Samsung Electronics Co., Ltd.

Lockheed Martin Corporation

Ryan M Hatcher

Titash Rakshit

Jorge A Kittl

Borna Josip Obradovic

Joon Goo Hong

Rwik Sengupta

Mark Stephen Rodder

Dharmendar Reddy Palle

Robert Christopher Bowen

Malik Aqeel Anwar

Brent M Segal

Vladimir Nikitin

Dmytro Apalkov

Jonathan Wesley Ward

Wei-E Wang

Engin Ipek

Chris Bowen

Robert Chris Bowen

A method of pipelining inference of a neural network, which includes an i-th layer (i being an integer greater than zero), an (i+1)-th layer, and an (i+2)-th layer, includes processing a first set of i-th values of the i-th layer to generate (i+1)-th values for the (i+1)-th layer, determining a quantity of the (i+1)-th values as being sufficient for processing, and in response to the determining, processing the (i+1)-th values to generate an output value for the (i+2)-th layer while concurrently processing a second set of i-th values of the i-th layer.

Ultra pipelined accelerator for machine learning inference

A neuromorphic weight cell (NWC) including a resistor ladder including a plurality of resistors connected in series, and a plurality of shunting nonvolatile memory (NVM) elements, each of the shunting NVM elements being coupled in parallel to a corresponding one of the resistors.

Monolithic multi-bit weight cell for neuromorphic computing

A method of storing a sparse weight matrix for a trained artificial neural network in a circuit including a series of clusters. The method includes partitioning the sparse weight matrix into at least one first sub-block and at least one second sub-block. The first sub-block includes only zero-value weights and the second sub-block includes non-zero value weights. The method also includes assigning the non-zero value weights in the at least one second sub-block to at least one cluster of the series of clusters of the circuit. The circuit is configured to perform matrix-vector-multiplication (MVM) between the non-zero value weights of the at least one second sub-block and an input vector during an inference process utilizing the artificial neural network. The sub-blocks containing all zero elements are power gated, thereby reducing overall energy consumption for inference.

Method of enabling sparse neural networks on memresistive accelerators

A method of pipelining inference of a neural network, which includes an i-th layer (i being an integer greater than zero) and an (i+1)-th layer, includes processing, for a first input image, first i-th values of the i-th layer to generate first (i+1)-th values for the (i+1)-th layer, processing, for the first input image, the first (i+1)-th values of the (i+1)-th layer to generate output values, and concurrently with processing, for the first image, the (i+1)-th values, processing, for a second input image, second i-th values of the i-th layer to generate second (i+1)-th values.

Batch size pipelined PIM accelerator for vision inference on multiple images

A non-volatile data retention circuit includes a complementary latch configured to generate and store complementary non-volatile spin states corresponding to an input signal when in a write mode, and to concurrently generate a first charge current signal and a second charge current corresponding to the complementary non-volatile spin states when in read mode, and a differential amplifier coupled to the complementary latch and configured to generate an output signal based on the first and second charge current signals.

Giant spin hall-based compact neuromorphic cell optimized for differential read inference

A neuromorphic multi-bit digital weight cell configured to store a series of potential weights for a neuron in an artificial neural network. The neuromorphic multi-bit digital weight cell includes a parallel cell including a series of passive resistors in parallel and a series of gating transistors. Each gating transistor of the series of gating transistors is in series with one passive resistor of the series of passive resistors. The neuromorphic cell also includes a series of programming input lines connected to the series of gating transistors, an input terminal connected to the parallel cell, and an output terminal connected to the parallel cell.

Multi-bit, SoC-compatible neuromorphic weight cell using ferroelectric FETs

A method and system for training a neural network are described. The method includes providing at least one continuously differentiable model of the neural network. The at least one continuously differentiable model is specific to hardware of the neural network. The method also includes iteratively training the neural network using the at least one continuously differentiable model to provide at least one output for the neural network. Each iteration uses at least one output of a previous iteration and a current continuously differentiable model of the at least one continuously differentiable model.

Method and system for training of neural networks using continuously differentiable models

A method and system are provided. The method includes mapping a binary matrix to an undirected graph form, applying a two-way graph partition algorithm to the mapped binary matrix that minimizes edge cuts between partitions in the mapped binary matrix, applying a greedy algorithm recursively to find a set of row or column permutations that maximizes a transfer of non-zeros from sparse blocks to nonsparse blocks, and sparsifying or densifying the binary matrix according to the applied greedy algorithm.

Optimization of sparsified neural network layers for semi-digital crossbar architectures

A method, system and electronic device for mitigating variance in a two transistor two resistive memory element (2T2R) circuit is provided. The method includes calculating a sum of a number of logical 1's in a column of bitcells in the 2T2R circuit, N, of an input vector, sensing output current values from each current line in the column of bitcells and calculating an inner product, M, of the input vector and the bitcells in the column in the 2T2R circuit based on the sensed output current values.

Variation mitigation scheme for semi-digital mac array with a 2T-2 resistive memory element bitcell

A hardware cell and method for performing a digital XNOR of an input signal and weights are described. The hardware cell includes input lines, a plurality of pairs of magnetic junctions, output transistors and at least one selection transistor coupled with the output transistors. The input lines receive the input signal and its complement. The magnetic junctions store the weight. Each magnetic junction includes a reference layer, a free layer and a nonmagnetic spacer layer between the reference layer and the free layer. The free layer has stable magnetic states and is programmable using spin-transfer torque and/or spin-orbit interaction torque. The first magnetic junction of a pair receives the input signal. The second magnetic junction of the pair receives the input signal complement. The output transistors are coupled with the magnetic junctions such that each pair of magnetic junctions forms a voltage divider. The output transistors form a sense amplifier.

Method and system for providing a variation resistant magnetic junction-based XNOR cell usable in neuromorphic computing

A circuit element. In some embodiments, the circuit element includes a first terminal, a second terminal, and a layered structure. The layered structure may include a first conductive layer connected to the first terminal, a first piezoelectric layer on the first conductive layer, a second piezoelectric layer on the first piezoelectric layer, and a second conductive layer connected to the second terminal. The first piezoelectric layer may have a first piezoelectric tensor and a first permittivity tensor, and the second piezoelectric layer may have a second piezoelectric tensor and a second permittivity tensor, one or both of the second piezoelectric tensor and a second permittivity tensor differing, respectively, from the first piezoelectric tensor and the first permittivity tensor.

Composite piezoelectric capacitor

A weight cell and device are herein disclosed. The weight cell includes a first field effect transistor (FET) and a first resistive memory element connected to a drain of the first FET, a second FET and a second resistive memory element connected to a drain of the second FET, the drain of the first FET being connected to a gate of the second FET and the drain of the second FET is connected to a gate of the first FET, a third FET and a third resistive memory element connected to a drain of the third FET, and a fourth FET and a fourth resistive memory element connected to a drain of the fourth FET, the drain of the third FET is connected to a gate of the fourth FET and the drain of the fourth FET being connected to a gate of the third FET.

4T4R ternary weight cell with high on/off ratio background

A weight cell, an electronic device and a device are provided. The weight cell includes a first resistive memory element and a second resistive memory element, a select transistor, and a layer of Spin Hall (SH) material disposed between the first resistive memory element and the second resistive memory element, the layer of the SH material including a first contact and a second contact. The first contact of the SH material is connected to a drain of the select transistor and the second contact of the SH material is connected to an external word line.

Growing community of inventors

Austin, TX, United States of America

Ryan M Hatcher