International Business Machines Corporation

Advanced Micro Devices Corporation

Globalfoundries Inc.

Zhibin Ren

Xinhui Wang

Jeffrey W Sleight

Kevin K Chan

Dinkar Virendra Singh

Haizhou Yin

Amlan Majumdar

Bruce Bennett Doris

Ghavam G Shahidi

Effendi Leobandung

Toshiharu Furukawa

Wilfried Haensch

Renee Tong Mo

Anthony I Chou

Huilong Zhu

Paul Michael Solomon

Meikei Ieong

Dae-Gyu Park

Jing Wang

Tak H Ning

Haining S Yang

Judson Robert Holt

Katherine Lynn Saenger

Jin Cai

Min Yang

Thomas Safron Kanarsky

Jinghong Li

Richard Andre Wachnik

Malte Johannes Rasch

John Michael Hergenrother

Yue Tan

Christine Qiqing Ouyang

MeiKei Ieong

Alexander Reznicek

Devendra K Sadana

Stephen W Bedell

Shom Ponoth

Keith Kwong Hon Wong

Ying Di Zhang

Chengwen Pei

Dominic Joseph Schepis

Chung-Hsun Lin

Brian Joseph Greene

Robert H Dennard

Seyoung Kim

Xiao Lin Sun

Henry K Utomo

Leathen Shi

Xinlin Wang

Ashima B Chakravarti

Diane Catherine Boyd

Christopher D Sheraw

Yulong Li

Junedong Lee

Yue Ke

Woo-Hyeong Lee

Gen Pei

Joel P Desouza

Suk Hoon Ku

Ryan Matthew Mitchell

Yinxiao Yang

Jeremy John Kempisty

Eric T Harley

Methods and systems of implementing a convolutional neural network are described. In an example, a structure may receive input signals and distribute the input signals to a plurality of unit cells. The structure may include a plurality of multi-kernel modules that may include a respective set of unit cells. A unit cell may correspond to an element of a kernel being implemented in the convolutional neural network and may include a storage component configured to store a weight of a corresponding element of the kernel. A first pass gate of the unit cell may be activated to pass a stored weight of the unit cell to a plurality of operation circuits in the corresponding unit cell, such that the stored weight may be applied to the input signals. The structure may generate a set of outputs based on the application of the stored weights to the input signals.

Multi-kernel configuration for convolutional neural networks

A resistive processing unit cell includes a weight storage device to store a weight value of the resistive processing unit cell, and multiple circuit blocks. Each circuit block includes a weight update circuit coupled to dedicated update control lines, and a weight read circuit coupled to dedicated read control lines. The circuit blocks are configured to operate in parallel to (i) perform separate weight read operations in which each read circuit generates a read current based on a stored weight value, and outputs the read current on the dedicated read control lines of the read circuit, and (ii) perform separate weight update operations in which each update circuit receives respective update control signals on the dedicated update control lines, generates update currents based on the respective update control signals, and applies the update current to the weight storage device to adjust the weight value based on the update current.

Resistive processing unit cell having multiple weight update and read circuits for parallel processing of data using shared weight value

Method, systems, crosspoint arrays, and systems for tuning a neural network. A crosspoint array includes: a set of conductive rows, a set of conductive columns intersecting the set of conductive rows to form a plurality of crosspoints, a circuit element coupled to each of the plurality of crosspoints configured to store a weight of the neural network, a voltage source associated with each conductive row, a first integrator attached at the end of at least one of the conductive column, and a first variable resistor attached to the integrator and the end of the at least one conductive column.

Aggregate adjustments in a cross bar neural network

Fin-type transistor fabrication methods and structures are provided which include, for example, providing a gate structure extending at least partially over a fin extended above a substrate structure, the gate structure being disposed adjacent to at least one region of the fin; disposing a protective film conformally over the gate structure and over the at least one region; modifying the protective film over the at least one region of the fin to form a conformal buffer layer, wherein the modifying selectively alters a crystalline structure of the protective film over the at least one region which thereby becomes the conformal buffer layer, without altering the crystalline structure of the protective film disposed over the gate structure; and removing the un-altered protective film over the gate structure, leaving the conformal buffer layer over the at least one region to form a source region and a drain region of the fin-type transistor.

Conformal buffer layer in source and drain regions of fin-type transistors

FinFETs are merged together by a metal. The method of manufacturing the FinFETs include forming a plurality of fin bodies on a substrate and merging the fin bodies with a metal. The method further includes implanting source and drain regions through the metal.

Merged FinFETs and method of manufacturing the same

Contact with a floating body of an FET in SOI may be formed in a portion of one of the two diffusions of the FET, wherein the portion of the diffusion (such as N&#x2212;, for an NFET) which is 'sacrificed' for making the contact is a portion of the diffusion which is not immediately adjacent (or under) the gate. This works well with linked body FETs, wherein the diffusion does not extend all the way to BOX, hence the linked body (such as P&#x2212;) extends under the diffusion where the contact is being made. An example showing making contact for ground to two NFETs (PG and PD) of a 6T SRAM cell is shown.

Body contacts for FET in SOI SRAM array

A method for forming a field effect device includes forming a gate portion on a silicon-on-insulator layer (SOI), forming first spacer members on the SOI layer adjacent to the gate portion, depositing a layer of spacer material on the SOI layer, the first spacer members, and the gate portion, removing portions of the layer of spacer material to form second spacer members on the SOI layer adjacent to the first spacer members, forming a source region and a drain region on the SOI layer by implanting ions in the SOI layer, and etching to remove the second spacer members.

Fabrication of field effect devices using spacers

Aspects of the invention provide for preventing undercuts during wafer etch processing and enhancing back-gate to channel electrical coupling. In one embodiment, aspects of the invention include a semiconductor structure, including: a high-k buried oxide (BOX) layer atop a bulk silicon wafer, the high-k BOX layer including: at least one silicon nitride layer; and a high-k dielectric layer; and a silicon-on-insulator (SOI) layer positioned atop the high-k BOX layer.

High-k dielectric and silicon nitride box region

SOI structures with silicon layers less than 20 nm thick are used to form ETSOI semiconductor devices. ETSOI devices are manufactured using a thin tungsten backgate encapsulated by thin nitride layers to prevent metal oxidation, the tungsten backgate being characterized by its low resistivity. The structure includes at least one FET having a gate stack formed by a high-K metal gate and a tungsten region superimposed thereon, the footprint of the gate stack utilizing the thin SOI layer as a channel. The SOI structure thus formed controls the Vt variation from the thin SOI thickness and dopants therein. The ETSOI high-K metal backgate fully depleted device in conjunction with the thin BOX provides an excellent short channel control and lowers the drain induced bias and sub-threshold swings. The structure supports the evidence of the stability of the wafer having a tungsten film during thermal processing, during STI and contact formation.

Thin box metal backgate extremely thin SOI device

A semiconductor-on-insulator structure includes a buried dielectric layer interposed between a base semiconductor substrate and a surface semiconductor layer. The buried dielectric layer comprises an oxide material that includes a nitrogen gradient that peaks at the interface of the buried dielectric layer with at least one of the base semiconductor substrate and surface semiconductor layer. The interface of the buried dielectric layer with the at least one of the base semiconductor substrate and surface semiconductor layer is abrupt, providing a transition in less than about 5 atomic layer thickness, and having less than about 10 angstroms RMS interfacial roughness. A second dielectric layer comprising an oxide dielectric material absent nitrogen may be located interposed between the buried dielectric layer and the surface semiconductor layer.

Semiconductor-on-insulator (SOI) structures including gradient nitrided buried oxide (BOX)

Silicon-on-insulator (SOI) structures with silicon layers less than 20 nm thick are used to form extremely thin silicon-on-insulator (ETSOI) semiconductor devices. ETSOI devices are manufactured using a thin tungsten backgate encapsulated by thin nitride layers to prevent metal oxidation, the tungsten backgate being characterized by its low resistivity. The structure further includes at least one FET having a gate stack formed by a high-K metal gate and a tungsten region superimposed thereon, the footprint of the gate stack utilizing the thin SOI layer as a channel. The SOI structure thus formed controls the Vt variation from the thin SOI thickness and dopants therein. The ETSOI high-K metal backgate fully depleted device in conjunction with the thin BOX provides an excellent short channel control and significantly lowers the drain induced bias and sub-threshold swings. The present structure supports the evidence of the stability of the wafer having a tungsten film during thermal processing, and especially during STI and contact formation.

Thin-BOX metal backgate extremely thin SOI device

A method of fabricating and a structure of a merged multi-fin finFET. The method includes forming single-crystal silicon fins from the silicon layer of an SOI substrate having a very thin buried oxide layer and merging the end regions of the fins by growing vertical epitaxial silicon from the substrate and horizontal epitaxial silicon from ends of the fins such that vertical epitaxial silicon growth predominates.

finFETs and methods of making same

A method is provided to fabricate a semiconductor device, where the method includes providing a substrate comprised of crystalline silicon; implanting a ground plane in the crystalline silicon so as to be adjacent to a surface of the substrate, the ground plane being implanted to exhibit a desired super-steep retrograde well (SSRW) implant doping profile; annealing implant damage using a substantially diffusionless thermal annealing to maintain the desired super-steep retrograde well implant doping profile in the crystalline silicon and, prior to performing a shallow trench isolation process, depositing a silicon cap layer over the surface of the substrate. The substrate may be a bulk Si substrate or a Si-on-insulator substrate. The method accommodates the use of an oxynitride gate stack structure or a high dielectric constant oxide/metal (high-K/metal) gate stack structure. The various thermal processes used during fabrication are selected/controlled so as to maintain the desired super-steep retrograde well implant doping profile in the crystalline silicon.

Method for fabricating super-steep retrograde well MOSFET on SOI or bulk silicon substrate, and device fabricated in accordance with the method

Growing community of inventors

Hopewell Junction, NY, United States of America

Zhibin Ren