Google Patente

Aspects of the disclosure are directed to an error detection approach performed in a processing cycle with improved accuracy and reduced latency. The approach involves determining whether the size of a received input value is greater than, equal to, or less than the size of an accumulator value and performing error detection accordingly using parallel lookup tables corresponding to portions of the input value. To avoid bit reversals in the error detection approach, results stored in the lookup tables may be stored in a bit reversed order. For input values that are less than the size of an accumulator value, the accumulator value is split into two portions to increase accuracy of the error detection result.

Implementations described herein relate to receiving user input directed to an automated assistant, processing the user input to determine whether data from a server and/or third-party application is needed to perform certain fulfillment of an assistant command included in the user input, and generating a prompt that requests a user consent to transmitting of a request to the server and/or the third-party application to obtain the data needed to perform the certain fulfillment. In implementations where the user consents, the data can be obtained and utilized to perform the certain fulfillment. In implementations where the user does not consent, client data can be generated locally at a client device and utilized to perform alternate fulfillment of the assistant command. In various implementations, the request transmitted to the server and/or third-party application can be modified based on ambient noise captured when the user input is received.

A method and system for controlling performance of a workload partitioned among a plurality of accelerator chips (301-304) of a multi-chip system (300). One or more processors (110) may receive performance speed data for each of the accelerator chips (610), obtain a model of the partitioned workload (620), determine a portion of the workload that is either overworked or underworked based on the model of the partitioned workload and the performance speed data for each of the plurality of accelerator chips (630), and adjust a performance speed of an accelerator chip that performs the portion of the partitioned workload that is either overworked or underworked (640).

Generally disclosed herein is an approach to mitigating hardware degradation of server machines caused by frequent chip temperature fluctuations based on controlling the power consumption level, changes in xPU temperature of server machines, and the job start latency for the server machines altogether. According to some examples, a power and temperature optimization system may monitor xPU temperature fluctuations caused by inter-job fluctuations related to the xPU's deep idle state. The xPU's deep idle state may refer to a state where the xPU turns off or reduces the voltage of the xPU components to save power when a job or a unit of work assigned to the xPU stops. The xPU's deep idle state may continue until the next job or unit of work starts.

Methods, systems, and apparatus, including computer-readable storage media and computer program products for determining thresholds for guaranteeing the availability of partitions of different quantities of computing devices and using those thresholds to more accurately provision computing resources to users of a computing platform. A partition (also called a "slice") of computing devices can be provisioned to a user of a computing platform. Slice-size thresholds can indicate how many computing devices on the platform can be provisioned as part of a single slice. Determining slice-size thresholds before provisioning computing resources reduces or eliminates the chance that requests to provision a slice of computing devices are approved when the platform is not able to fulfill those requests with the largest slice-size possible for a given request.

Methods, systems, and apparatus, including computer-readable storage media for hardware-accelerated fine-grained sparse computation. The accelerator provides for improved performance for structured fine-grained sparse AI workloads, for example by accelerating sparse matrix multiplication required to execute or train AI models. Sparse data is compressed to remove zero-valued elements before being streamed into a matrix multiplication unit (MXU) of the accelerator. The accelerator stores a gains matrix, which can be the matrix for multiplying with the received input matrix. The accelerator uses an index array mapping locations of elements in the compressed matrix with locations of elements in the matrix's precompressed form, to generate a multiplier matrix from the gains matrix. Aspects of the disclosure also provide for accelerated gains matrix loading in a hardware accelerator or other type of processor. The accelerator can load the gains matrix more efficiently in a compressed form, and then un-compress the matrix once loaded.

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a diffusion neural network using a region-aware fine-tuning process. After training, the diffusion neural network can be used to generate an image conditioned on a conditioning input.

Methods, systems, and apparatuses, including computer programs encoded on computer storage media, for training a target generative neural network over a plurality of training iterations. At each iteration, a first data item is generated by processing a conditioning input using the target generative neural network. An improvement generative neural network then processes the first data item and the conditioning input to generate a second, preferred data item. A training example is generated that includes the first and second data items and indicates that the second data item is preferred over the first. The target generative neural network is then trained on this training example. By using this iterative process to dynamically generate preference data, the described techniques improve the performance of the generative neural network beyond the limitations of static, offline datasets without requiring computationally expensive reward models or external human annotation.

A media application receives an initial image and a textual request to change the initial image, the initial image including a subject with a face. The media application generates, from the initial image, a preserving mask that corresponds to the face of the subject. The media application provides the textual request, the initial image, and the preserving mask as input to a diffusion model. The diffusion model outputs a denoised initial image based on the initial image; performs text conditioning of the textual request and forward diffusion to generate a noisy translated image that satisfies the textual request; and outputs based on the noisy translated image, extracted features, and self-attention maps, a denoised translated image. The media application blends the denoised initial image, the preserving mask, and the denoised translated image to form an output image, wherein the preserving mask prevents modification to the face from the initial image.

The technology is directed to error detection and correction in a memory device containing dynamic random-access memory (DRAM) elements. The DRAM elements are configured to store data in a first number of elements, store error correction code (ECC) information in a second number of elements and storing an error correction code in a third number of elements. The amount of memory allocated to ECC is reduced without adversely affecting the ECC capability by storing an error correction code, such as a cyclic redundancy code (CRC) in the freed-up memory space. A location of an uncorrectable error is identified assuming one of the elements contains an error then recalculating the selected data using the ECC, recalculating the CRC and comparing it to the original CRC. A match indicates the faulty value has been found and corrected. A mismatch indicates that the selected element is not the location of the error.

The technology is generally directed to error detection and correction in memory devices using dynamic random-access memory (DRAM) elements. Error correction codes (ECC) provide the ability to correct errors in memory elements. The ability to extend ECC capabilities to correct errors in defective elements is achievable when the location of the error(s) is known. Error correction metadata is calculated in the memory device when a write operation is performed. The metadata is stored in available space on the memory device. To locate uncorrectable errors, a pair of elements is selected and ECC is performed to recreate the data in the pair of elements. The metadata is recalculated on the recreated data and compared to the originally stored metadata. A match in metadata indicates the chosen pair contained the error. A mismatch indicates that the recreated data does not contain the error. Locating errors extends ECC correction capability.

Implementations relate to determination and visual display of spoken menus for calls. In some implementations, a computer-implemented method includes receiving audio data output in a call between a call device and a device associated with a target entity. The audio data includes speech indicating one or more selection options for a user of the call device to navigate through a call menu provided by the target entity in the call. Text is determined by programmatically analyzing the audio data, the text representing the speech. The selection options are determined based on programmatically analyzing at least one of the text or the audio data. At least a portion of the text is displayed by the call device during the call, as one or more visual options that correspond to the selection options. The visual options are each selectable via user input to cause corresponding navigation through the call menu.

The present disclosure describes apparatuses and techniques of adaptive PLMN management for varying network conditions. In some aspects, a PLMN manager (144) of a user equipment (UE, 102) searches, as part of a registration procedure, a data repository of PLMN information that includes a PLMN (104, 106) blocked from registration for a duration of time and determines that the PLMN is blocked due to a signal-related condition with a cell (110, 112) of the PLMN (e.g., weak signal). The PLMN manager (144) then determines, during the duration of time, that the signal-related condition between the UE (102) and the cell (110, 112) has improved. In response to the improvement, the PLMN manager (144) unblocks the PLMN (104, 106) before expiration of the duration of time. By so doing, the UE (102) may reattempt registration with the unblocked PLMN (104, 106) at least once before the duration of time expires, which may allow the UE to register more quickly and reestablish network services.

A method for on-device artificial intelligence (AI) includes obtaining, at a client device, a request to perform a task. The method includes providing the request as input to a generative AI model operating on the client device. The generative AI model may be trained to identify one or more applications capable of performing the task and to generate instructions related to the task for the one or more applications. The method includes causing the instructions related to the task to be provided to the one or more applications. The method includes receiving an indication of completion of at least a portion of the task. The method includes providing, at the client device, information regarding the completion of the at least a portion of the task.

A method (500) includes obtaining proximity information (250) for each of a plurality of assistant-enabled devices (210) within an environment (32) of a user device (200). Each assistant-enabled device is controllable by an assistant application (106) to perform a respective set of available actions (120) associated with the assistant-enabled device. For each assistant-enabled device, the method also includes determining a proximity score (260) based on the proximity information indicating a proximity estimation of the corresponding assistant-enabled device relative to the user device. The method further includes generating, using the proximity scores determined for the assistant-enabled devices, a ranked list (310) of candidate assistant-enabled devices, and for each corresponding assistant-enabled device in the ranked list, displaying, in a graphical user interface (GUI) (400), a respective set of controls (220) for performing the respective set of actions associated with the corresponding assistant-enabled device.

Generally disclosed herein is a mechanism to mitigate power fluctuations of a data center by dynamically charging and discharging a battery energy storage system (BESS). According to some examples, a BESS control system can be configured to monitor power demand fluctuations and initiate charging of the BESS during periods of low power demand. One example of a period of low power demand includes active idle, wherein the workloads of the server machines frequently decrease by a small magnitude. Another example of a period of low power demand includes deep idle, wherein the workloads of the server machines decrease by a larger magnitude than the active idle for a longer duration. The system may discharge the power from the BESS during the peak power demand. The BESS control system may stabilize the data center power system through fast-acting voltage and frequency control at a system level.

A system for repair of an accelerator architecture including a plurality of accelerator host having a plurality of accelerator elements in each host and high-speed interconnects between accelerator elements and including a link health monitor service in the host determining a status of each interconnect internal and external to the host. The link health monitor service detects link status of the interconnects and communicates this to a pod manager responsible for managing a number of interconnected hosts. The pod manager communicates with a scheduler for scheduling processing tasks to each accelerator element or group of elements. Upon detection of an interconnect problem, the pod manager flags a least common ancestor in a binary tree of elements containing elements adjacent to the failed interconnect and flags the ancestor and its parent nodes as unavailable. Pod manager further generates a corrective action repairing the problem interconnect communicating the action to a technician.

The technology is directed to systems and methods of high-bandwidth memory allocation. High-bandwidth memory may include a plurality of data channels and shareable memory that can be selectively allocated to particular data channels. In addition, bandwidth may be selectively allocated to the data channels independent of the shareable memory. The allocation of memory and bandwidth to particular data channels may be based on identified attributes of workloads that are to be associated with each data channel.

A liquid-cooled rack for liquid-cooling trays of computing equipment includes an inlet manifold including an inlet for receiving a coolant at a first temperature and at least one outlet interface for discharging the coolant to one of the liquid-cooled trays; and an outlet manifold including at least one inlet interface for receiving the coolant at a second temperature higher than the first temperature from one of the liquid-cooled trays and an outlet for discharging the coolant. At least one of the inlet manifold or the outlet manifold includes at least one coolant drain port to facilitate the draining and purge drying process. The drain ports can also be added to each tray. The draining and purge drying process can be enhanced by using high pressure air (desiccated and/or heated), flushing with volatile fluid, tilting, vibration, vacuuming.

Generally disclosed herein is a mechanism to mitigate power fluctuations of a data center by dynamically charging and discharging a capacitor-based energy storage device. According to some examples, a supercapacitor control system can be configured to monitor power demand fluctuations and initiate charging of the supercapacitor during a workload ramp-up period for training machine learning (ML) models with massive training data. As the system starts discharging the supercapacitors during the ramp-up period, the system is capable of reducing the rate of the power demand increase. Similarly, during a ramp-down period, the system is configured to charge the supercapacitor as the training of the ML model is being finished and the power demand decreases. The power stored in the supercapacitors may be used to reduce the rate of the power demand decrease when the training of the ML model is being completed, thereby maintaining the stability of the power utility.

Various implementations include reducing latency when interacting with a generative model system based on generating predicted complete text based on natural language (NL) text input, where the NL text input is a portion of a user query. In many implementations, predicted completion text can be generated by processing NL text input using a language model. In several implementations, the system can perform initial processing of the NL input text and the predicted completion text (e.g., preform initial preprocessing of the NL input text and predicted completion text for processing using the generative model, performing an initial limited decoding of output using the generative model, etc.). The user can confirm the predicted completion text before the system continues processing the NL input and predicted completion text using the generative model to generate output.

An insect storage and release device includes suitably rigid container that may be used to store and transport insects. The insect storage and release device may hold compressed insects or non-compressed insects. The insect storage and release device includes a removable enclosure and a reticulated foam for adding compliance during packing of the insects.

A media application receives user input that indicates one or more objects to be erased from a media item. The media application translates the user input to a bounding box. The media application provides a crop of the media item based on the bounding box to a segmentation machine-learning model. The segmentation machine-learning model outputs a segmentation mask for one or more segmented objects in the crop of the media item and a corresponding segmentation score that indicates a quality of the segmentation mask.

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …

Zusammenfassung wird geladen …