Abstract: Systems and methods are provided performing for low compute depth map generation by implementing acts of obtaining a stereo pair of images of a scene, downsampling the stereo pair of images, generating a depth map by stereo matching the downsampled stereo pair of images, and generating an upsampled depth map based on the depth map using an edge-preserving filter for obtaining at least some data of at least one image of the stereo pair of images.

Abstract: One example provides a method of operating a time-of-flight camera system comprising an illumination source and an image sensor. The method comprises operating the illumination source and the image sensor to control a plurality of integration cycles and a plurality of readout cycles. In each integration cycle, the method comprises performing a plurality of pulse width modulated (PWM) illumination cycles where each PWM illumination cycle is separated from one or more adjacent PWM illumination cycles by a non-illumination cycle. For each PWM illumination cycle, the method comprises directing photocharge to in-pixel memory for each pixel that is performing image integration and for each non-illumination cycle conducting photocharge away from the in-pixel memory for each pixel that is performing image integration. The readout cycle comprises, for each pixel that performed image integration, reading a charge stored in the in-pixel memory after the integration cycle.

Abstract: Embodiments are associated with a speaker-independent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-independent attention-based encoder-decoder model associated with a first output distribution, and a speaker-dependent attention-based encoder-decoder model to classify output tokens based on input speech frames, the speaker-dependent attention-based encoder-decoder model associated with a second output distribution. The second attention-based encoder-decoder model is trained to classify output tokens based on input speech frames of a target speaker and simultaneously trained to maintain a similarity between the first output distribution and the second output distribution.

Abstract: A method for human movement classification includes a plurality of radio frequency (RF) antennas to expose a body surface of a human user to an E-field, at least a part of the human user positioned within a near-field region relative to the RF antennas. The plurality of RF antennas are scanned to determine ground-relative changes in electrical conditions. A differential-scanning subset of RF antennas are selected, including two or more RF antennas for which the changes in electrical conditions exceed a threshold. For each RF antenna in the differential-scanning subset, a set of antenna-relative changes in electrical conditions are determined by comparing the RF antenna to one or more other RF antennas in the differential-scanning subset as a plurality of pairs. Based at least in part on the ground-relative changes in electrical conditions and the antenna-relative changes in electrical conditions, a movement performed by the human user is classified.

Abstract: Methods for data visibility in nested transactions in distributed systems are performed by systems and devices. Distributed executions of queries are performed in processing systems according to isolation level protocols with unique nested transaction identifiers for data management and versioning across one or more data sets, one or more compute pools, etc., within a logical server via a single transaction manager that oversees the isolation semantics and data versioning. A distributed query processor of the systems and devices performs nested transaction versioning for distributed tasks by generating nested transaction identifiers, encoded in data rows, which are used to enforce correct data visibility. Data visibility is restricted to previously committed data from distributed transactions and tasks, and is blocked for distributed transactions and tasks that run concurrently.

Abstract: Providing extended branch target buffer (BTB) entries for storing trunk branch metadata and leaf branch metadata is disclosed herein. In one aspect, a processor comprises a BTB circuit comprising a BTB comprising a plurality of extended BTB entries. The BTB circuit is configured to store trunk branch metadata for a first branch instruction in an extended BTB entry of the plurality of extended BTB entries, wherein the extended BTB entry corresponds to a first aligned memory block containing an address of the first branch instruction. The BTB circuit is also configured to store leaf branch metadata for a second branch instruction in the extended BTB entry in association with the trunk branch metadata, wherein an address of the second branch instruction is subsequent to a target address of the first branch instruction within a second aligned memory block.

Abstract: The resolving of a decentralized identifier to a corresponding data structure using multiple resolvers. This allows for the use of a consensus of resolvers to improve trust in the resolution process. In order to resolve, a decentralized identifier is sent to multiple resolvers. In response, each of at least some of those resolvers will return a data structure of a particular type (e.g., a decentralized identifier document) that is associated with the decentralized identifier. Then, it is determined whether the data structure for at least some number of resolvers matches each other. That is, it is determined whether at least some predetermined threshold of resolvers is returning the same data structure (e.g., the same decentralized identifier document). If so, then it is determined that the matching data structure is indeed associated with the decentralized identifier. Otherwise, the resolution process has failed.

Abstract: An approach for clustering large sets of categorical data involves iteratively ordering the data points, partitioning the data into blocks based on the ordering, and clustering the data points within each block, where different iterations use different orderings and, thus, different partitionings. In some embodiments, the data points are represented by multi-dimensional categorical vectors, and the orderings are based on permutations of the categorical dimensions. The iterative clustering may be repeated for multiple successive time windows to track the clusters. Various applications of the disclosed clustering approach, including for cyber security, are also described.

Abstract: In a see-through waveguide-based HMD device configured to display holographic virtual images within a field of view (FOV) of a device user, a glint-based eye tracker illumination system provides infrared (IR) point sources at known locations having a predetermined angular distribution using optical components—including input and output couplers and diffusers—on a waveguide that is located in front of the user's eyes. An input coupler couples light from an IR source into the illumination system waveguide which is propagated to one or more output couplers. Separate diffuser elements aligned with the output couplers distribute the out-coupled IR light into a deterministic range of divergent angles to function as point sources for eye tracker glints. Various illustrative illumination system waveguide architectures are disclosed in which the optical components can be disposed on the same or opposite sides of the waveguide in dual-layered and dual-sided arrangements.

Abstract: An operating system service for persistently executing programs comprises a registration service configured to register a manifest defining a persistence mode of a user-controllable program. The operating system service further comprises a runtime service configured, based on the persistence mode defined via the manifest, to detect and reject termination requests for the program, and to detect that the program has terminated and restart execution of the program.

Abstract: In non-limiting examples of the present disclosure, systems, methods, and devices for a graphic search bar with responsive results is provided. A search bar may be generated. The search bar may include a text input area and a number of dropdown menus, each of which contain a selectable list of options. The search bar may be rendered in an application user interface that provides content options that are searchable using the search bar. A selection of an option from a first of the dropdown menus may be received. In response to receiving the selection, the content options may be filtered in real time based on the selected option to generate search results. The search results may be rendered for immediate display to the user.

Abstract: Examples of augmented reality (AR) environment control advantageously employ multi-factor intention determination and include: performing a multi-factor intention determination for summoning a control object (e.g., a menu, a keyboard, or an input panel) using a set of indications in an AR environment, the set of indications comprising a plurality of indications (e.g., two or more of a palm-facing gesture, an eye gaze, a head gaze, and a finger position simultaneously); and based on at least the set of indications indicating a summoning request by a user, displaying the control object in a position proximate to the user in the AR environment (e.g., docked to a hand of the user). Some examples continue displaying the control object while at least one indication remains, and continue displaying the control object during a timer period if one of the indications is lost.

Abstract: An example method of parsing data in a distributed computing system includes assigning a worker k to scan a data chunk k at a data storage node at which data chunk k is stored, each of one or more data chunks including a partial record, receiving, from each worker k, a number of instances of a context-varying symbol, a position in the data chunk k of a first instance of a context-dependent symbol after an even number of instances of the context-varying symbol, and a position in the data chunk k of a first instance of the context-dependent symbol after an odd number of instances of the context-varying symbol, for each data chunk that includes a partial record, computing an adjusted data chunk including data from a later data chunk to complete the partial record, parsing the adjusted data chunks and executing a query against the plurality of records.

Abstract: Examples of the present disclosure describe systems and methods for generating multiple semantic hypotheses for search query intent understanding. In aspects, a search query may be received by a query analysis component associated with a search system. The query analysis component may be used to evaluate the search query for ambiguity in the domain, intent, and/or slot(s) of the search query. A set of hypotheses representing for one or more combinations of the domain, intent, and/or slot(s) of the search query may be generated. The set of hypotheses may be scored and/or ranked. Based on the scores/ranks, one or more of the hypotheses in the set of hypotheses may be provided to a user and/or one or more processing components accessible to the search system.

Abstract: Solutions for enforcing both syntactic and semantic correctness of domain-specific data queries include: receiving a data query; constraining an operation of the data query to enforce semantic correctness, wherein enforcing semantic correctness comprises: determining semantic information associated with each of a plurality of data entities; based on at least the semantic information, determining whether combining the two data entities of the plurality of data entities is allowed or is not allowed; based on at least determining that combining the two data entities is allowed, permitting combining the two data entities in the data query; and based on at least determining that combining the two data entities is not allowed, preventing combining the two data entities in the data query; and executing the data query. Some examples further include basing the permission on (at least) an ontology between the semantic information of the two data entities and a set of axioms.

Abstract: Techniques for auto-starting a VPN in a MAM environment are disclosed. A MAM-controlled application is launched on a computer system. Policy is queried and a determination is made as to whether to auto-start a VPN application based on the policy. Based on the policy, the VPN application is auto-started, and the VPN application initiates a VPN tunnel that is usable by at least the MAM-controlled application. Network communications transmitted to or from the MAM-controlled application then pass through the VPN tunnel.

Abstract: Techniques are disclosed for performing operations in a distributed computing environment comprising a plurality of computing devices implementing a Kubernetes environment. An indication is received of one or more custom resources of the Kubernetes environment. The custom resources are grouped into a commit set. The commit set is defined as an atomic unit which either succeeds or fails atomically. It is determined that at least one of the custom resources has failed to complete. In response to the determination, the custom resources of the commit set are reverted to a previous consistent state.

Abstract: Embodiments of the disclosed technologies include receiving, from a first user device, a first user interface event that includes an upload of a media asset but does not indicate an intended usage for the media asset, determining an intended usage based on a second user interface event, mapping the intended usage to a usage-based specification, converting the media asset to a usage-specific version of the media asset by applying the usage-based specification to the media asset, and initiating a distribution of the usage-specific version of the media asset instead of the media asset to at least one second user device different from the first user device.

Abstract: A computer-implemented method includes generating, using a transaction management computing subsystem of an advertising platform, a set of primary bid requests responsive to receipt of an advertising call. Each primary bid request includes information sufficient to characterize an impression consumer and information sufficient to characterize each of one or more impressions identified in the advertising call. The method also includes sending the set of primary bid requests from the transaction management computing subsystem to a first set of decisioning computing subsystems of the advertising platform. Each decisioning computing subsystem of the first set being operable to generate a bid response based on the information included in a primary bid request.

Abstract: A dormant account identifier is disclosed. An inactive account can be determined based on whether a user activity of the account is outside a threshold amount. A determination can be made as to whether the inactive account is a dormant account based on account activity of a peer account to the inactive account.

Abstract: A deep neural network (“DNN”) module compresses and decompresses neuron-generated activation data to reduce the utilization of memory bus bandwidth. The compression unit receives an uncompressed chunk of data generated by a neuron in the DNN module. The compression unit generates a mask portion and a data portion of a compressed output chunk. The mask portion encodes the presence and location of the zero and non-zero bytes in the uncompressed chunk of data. The data portion stores truncated non-zero bytes from the uncompressed chunk of data. A decompression unit receives a compressed chunk of data from memory in the DNN processor or memory of an application host. The decompression unit decompresses the compressed chunk of data using the mask portion and the data portion.

Abstract: A public-private key cryptographic scheme is described for granting authenticating a client to a remote device or service in order to access a secure resource. The client is provided the public key, but the private key is stored in a hardware security module (HSM) that the client is not able to access. The client requests a digital signature be generated from the private key from a secure vault service. The secure vault service accesses the HSM and generates the digital certificate, which is then passed to the client. The digital certificate may be added to a security token request submitted to an identity provider. The identity provider determines whether the digital signature came from the private key.

Abstract: Techniques are disclosed for implementing direct memory access in a virtualized computing environment. A memory access policy of the virtualized computing environment is applied to a direct memory access connection request received from a first virtual machine via an exception path. The request is flagged to indicate that the request has been processed and the request is forwarded to a network interface device configured to execute offloaded network functions for one or more virtual machines. A memory access policy of the virtualized computing environment is applied to a direct memory access connection reply received from a second virtual machine on the exception path. The reply is flagged to indicate that the reply has been processed and the reply is forwarded to the network interface device. A direct memory access connection is established between first and second virtual machines in accordance with the request.

Abstract: Systems and methods disclosed herein are related to an intelligent UI element selection system using eye-gaze technology. In some example aspects, a UI element selection zone may be determined. The selection zone may be defined as an area surrounding a boundary of the UI element. Gaze input may be received and the gaze input may be compared with the selection zone to determine an intent of the user. The gaze input may comprise one or more gaze locations. Each gaze location may be assigned a value according to its proximity to the UI element and/or its relation to the UI element's selection zone. Each UI element may be assigned a threshold. If the aggregated value of gaze input is equal to or greater than the threshold for the UI element, then the UI element may be selected.

Abstract: An electronic mail (email) system generates an email user interface display, with a timeslot actuator, for composing an email message. Actuation of the timeslot actuator is detected and the email system accesses available timeslots from a calendar system corresponding to the user. User selectable display elements, corresponding to each available timeslot, are displayed on the email user interface display, along with message content. When a user selectable display element corresponding to an available timeslot is selected, the email message content is populated with a selectable timeslot element, corresponding to the available timeslot, that can be sent to recipients in the email message.

Abstract: A method of enhancing a link in collaboration applications includes determining, from content entered into a collaboration application, a string of characters indicating a uniform resource locator (URL). Upon determining the string of characters indicating the URL from the content entered into the collaboration application, the method includes automatically calling a first web service to obtain a representation of content referenced by the URL by sending a request, at least including the URL, to the first web service. The representation of the content referenced by the URL can then be received from the first web service in a format customized for the collaboration application, a computing device on which the collaboration application is implemented, or both.

Abstract: Examples are disclosed that relate to pixel-shifting devices for increasing display resolution. An example pixel-shifting device comprises an outer frame, an inner frame coupled to the outer frame via a flexure, a path-shifting optical element mounted to the inner frame, and one or more piezoelectric actuators configured to drive motion of the inner frame.

Abstract: A liquid-submersible thermal management system includes a cylindrical outer shell and an inner shell positioned in an interior volume of the outer shell. The cylindrical outer shell has a longitudinal axis oriented vertically relative to a direction of gravity, and the inner shell defines an immersion chamber. The liquid-submersible thermal management system a spine positioned inside the immersion chamber and oriented at least partially in a direction of the longitudinal axis with a heat-generating component located in the immersion chamber. A working fluid is positioned in the immersion chamber and at least partially surrounding the heat-generating component. The working fluid receives heat from the heat-generating component.

Abstract: An electronic device includes a substrate having a first surface and an opposite second surface; a photonic transmitter supported by the first surface of the substrate; a photonic receiver supported by the first surface of the substrate; a microfluidic volume positioned in the second surface of the substrate; a waveguide positioned to direct photonic signal from the photonic transmitter to the photonic receiver, wherein at least a portion of the waveguide is positioned between the first surface of the substrate and at least a portion of the microfluidic volume; and a working fluid in the microfluidic volume to receive heat from the waveguide.

Abstract: The present disclosure relates to processing operations configured to provide processing that automatically analyzes acoustic signals from attendees of a live presentation and automatically triggers corresponding reaction indications from results of analysis thereof. Exemplary reaction indications provide feedback for live presentations that can be presented in real-time (or near real-time) without requiring a user to manually take action to provide any feedback. As a non-limiting example, reaction indications may be presented in a form that is easy to visualize and understand such as emojis or icons. Another example of a reaction indication is a graphical user interface (GUI) notification that provides a predictive indication of user intent derived from analysis of acoustic signals.

Abstract: A congested network link is identified. Applications that are contributing to the congestion are also identified. The contribution of each of the identified applications to the congestion is then determined in order to determine whether the application's usage of the network should be throttled. A control signal is generated in order to indicate to the application that it is to perform an application-specific throttling operation to reduce the network bandwidth consumed by the application. If the application-specific throttling does not sufficiently reduce network congestion, a more aggressive throttling operation is performed.

Abstract: Methods, systems, and computer program products are provided for creating a resource management testing environment. An initial population of databases is established in a database ring, having an in initial count of databases and different types of databases that are determined based on an initial database population model. The initial population model receives ring classification information for the database ring from a ring grouping model. A sequence of database population-change events is generated based on a model, to change the population of the databases over time in the ring. An orchestration framework performs testing of resource manager operations based on the model-defined initial population of databases and the model-defined populations of databases changed over time. Model-defined resource usage metrics for each database are utilized to test the resource manager operations. Resource usage metrics and database add/drop events of a production system are used to train the models.

Abstract: An accessory device receives authentication information from a host computing device connected thereto and determines whether the authentication information is valid. If the authentication information is valid, the accessory device applies a first access policy that specifies whether the accessory device can provide the host computing device with access to none, some, or all of various computing resources of the accessory device. If the authentication information is not valid, the accessory device applies a second access policy that is different than the first access policy. The accessory device can also be provisioned with access policies by a host computing device if the host computing device successfully authenticates with the accessory device. In either case, authenticating the host computing device may include verifying a digital signature of a certificate provided by the host computing device using a public key of a certificate authority that has been provisioned to the accessory device.

Abstract: A computing device is provided, including a logic subsystem with one or more processors, and memory storing instructions executable by the logic subsystem. These instructions are executed to obtain one or more source images, segment the one or more source images to generate a plurality of segments, determine a priority order for the plurality of segments, and transmit the plurality of segments to a remote computing device in the priority order. The plurality of segments are spatial components generated by spatial decomposition of the one or more source images and/or frequency components that are generated by frequency decomposition of the one or more source images. A remote computing device may receive these components in priority order, and perform certain algorithms on individual components without waiting for the entire image to upload.

Abstract: Examples are disclosed that relate to backlight modules, displays including a backlight module, and methods for assembling a backlight module. In one example, a backlight module for a display comprises light sources affixed to a printed circuit and a light guide plate. An adhesive affixes the printed circuit to the light guide plate. The light guide plate comprises a body and a plurality of projections extending from the body, with the plurality of projections retaining two or more of the light sources between the projections and the body of the light guide plate.

Abstract: A cloud-connected baseboard management controller (BMC) enables remote management of all management layers of edge-based host machines from the cloud without imposition of costs associated with firewalls and private data connections. A remote management agent incorporated into the BMC enables creation of a remote management-enabled bare-metal server paradigm in which remote management from the cloud is supported at the lowest hardware levels which provides for cost-effective utilization of network resources down to even a single isolated node.

Abstract: To increase the speed with which a Second Layer Address Table (SLAT) is traversed, memory having the same access permissions is contiguously arranged such that one or more hierarchical levels of the SLAT need not be referenced, thereby resulting in more efficient SLAT traversal. “Slabs” of memory are established whose memory range is sufficiently large that reference to a hierarchically lower level table can be skipped and a hierarchically higher level table's entries can directly identify relevant memory addresses. Such slabs are aligned to avoid smaller intermediate memory ranges. The loading of code or data into memory is performed based on a next available memory location within a slab having equivalent access permissions, or, if such a slab is not available, or if an existing slab does not have a sufficient quantity of available memory remaining, a new slab with the proper access permissions is established.

Abstract: Innovations in the area of prediction of block vector (“BV”) values improve encoding or decoding of blocks using intra block copy (“BC”) prediction. For example, some of the innovations relate to use of a default BV predictor with a non-zero value. Other innovations relate to use of a selected one of multiple BV predictor candidates for a current block. Still other innovations relate to use of a skip mode in which a current intra-BC-predicted block uses a predicted BV value.

Abstract: Systems and methods for providing resilient persistent connections for communication are disclosed. Methods include identifying, at a first device, a second device registered with a client connection service; obtaining, from the client connection service, a first resource identifier for delivering request messages to the second device via the client connection service; identifying, based on the obtained first resource identifier, a first target resource for a first request message directed to the second device, wherein the first target resource specifies a first host included in the client connection service; sending, to the client connection service via the data communication network, the first request message to the first target resource for delivery to the second device by the client connection service; and receiving, from the client connection service via the data communication network, a first response message provided by the second device as a response to the first request message.

Abstract: Described are examples for deploying workloads in a cloud-computing environment. In an aspect, based on a desired number of workloads of a process to be executed in a cloud-computing environment and based on one or more failure probabilities, an actual number of workloads of the process to execute in the cloud-computing environment to provide a level of service can be determined and deployed. In another aspect, a standby workload can be executed as a second instance of the process without at least a portion of the separate configuration used by the multiple workloads, and based on detecting termination of one of multiple workloads, the standby workload can be configured to execute based on the separate configuration of the separate instance of the process corresponding to the one of the multiple workloads.

Abstract: Trace recording based on data influxes to an outer-level cache and cache coherence protocol (CCP) transitions between inner caches. Example computing device(s) include a plurality of processing units, a plurality of (N-1)-level caches, and an N-level cache that is associated with two or more of the (N-1)-level caches and that is a backing store for the two or more (N-1)-level caches. Based at least on detecting influx(es) of data to a location in the N-level cache during execution across the processing units, the computing device(s) causes the influx(es) of data to be logged. The computing device(s) also causes one or more (N-1)-level CCP transitions between the two or more (N-1)-level caches to be logged. The (N-1)-level CCP transitions result from the location being accessed by two or more of the processing units.

Abstract: Interaction between a user and a conversational digital assistant executing on a computing device is controlled. Multiple interaction pairs are stored in one or more datastores accessible by the conversational digital assistant. Each interaction pair includes an interaction query and an associated assistance operation. An interactive engagement event is detected between the user and the conversational digital assistant, responsive to the storing operation. An interaction pair is selected from the one or more datastores, responsive to the operation of detecting an interactive engagement event. The interaction query of the selected interaction pair is communicated to the user. The assistance operation associated with the communicated interaction query is executed, responsive to receipt of a response from the user to the interaction query.

Abstract: Disclosed in some examples are methods, systems, machine-readable media, and devices which provide for real-time personalized suggestions for participants in a network-based communication service. The personalized suggestions may include options for taking actions, content suggestions, and smart replies. These suggestions may be based upon the current conversation and are delivered personally to each participant.

Abstract: Examples are disclosed relating to friction hinges. In one example, a friction hinge comprises a first stub and adjacent second stub extending from a stub substrate. A first finger extends from a finger substrate between the first stub and the second stub and comprises a distal contacting face that is biased into contact with a first biasing face of the first stub. A second finger adjacent to the first finger extends from the finger substrate between the first stub and the second stub and comprises a distal contacting face that is biased into contact with a second biasing face of the second stub. A pin extends through apertures in the first stub, first finger, second finger, and second stub.

Abstract: A computing device is provided, including one or more processing devices, one or more temperature sensors, a fan, and a fan tachometer. The one or more processing devices may be configured to execute an application program. While executing the application program, the one or more processing devices may be further configured to collect performance data including temperature data received from the one or more temperature sensors and fan speed data received from the fan tachometer. The one or more processing devices may be further configured to generate a fan control signal at least in part by applying a machine learning model to the performance data. The one or more processing devices may be further configured to control the fan according to the fan control signal.

Abstract: Examples are disclosed related to forming solder joints between printed circuits by using radiant heat. One example provides a method of manufacturing an electronic device, the method comprising aligning a contact of a first printed circuit with a via of a second printed circuit. The method further comprises applying radiant heat via an infrared light source to a second surface of the second printed circuit, the radiant heat incident on the via to cause the via to conduct heat to solder located at an interface of the contact and the via, and after heating the solder to reflow, cooling the solder, thereby forming a solder joint between the contact of the first printed circuit and the via of the second printed circuit.

Abstract: The automatic detection of inconsistencies in a database system is described. A first signature and a second signature are received. The first signature is a signature of a result of a first execution of the query against a database by a first version of database engine program code. The second signature is a signature of a result of a second execution of the query by a second version of the database engine program code. A determination is made of whether the first signature and the second signature match. In response to the first signature and the second signature failing to match, an inconsistency report regarding at least one of the first or second versions of the database engine program code is generated and remediation regarding at least one of the first or second versions of the database engine program code is performed.

Abstract: Embeddable extensions of the functionality of email and calendaring applications are described herein. An originating email and calendaring application formats meeting items to include extensible content, e.g., within the body field of the meeting item. Recipient email and calendaring applications receive the meeting items and, depending on whether they are configured to provide an extended functionality, will either deploy the extensible content item according to the extended functionality or present the extensible content item according to a standardized functionality. The body field includes the extensible content item according to the standardized functionality that is interpretable by configured clients to provide the extended functionality.

Abstract: A computing apparatus comprising one or more computer readable storage media, one or more processors operatively coupled with the one or more computer readable storage media, and an application comprising program instructions stored on the one or more computer readable storage media that direct the computing apparatus to at least display electronic document in a user interface to an application and send a request to an energy analysis service to analyze an energy footprint of the electronic document, wherein the request identifies a characteristic of the electronic document. The program instructions further direct the computing apparatus to receive a reply to the request, wherein the reply includes a recommendation to decrease the energy footprint of the electronic document and modify the electronic document in accordance with the recommendation.

Abstract: Methods, systems and computer program products are provided for improving performance (e.g., reducing power consumption) of a hardware accelerator (e.g., neural processor) comprising hybrid or analog multiply and accumulate (MAC) processing elements (PEs). Selective variation of the precision of an array of MAC PEs may reduce power consumption of a neural processor. Power may be conserved by dynamically controlling the precision of analog to digital (ADC) output bits for one or more MAC PEs. Dynamic control of ADC output bit precision may be based on precision information determined during training and/or post-training (e.g., quantization) of an artificial intelligence (AI) neural network (NN) model implemented by the neural processor. Precision information may include a range of dynamic precision for each of a plurality of nodes of a computation graph for the AI NN model.