Abstract: Imaging circuitry may include circuits for implementing feature extraction in the analog domain. The imaging circuitry may include pixels configured to generate pixel values. The pixel values may then be weighted using variable charge integration times, variable resistors in the readout path, and/or variable switch on times in the readout path. The weighted pixels values may be binned and combined to obtain an output neuron voltage for at least one layer in a neural network. Performing feature extraction in the analog domain for each layer of results in the neural network saves power and area by avoiding the need to move data around to conventional digital memories.

Abstract: Imaging circuitry may include circuits for implementing feature extraction in the analog domain. The imaging circuitry may include pixels configured to generate pixel values. The pixel values may then be weighted using variable charge integration times, variable resistors in the readout path, and/or variable switch on times in the readout path. The weighted pixels values may be binned and combined to obtain an output neuron voltage for at least one layer in a neural network. Performing feature extraction in the analog domain for each layer of results in the neural network saves power and area by avoiding the need to move data around to conventional digital memories.

Abstract: A micro-code sequence to reduce the rate of change of current required by a processor coming out of a sleep mode when the processor clock is resumed. After stopping the instruction fetch unit, an instruction with a long latency, or execution time, can be initiated by the micro-code before the processor clock is stopped to enter a sleep mode. When the sleep mode is exited by resuming the processor clock, the instruction with the long execution time is completed before restarting the instruction fetch unit. This prevents a portion of the processor circuitry from resuming operation immediately when the clock is resumed, which also delays some of the current demands made by that portion of the circuitry. This creates a more gradual increase in the current required by the processor when exiting a sleep mode.

Abstract: A micro-code sequence to reduce the rate of change of current required by a processor coming out of a sleep mode when the processor clock is resumed. After stopping the instruction fetch unit, an instruction with a long latency, or execution time, can be initiated by the micro-code before the processor clock is stopped to enter a sleep mode. When the sleep mode is exited by resuming the processor clock, the instruction with the long execution time is completed before restarting the instruction fetch unit. This prevents a portion of the processor circuitry from resuming operation immediately when the clock is resumed, which also delays some of the current demands made by that portion of the circuitry. This creates a more gradual increase in the current required by the processor when exiting a sleep mode.

Abstract: The present invention provides an improved cache memory architecture with way prediction. The improved architecture entails placing the address tag array of a cache memory on the central processing unit core (i.e. the microprocessor chip), while the cache data array remains off the microprocessor chip. In addition, a way predictor is provided in conjunction with the improved memory cache architecture to increase the overall performance of the cache memory system.

Abstract: A circuit for providing an OR function on the outputs of at least two MOS logic circuits. The circuit has an output node capable of being in a first or second logic state and being responsive to a first or second path. The first path includes multiple WIRED-OR logic circuits which function as an OR gate on the outputs of MOS logic circuits. The results of the operation cause the architecture output to transition into the first state. The second path is skewed for the second state, such that the transition into the second state occurs fast. Thus, the transition of the output node from the second state to the first state and vice versa is provided by one path, such that the overall ORing function occurs faster.