Abstract: A hardware accelerator can receive, from a host processor, a slice of input data at a time-step. The hardware accelerator can process the input data using a machine learning model deployed on the hardware accelerator to compute a respective probability among multiple probabilities for each of multiple classes. The respective probability for each class being a likelihood that content in the slice belongs to the class. The hardware accelerator can determine, from the multiple probabilities, a preset number of highest probabilities for the slice of input data. The hardware accelerator can transmit the preset number of highest probabilities for the slice to the host processor. Related apparatus, systems, techniques and articles are also described.

Abstract: This disclosure describes a system and method for compiling and executing machine learning inferences in an array of multi-core computing devices. Each multi-core computing device can be an application specific integrated circuit (ASIC) or group of ASICS. In many applications, the array of computing devices changes from inference to inference, and can be adjusted based on the requirements of the inference. Additionally, each ASIC can have multiple processing cores, and multiple types of processing cores. Therefore, performing optimizations and scheduling at compile time, can dramatically increase the efficiency of the array in executing the inference. In some implementations, it is possible to select an amount of time or effort to be spent optimizing during compiling, giving the user flexibility in determining whether to spend time during compilation or during execution.

Abstract: Some embodiments of the present disclosure relate to a memory device. The memory device includes an active current path including a data storage element; and a reference current path including a reference resistance element. The reference resistance element has a resistance that differs from a resistance of the data storage element. A delay-sensing element has a first input coupled to the active current path and a second input coupled to the reference current path. The delay-sensing element is configured to sense a timing delay between a first signal on the active current path and a second signal on the reference current path. The delay-sensing element is further configured to determine a data state stored in the data storage element based on the timing delay.

Abstract: Some embodiments of the present disclosure relate to a memory device. The memory device includes an active current path including a data storage element; and a reference current path including a reference resistance element. The reference resistance element has a resistance that differs from a resistance of the data storage element. A delay-sensing element has a first input coupled to the active current path and a second input coupled to the reference current path. The delay-sensing element is configured to sense a timing delay between a first signal on the active current path and a second signal on the reference current path. The delay-sensing element is further configured to determine a data state stored in the data storage element based on the timing delay.

Abstract: An integrated circuit includes a device, a first interconnect structure disposed above the device and a second interconnect structure positioned below the device. The first interconnect structure includes multiple frontside metal layers. The second interconnect structure includes multiple backside metal layers, where each backside metal layer includes metal conductors routed according to diagonal routing. In some embodiments, a backside interconnect structure can include another backside metal layer that includes metal conductors routed according to mixed-Manhattan-diagonal routing. A variety of techniques can be used to route signals between metal conductors in the backside interconnect structure and cells on one or more frontside metal layers.

Abstract: Methods and systems, including computer programs encoded on a computer storage medium. In one aspect, a method includes obtaining data specifying one or more neural networks to be deployed on a neural network hardware accelerator, each of the one or more neural networks having a respective set of parameters, and the neural network hardware accelerator having one or more memories having a memory capacity; determining a maximum amount of the memory capacity that will be in use at any one time during a processing of any of the one or more neural networks by the neural network hardware accelerator; identifying a subset of the parameters of the one or more neural networks that consumes an amount of memory that is less than a difference between the memory capacity and the determined maximum amount of the memory capacity; and storing the identified subset of the parameters.

Abstract: An integrated circuit includes a device, a first interconnect structure disposed above the device and a second interconnect structure positioned below the device. The first interconnect structure includes multiple frontside metal layers. The second interconnect structure includes multiple backside metal layers, where each backside metal layer includes metal conductors routed according to diagonal routing. In some embodiments, a backside interconnect structure can include another backside metal layer that includes metal conductors routed according to mixed-Manhattan-diagonal routing. A variety of techniques can be used to route signals between metal conductors in the backside interconnect structure and cells on one or more frontside metal layers.

Abstract: A hardware accelerator can store, in multiple memory storage areas in one or more memories on the accelerator, input data for each processing time step of multiple processing time steps for processing sequential inputs to a machine learning model. For each processing time step, the following is performed. The accelerator can access a current value of a counter stored in a register within the accelerator to identify the processing time step. The accelerator can determine, based on the current value of the counter, one or more memory storage areas that store the input data for the processing time step. The accelerator can facilitate access of the input data for the processing time step from the one or more memory storage areas to at least one processor coupled to the one or more memory storage areas. The accelerator can increment the current value of the counter stored in the register.

Abstract: A method includes: providing a cell of a first group that supplies a first potential from a backside of a substrate, multiple cells of a second group, and two cells of a third group that supply a second potential from the backside; determining a distance in a row direction between the cell of the first group and each of the two cells of the third group; determining a placement of the cell of the first group, the two cells of the third group, and each of the cells of the second group; and counting a number of pins of the cells of the second group. The cell of the first group is located between the two cells of the third group. Each of the cells of the second group is located between the two cells of the third group.

Abstract: Disclosed embodiments herein relate to an integrated circuit including power switches with active regions connected to form a contiguous region. In one aspect, the integrated circuit includes a first layer including a first metal rail extending in a first direction. In one aspect, the integrated circuit includes a second layer above the first layer along a second direction perpendicular to the first direction. The second layer may include active regions for power switches. In one aspect, the active regions of the power switches are connected to form a contiguous region extending in the first direction. The first metal rail may be electrically coupled to the active regions through via contacts. In one aspect, the integrated circuit includes a third layer above the second layer along the second direction. The third layer may include a second metal rail electrically coupled to some of the power switches through additional via contacts.

Abstract: A method includes: providing a cell of a first group that supplies a first potential from a backside of a substrate, multiple cells of a second group, and two cells of a third group that supply a second potential from the backside; determining a distance in a row direction between the cell of the first group and each of the two cells of the third group; determining a placement of the cell of the first group, the two cells of the third group, and each of the cells of the second group; and counting a number of pins of the cells of the second group. The cell of the first group is located between the two cells of the third group. Each of the cells of the second group is located between the two cells of the third group.

Abstract: One aspect of this description relates to an integrated circuit (IC) structure including a first layer and a second layer. The first layer includes a first metal structure coupled to a first power supply having a first voltage level and a second metal structure coupled to a second power supply having a second voltage level different from the first voltage level. The second layer is formed over the first layer. The second layer includes a first nano-sheet device coupled to the first metal structure and a second nano-sheet device adjacent to the first nano-sheet device. The second nano-sheet device is coupled to the second metal structure. A distance between the first nano-sheet device and the second nano-sheet device is less than a minimum n-well to n-well spacing.

Abstract: An IC device includes a first interconnect structure configured to distribute a power supply voltage, a second interconnect structure configured to distribute a reference voltage, a third interconnect structure configured to distribute a first power gating signal, and a plurality of logic circuits. Each logic circuit of the plurality of logic circuits includes a power switch coupled in series with a logic gate between the first and second interconnect structures, and the power switch includes a control terminal coupled to the third interconnect structure.

Abstract: In some embodiments, a semiconductor device is provided. The semiconductor device includes a semiconductor substrate having a first semiconductor material layer separated from a second semiconductor material layer by an insulating layer. A first access transistor is arranged on the first semiconductor material layer, where the first access transistor has a pair of first source/drain regions having a first doping type. A second access transistor is arranged on the first semiconductor material layer, where the second access transistor has a pair of second source/drain regions having a second doping type opposite the first doping type. A resistive memory cell having a bottom electrode and an upper electrode is disposed over the semiconductor substrate, where one of the first source/drain regions and one of the second source/drain regions are electrically coupled to the bottom electrode.

Abstract: A helmet comprising first and second parts and a connector connecting the first and second parts of the helmet, the connector comprising: a first plate; a first anchor point on one side of the first plate, configured to be connected to the first part; a second plate located on an opposite side of the first plate from the first anchor point; a second anchor point on an opposite side of the second plate from the first plate, configured to be connected to the second part; a low friction interface provided between opposing surfaces of the first and second plates; and a cuff of deformable material provided around the first and second plates, and configured to at least partially cover the side of the first plate on which the first anchor point is located and to at least partially cover the side of the second plate on which the second anchor point is located.

Abstract: Disclosed embodiments herein relate to an integrated circuit including power switches with active regions connected to form a contiguous region. In one aspect, the integrated circuit includes a first layer including a first metal rail extending in a first direction. In one aspect, the integrated circuit includes a second layer above the first layer along a second direction perpendicular to the first direction. The second layer may include active regions for power switches. In one aspect, the active regions of the power switches are connected to form a contiguous region extending in the first direction. The first metal rail may be electrically coupled to the active regions through via contacts. In one aspect, the integrated circuit includes a third layer above the second layer along the second direction. The third layer may include a second metal rail electrically coupled to some of the power switches through additional via contacts.

Abstract: Disclosed embodiments herein relate to an integrated circuit including power switches with active regions connected to form a contiguous region. In one aspect, the integrated circuit includes a first layer including a first metal rail extending in a first direction. In one aspect, the integrated circuit includes a second layer above the first layer along a second direction perpendicular to the first direction. The second layer may include active regions for power switches. In one aspect, the active regions of the power switches are connected to form a contiguous region extending in the first direction. The first metal rail may be electrically coupled to the active regions through via contacts. In one aspect, the integrated circuit includes a third layer above the second layer along the second direction. The third layer may include a second metal rail electrically coupled to some of the power switches through additional via contacts.

Abstract: One aspect of this description relates to an integrated circuit (IC) structure including a first layer and a second layer. The first layer includes a first metal structure coupled to a first power supply having a first voltage level and a second metal structure coupled to a second power supply having a second voltage level different from the first voltage level. The second layer is formed over the first layer. The second layer includes a first nano-sheet device coupled to the first metal structure and a second nano-sheet device adjacent to the first nano-sheet device. The second nano-sheet device is coupled to the second metal structure. A distance between the first nano-sheet device and the second nano-sheet device is less than a minimum n-well to n-well spacing.

Abstract: In some embodiments, a semiconductor device is provided. The semiconductor device includes a semiconductor substrate having a first semiconductor material layer separated from a second semiconductor material layer by an insulating layer. A source region and a drain region are disposed in the first semiconductor material layer and spaced apart. A gate electrode is disposed over the first semiconductor material layer between the source region and the drain region. A first doped region having a first doping type is disposed in the second semiconductor material layer, where the gate electrode directly overlies the first doped region. A second doped region having a second doping type different than the first doping type is disposed in the second semiconductor material layer, where the second doped region extends beneath the first doped region and contacts opposing sides of the first doped region.

Abstract: A hardware accelerator can store, in multiple memory storage areas in one or more memories on the accelerator, input data for each processing time step of multiple processing time steps for processing sequential inputs to a machine learning model. For each processing time step, the following is performed. The accelerator can access a current value of a counter stored in a register within the accelerator to identify the processing time step. The accelerator can determine, based on the current value of the counter, one or more memory storage areas that store the input data for the processing time step. The accelerator can facilitate access of the input data for the processing time step from the one or more memory storage areas to at least one processor coupled to the one or more memory storage areas. The accelerator can increment the current value of the counter stored in the register.