumulate value zone detector is configured to generate a zone switch signal by comparing the maximum partial multiply-and-accumulate value and the global reference voltages. The analog-to-digital converter is configured to convert an analog multiply-and-accumulate output value of the memory array to a digital multiply-and-accumulate output value according to the maximum reference voltage, the selected minimum reference voltage and the zone switch signal.

Memory structure with input-aware maximum multiply-and-accumulate value zone prediction for computing-in-memory applications and operating method thereof

A transpose memory unit for a plurality of multi-bit convolutional neural network based computing-in-memory applications includes a memory cell and a transpose cell. The memory cell stores a weight. The transpose cell is connected to the memory cell and receives the weight from the memory cell. The transpose cell includes an input bit line, at least one first input word line, a first output bit line, at least one second input word line and a second output bit line. One of the at least one first input word line and the at least one second input word line transmits at least one multi-bit input value, and the transpose cell is controlled by the second word line to generate a multiply-accumulate output value on one of the first output bit line and the second output bit line according to the at least one multi-bit input value multiplied by the weight.

Transpose memory unit for multi-bit convolutional neural network based computing-in-memory applications, transpose memory array structure for multi-bit convolutional neural network based computing-in-memory applications and computing method thereof

A computation operator in memory and an operation method thereof are provided. The computation operator in memory includes a word line calculator, a decision-maker and a sense amplifier. The word line calculator calculates a number of enabled word lines of a memory. The decision-maker generates a plurality of reference signals according to at least one of the number of enabled word lines and a used size of the memory, the reference signals are configured to set a distribution range. The sense amplifier receives a readout signal of the memory, and obtains a computation result by converting the readout signal according to the reference signals.

Growing community of inventors

Hsinchu, Taiwan

Jian-Wei Su