Abstract: A hybrid processing fabric for doing training and inference on neural networks comprises heterogeneous computational resources for performing both arithmetic and logic operations such that some layers of the neural network are implemented with arithmetic multiply-and-accumulate processing elements whereas other layers are implemented with fixed-function logic processing elements. Support circuitry including memory, data transformation modules, and I/O interfaces are also included in the fabric. A compiler is introduced to optimize and map high level descriptions of the neural network model to the said hybrid fabric.

Abstract: This patent discloses novel design and implementations of high-performance and scalable Montgomery modular multiplier circuits by utilizing and developing a combination of optimization techniques and dataflow transformations including use of carry-save compressions, multiplier decomposition using a radix of 2m, multiplicand decomposition using a radix of 2w, parallelization of computations of the quotient and intermediate results in each iteration of the Montgomery modular multiplication, replacement of multiplications and additions in each iteration with simple encoding and compression operations, and correction of potential overflows in intermediate results by doing a simple 2-bit addition.

Abstract: An SFQ circuit system includes at least one SFQ block having a plurality of SFQ logic gates. Characteristically, at least a portion of the SFQ logic gates are arranged in series. The SFQ circuit system includes a timing system configured to provide a first set of inputs and collect a first set of outputs of the at least one SFQ block at a rate defined by a slow clock frequency while the SFQ logic gates are clocked at a fast clock frequency. Advantageously, the rate is sufficiently slow to allow the first set of inputs to propagate through all levels of the SFQ logic gates to produce the first set of outputs of the at least one SFQ block without colliding with a second set of inputs to the at least one SFQ block.

Abstract: An SFQ circuit system includes at least one SFQ block having a plurality of SFQ logic gates. Characteristically, at least a portion of the SFQ logic gates are arranged in series. The SFQ circuit system includes a timing system configured to provide a first set of inputs and collect a first set of outputs of the at least one SFQ block at a rate defined by a slow clock frequency while the SFQ logic gates are clocked at a fast clock frequency. Advantageously, the rate is sufficiently slow to allow the first set of inputs to propagate through all levels of the SFQ logic gates to produce the first set of outputs of the at least one SFQ block without colliding with a second set of inputs to the at least one SFQ block.

Abstract: A superconducting field programmable gate array (SuperFPGA) apparatus for implementing a superconducting electronic circuit includes a superconducting logic core that includes a plurality of superconducting single flux quantum configurable logic blocks having regular Josephson junctions and inductors that are interconnectible to each other and to input/output terminals of the superconducting electronic circuit. The SuperFPGA apparatus also includes a superconducting routing network, a zero-static-power dissipation biasing network, magnetic Josephson junctions, and a magnetic Josephson junction programming layer.

Abstract: A superconducting field programmable gate array (SuperFPGA) apparatus for implementing a superconducting electronic circuit includes a superconducting logic core that includes a plurality of superconducting single flux quantum configurable logic blocks having regular Josephson junctions and inductors that are interconnectible to each other and to input/output terminals of the superconducting electronic circuit. The SuperFPGA apparatus also includes a superconducting routing network, a zero-static-power dissipation biasing network, magnetic Josephson junctions, and a magnetic Josephson junction programming layer.

Abstract: An embodiment of the present invention is directed to a method for determining a pixel transformation function that maximizes backlight dimming while maintaining a pre-specified distortion level. The method includes determining a minimum dynamic range of pixel values in a transformed image based on an original image and the pre-specified distortion level and determining the pixel transformation function. The pixel transformation function takes a histogram of the original image to a uniform distribution histogram having the minimum dynamic range.

Abstract: In one embodiment, a circuit includes a first row of circuit blocks that are each connected to a supply directly and to ground via a first sleep transistor. A connection between the first circuit block and the first sleep transistor is a virtual ground node. The circuit includes a second row of circuit blocks that are each connected to ground directly and to the supply via a second sleep transistor. A connection between the second circuit block and the second sleep transistor is a virtual supply node. The circuit includes a transmission gate (TG) or pass transistor connecting the virtual ground nodes to the virtual supply nodes to enable charge recycling between circuit blocks in the first row and circuit blocks in the second row during transitions by the circuit from active mode to sleep mode, from sleep mode to active mode, or both.

Abstract: In one embodiment, a method for power mode transition in a multi-threshold complementary metal oxide semiconductor (MTCMOS) circuit includes clustering logic cells in the circuit to a number of logic clusters and optimizing wake-up times of the logic clusters to reduce a total turn-on time of the circuit while keeping below a predetermined threshold a sum of currents flowing from the circuit to ground, a sum of currents flowing from a supply voltage to the circuit, or both during a transition by the circuit from sleep mode to active mode.

Abstract: In one embodiment, a memory block includes one or more bit lines that each include two or more cells. Each cell in each bit line has a distance from a sense amplifier coupled to the bit line, and each of one or more of the cells in each of one or more of the bit lines has a delay particularly set according to the distance of the cell from the sense amplifier coupled to the bit line.

Abstract: In one embodiment, a circuit includes a first row of circuit blocks that are each connected to a supply directly and to ground via a first sleep transistor. A connection between the first circuit block and the first sleep transistor is a virtual ground node. The circuit includes a second row of circuit blocks that are each connected to ground directly and to the supply via a second sleep transistor. A connection between the second circuit block and the second sleep transistor is a virtual supply node. The circuit includes a transmission gate (TG) or pass transistor connecting the virtual ground nodes to the virtual supply nodes to enable charge recycling between circuit blocks in the first row and circuit blocks in the second row during transitions by the circuit from active mode to sleep mode, from sleep mode to active mode, or both.

Abstract: In one embodiment, a circuit includes a first circuit block connected to ground via a first sleep transistor, a first virtual ground node between the first circuit block and the first sleep transistor, a second circuit block connected to ground via a second sleep transistor, a second virtual ground node between the second circuit block and the second sleep transistor, and a transmission gate (TG) or a pass transistor connecting the first virtual ground node to the second virtual ground node to enable charge recycling between the first circuit block and the second circuit block during transitions by the first circuit block from active mode to sleep mode and the second circuit block from sleep mode to active mode or vice versa.

Abstract: A method of forming a memory cell includes coupling a first transistor between a supply rail of a memory cell and a node operable to accept a supply voltage. The method further includes coupling a second transistor between a ground rail of the cell and a node operable to accept a ground. In one embodiment, the method includes forming the cell to accept selectively applied external voltages, wherein the external voltages are selected to minimize leakage current in the cell. In another embodiment, the method includes forming at least one of the first and the second transistors to have a channel width and/or a threshold voltage selected to minimize a total leakage current in the cell.

Abstract: In one embodiment, a circuit includes a first circuit block connected to ground via a first sleep transistor, a virtual ground node between the first circuit block and the first sleep transistor, a second circuit block connected to a supply via a second sleep transistor, and a virtual supply node between the second circuit block and the second sleep transistor. The circuit also includes a transmission gate (TG) or a pass transistor connecting the virtual ground node to the virtual supply node to enable charge recycling between the first circuit block and the second circuit block during transitions by the circuit between active mode and sleep mode.

Abstract: A method of forming a memory cell includes coupling a first transistor between a supply rail of a memory cell and a node operable to accept a supply voltage. The method further includes coupling a second transistor between a ground rail of the cell and a node operable to accept a ground. In one embodiment, the method includes forming the cell to accept selectively applied external voltages, wherein the external voltages are selected to minimize leakage current in the cell. In another embodiment, the method includes forming at least one of the first and the second transistors to have a channel width and/or a threshold voltage selected to minimize a total leakage current in the cell.

Abstract: In one embodiment, a method for power mode transition in a multi-threshold complementary metal oxide semiconductor (MTCMOS) circuit includes clustering logic cells in the circuit to a number of logic clusters and optimizing wake-up times of the logic clusters to reduce a total turn-on time of the circuit while keeping below a predetermined threshold a sum of currents flowing from the circuit to ground, a sum of currents flowing from a supply voltage to the circuit, or both during a transition by the circuit from sleep mode to active mode.

Abstract: In one embodiment, a circuit includes a first circuit block connected to ground via a first sleep transistor, a virtual ground node between the first circuit block and the first sleep transistor, a second circuit block connected to a supply via a second sleep transistor, and a virtual supply node between the second circuit block and the second sleep transistor. The circuit also includes a transmission gate (TG) or a pass transistor connecting the virtual ground node to the virtual supply node to enable charge recycling between the first circuit block and the second circuit block during transitions by the circuit between active mode and sleep mode.

Abstract: In one embodiment, a memory block includes one or more bit lines that each include two or more cells. Each cell in each bit line has a distance from a sense amplifier coupled to the bit line, and each of one or more of the cells in each of one or more of the bit lines has a delay particularly set according to the distance of the cell from the sense amplifier coupled to the bit line.

Abstract: A method for reducing transitions on a bus is provided that includes receiving an input trace and constructing a Markov source correlating to the input trace. The method also includes identifying an encoding technique, which can either minimize or maximize an objective function associated with the input trace.

Abstract: An embodiment of the present invention is directed to a method for determining a pixel transformation function that maximizes backlight dimming while maintaining a pre-specified distortion level. The method includes determining a minimum dynamic range of pixel values in a transformed image based on an original image and the pre-specified distortion level and determining the pixel transformation function. The pixel transformation function takes a histogram of the original image to a uniform distribution histogram having the minimum dynamic range.