Abstract: Information recommendation system usually involve a multitask problem, which tries to predict not only users' click-through rate (CTR) but also the post-click conversion rate (CVR). At the same time, for multi-functional information systems, there are commonly multiple services for users, such as news feed, search engine, and product suggestions. The prediction/ranking model should be conducted in a multi-scene manner. In the present patent document, embodiments of a unified ranking model for such a multi-task and multi-scene problem are disclosed. The disclosed model explores independent and non-shared embeddings for each task and scene, which reduces the coupling between tasks and scenes. Therefore, new tasks or scenes may be added easily. Besides, a simplified network may be chosen beyond the embedding layer, which largely improves the ranking efficiency for various online services. Extensive offline and online experiments demonstrated the superiority of model embodiments.

Abstract: Sound signals are received by one or more microphones disposed at an ADV. The sound signals are analyzed to extract a feature of a road on which the ADV is driving. A road condition of the road is determined based on the extracted feature. A path planning and speed planning is performed to generate a trajectory based on the road condition. The ADV is controlled to drive autonomously according to the generated trajectory.

Abstract: A cooling system includes a container and a number of battery packaging modules submerged in a first coolant fluid contained in the container. The cooling system includes a supply channel and a return channel coupled respectively to at least a supply end and at least a return end of the battery packaging modules to accelerate a second coolant fluid through any of the battery packaging modules. The cooling system includes one or more pairs of fluid valves secured to the supply and return channels, each pair to control the second coolant fluid flowing through a subset of the plurality of battery packaging modules. The cooling system includes a fluid pump disposed at the channels to accelerate the second coolant fluid supplied to any of the battery packaging modules which are in activated modes.

Abstract: Embodiments are disclosed of a fluid distribution apparatus. The fluid distribution apparatus includes a hot manifold including a hot chamber fluidly coupled to one or more fluid return inlets and a main fluid return outlet, and a cold manifold including a cold chamber fluidly coupled to a main fluid supply inlet and one or more fluid supply outlets. A thermoelectric device is sandwiched between the hot manifold and the cold manifold so that the thermoelectric device is in thermal contact with the hot chamber and in thermal contact with the cold chamber. The apparatus is connected to either a server energy storage unit or a rack mounted energy unit through dedicated busbar.

Abstract: An information technology (IT) enclosure includes an IT container having immersion coolant self-contained therein, one or more cooler trays and one or more dedicated battery spacings that are alternately arranged in a series manner within the IT container. Each cooler tray to house a cooler, each battery spacing to house one or more rows of battery packs. The IT enclosure includes a supply channel disposed at a first side of the IT container, and a return channel and a fluid pump disposed at another side of the IT container. Where, when in operation, the immersion coolant circulates amongst the alternate one or more rows of battery packs and coolers to transfer a thermal load from the one or more rows of battery packs to the coolers.

Abstract: The present disclosure provides apparatus, systems, methods, and techniques for cooling using a fluid management hardware that allows for directly recirculating a liquid portion of two-phase fluids. In a core unit, a merging region is designed for the liquid portion may be merged with supply coolant to return to cooling the system, while the vapor portion may be transferred to external cooling and condensed to liquid to return to the coolant supply source. A pump is directly intaking coolant from the merging region. In an embodiment, the supply coolant may directly connected to the pump intaking loop. The cooling techniques may be applied to an advanced server rack having high power density servers. In some cases, the disclosed design may be used for distributing two phase coolant to servers with either cooling plate/cooling module or the server is immersion based packaged.

Abstract: Deep neural networks (DNN) model quantization may be used to reduce storage and computation burdens by decreasing the bit width. Presented herein are novel cursor-based adaptive quantization embodiments. In embodiments, a multiple bits quantization mechanism is formulated as a differentiable architecture search (DAS) process with a continuous cursor that represents a possible quantization bit. In embodiments, the cursor-based DAS adaptively searches for a quantization bit for each layer. The DAS process may be accelerated via an alternative approximate optimization process, which is designed for mixed quantization scheme of a DNN model. In embodiments, a new loss function is used in the search process to simultaneously optimize accuracy and parameter size of the model. In a quantization step, the closest two integers to the cursor may be adopted as the bits to quantize the DNN together to reduce the quantization noise and avoid the local convergence problem.

Abstract: Systems and methods are disclosed for migrating a virtual machine (VM) having a virtual function that maps resources of an artificial intelligence (AI) accelerator to the VM. A driver for the AI accelerator can generate a checkpoint of VM processes that make calls to the AI accelerator, and can the checkpoint can include a list and configuration of resources mapped to the AI accelerator by the virtual function. The driver can also access the code, data, and memory of the AI accelerator to generate a checkpoint of the AI accelerator status. When the VM is migrated to a new host, then either, or both, of these checkpoint frames can be used to ensure that resuming the VM on a new host having appropriate AI accelerator resources, can be successful resumed on the new host. One or both checkpoint frames can be captured based upon an event, in anticipation of the need to migrate the VM.

Abstract: Embodiments are disclosed of an IT rack and a server including a cooling module. An integrated fluid distribution manifold set with a supply manifold is on the front of the rack and a return manifold is on the rear of the rack. The supply manifold extends from the front to the rear and rack connectors, which fluidly couple the rack to a data center facility system, are on the top of the rack. Supply and return connectors are positioned so that the connectors face each other within the rack. Each server includes a cooling module with a movable connector set that can extend outside the server chassis to connect with the manifolds. The cooling module's connector set includes a transmission structure that allows the supply and return connectors of the connector set to engage with connectors on the supply and return manifolds.

Abstract: Embodiments of the disclosure discloses a method and system of a virtualization environment for a data processing (DP) accelerator. In one embodiment, a data processing (DP) accelerator includes a resource management unit and one or more dynamically isolated resources managed by the resource management unit. The DP accelerator includes one or more virtual functions (VFs) each associated with one of the one or more dynamically isolated resources, where a virtual machine (VM) of a host is assigned one of the one or more VFs to access the dynamically isolated resources associated with the assigned VF, and where the VM has no access to the rest of the one or more dynamically isolated resources.

Abstract: A cooling plate for cooling high power density electronics has an internal cavity and an opening for fluid exchange with the cavity. A mounting structure is positioned within the opening. A coaxial port is attached to the mounting structure, the coaxial port having a center conduit and a ring conduit surrounding the central conduit such that rotational axis of the center conduit coincides with rotational axis of the ring conduit. A single coaxial port can serve to deliver cooling liquid to the cooling plate and return warmed fluid from the cooling plate. The coaxial port center conduit connected with a fluid distribution panel. Fluid distribution is regulated by the panel before it exits the port through the ring conduit.

Abstract: Modern deep neural network (DNN) models have many layers with a single layer potentially involving large matrix multiplications. Such heavy calculation brings challenges to deploy such DNN models on a single edge device, which has relatively limited computation resources. Therefore, multiple and even heterogeneous edge devices may be required for applications with stringent latency requirements. Disclosed in the present patent documents are embodiments of a model scheduling framework that schedules multiple models on a heterogeneous platform. Multiple-model heterogeneous computing is partitioned into a neural computation optimizer (NCO) part and a neural computation accelerator (NCA) part. The migration, transition, or transformation of DNN models from cloud to edge is handled by the NCO, while the deployment of the transformed DNN models on the heterogeneous platform is handled by the NCA. Such a separation of implementation simplifies task execution and improves the flexibility for the overall framework.

Abstract: Embodiments of a system/method is disclosed to operate an autonomous driving vehicle (ADV). In one embodiment, a system perceives a driving environment surrounding the ADV using a plurality of sensors mounted on the ADV including one or more obstacles. The system receives traffic signal information from one or more traffic indicators identified within a predetermined radius of the ADV. For each of the one or more obstacles, the system determines if the obstacle is situated on a lane with traffic flow coordinated by the one or more traffic indicators. The system predicts a behavior of the obstacle based on the traffic signal information for the lane. The system plans a trajectory based on the predicted behaviors for the one or more obstacles to control the ADV based on the planned trajectory.

Abstract: A sensor aggregation framework for autonomous driving vehicles is disclosed. In one embodiment, sensor data is collected from one or more sensors mounted on an autonomous driving vehicle (ADV) while the ADV is moving within a region of interest (ROI) that includes a number of obstacles. The sensor data includes obstacle information of the obstacles and vehicle data of the ADV. Each of the vehicle data is timestamped with a current time at which the vehicle data is captured to generate a number of timestamps that correspond to the vehicle data. The obstacle information, the vehicle data, and the corresponding timestamps are aggregated into training data. The training data is used to train a set of parameters that is subsequently utilized to predict at least in part future obstacle behaviors and vehicle movement of the ADV.

Abstract: Deep neural network (DNN) models have been widely used for user-relevance content prediction. Presented herein are embodiments of a new user-relevance framework, which may be referred as Gating-Enhanced Multi-task Neural Networks (GemNN) embodiments. Neural network-based multi-task learning model embodiments herein predict user engagement with content in a coarse-to-fine manner, which gradually reduces content candidates and allows parameter sharing from upstream tasks to downstream tasks to improve the training efficiency. Also, in one or more embodiments, a gating mechanism was introduced between embedding layers and multi-layer perceptions to learn feature interactions and control the information flow fed to MLP layers. Tested embodiments demonstrated considerable improvements over prior approaches.

Abstract: A liquid manifold can be assembled to an information technology (IT) rack to deliver and distribute fluid to IT equipment. The manifold can include a plurality of sections, each of the plurality of sections having one or more shut-off valves. One or more leak detection sensors can be arranged to detect leaks in any of the sections and in any of the IT equipment. A controller can control a shut-off valve to a closed position based on a detected leak. The design enables the manifold to better manage and control the fluid for mission critical IT equipment.

Abstract: Presented herein are embodiments of an improved asymmetric quantization, which may generally be referred to as improved asymmetric quantization (IAQ) embodiments. IAQ embodiments combine the benefits of conventional asymmetric quantization and symmetric quantization but also provide additional computation efficiencies. Embodiments of IAQ adopt an asymmetric range of the weights of a neural network layer, so they circumvent the limitation of symmetric range of symmetric quantization. Moreover, the inference process of a neural network quantized by an IAQ embodiment is much faster than that of the neural network quantized by conventional asymmetric quantization by quantizing an offset value of each layer.

Abstract: A cooling module includes a first cooling plate having a first internal channel and a second cooling plate having a second internal channel. The cooling module includes an interconnect frame coupled in between the first and second cooling plates, the interconnect frame includes a third internal channel that connects the first internal channel to the second internal channel. The cooling module includes a first injection plate attached to a bottom portion of the first cooling plate and a second injection plate attached to a bottom portion of the second cooling plate, the first and second injection plates manage a distribution of coolant fluid to dedicated areas of electronic chips adjacent to the first and second injection plates. The cooling module includes a first pump frame coupled to an inlet port at the first cooling plate, a first pump disposed at the first pump frame to directly intake a coolant fluid.

Abstract: Image super-resolution (SR) refers to the process of recovering high-resolution (HR) images from low-resolution (LR) inputs. Blind image SR is a more challenging task which involves unknown blurring kernels and characterizes the degradation process from HR to LR. In the present disclosure, embodiments of a variational autoencoder (VAE) are leveraged to train a kernel autoencoder for more accurate degradation representation and more efficient kernel estimation. In one or more embodiments, a kernel-agnostic loss is used to learn more robust kernel features in the latent space from LR inputs without using ground-truth kernel references. In addition, attention-based adaptive pooling is introduced to improve kernel estimation accuracy, and spatially non-uniform kernel features are passed into SR restoration resulting in additional kernel estimation error tolerance.

Abstract: Modem deep neural network (DNN) models have many layers with a single layer potentially involving large matrix multiplications. Such heavy calculation brings challenges to deploy such DNN models on a single edge device, which has relatively limited computation resources. Therefore, multiple and even heterogeneous edge devices may be required for applications with stringent latency requirements. Disclosed in the present patent documents are embodiments of a model scheduling framework that schedules multiple models on a heterogeneous platform. Two different approaches, model first scheduling (MFS) and hardware first scheduling (HFS), are presented to allocate a group of models for a service into corresponding heterogeneous edge devices, including CPU, VPU and GPU. Experimental results prove the effectiveness of the MFS and HFS methods for improving the inference speed of single and multiple AI-based services.