Abstract: A device can be used to implement a neural network in hardware. The device can include a processor, a memory, and a neural network accelerator. The neural network accelerator can be configured to implement, in hardware, a neural network by using a residue number system (RNS). At least one function of the neural network can have a corresponding approximation in the RNS system, and the at least one function can be provided by implementing the corresponding approximation in hardware.

Abstract: A method can be used to reduce the memory storage and energy used by deep neural networks. The method can include determining the weights associated with the deep neural network. An input feature map can be received and used with the weights to generate approximated weights. Using the approximated weights and the input feature map a convolution inference can be performed.

Abstract: Digital logic voltage regulators and related methods generate a regulated voltage via controlled switching of a power transistor. A digital logic voltage regulator includes a voltage level comparator, a power transistor, and a charge accumulator. The voltage level comparator generates a digital control signal that alternates between a first voltage level and a second voltage level in response to changes in relative voltage level between the regulated output voltage and the target voltage. The digital control signal causes the power transistor to switch from off to on in response to a reduction of the regulated output voltage relative to the target voltage and causes the power transistor to switch from on to off in response to an increase of the regulated output voltage relative to the target voltage. The charge accumulator decreases variation in the regulated output voltage that would occur without the charge accumulator.

Abstract: In an embodiment, a dot-product unit to perform single-precision floating-point product and addition operations is disclosed that includes a first multiplier tree unit adapted to multiply first and second significand operands to produce a first set of two partial products. The dot-product unit further includes a second multiplier tree unit adapted to multiply third and fourth significand operands to produce a second set of two partial products, a shared exponent compare unit adapted to compare exponents of the first, second, third and fourth operands to produce an alignment shift value, and an alignment unit adapted to shift the second set of two partial products based on the alignment shift value. The dot-product unit also includes an adder unit adapted to add or subtract the first set of two partial products and the second shifted set of two partial products to produce a dot-product value that is a single-precision floating-point value.

Abstract: In a particular embodiment, a method is disclosed that includes receiving first and second operands at a floating-point fused add-subtract circuit. The method further includes simultaneously performing add and subtract operations on the first and second operands via the floating-point fused add-subtract circuit to produce a sum result output and a difference result output. The floating-point fused add-subtract circuit includes sign logic, exponent adjustment logic, and shift logic that are shared by an add/round and post-normalize circuit and a subtract/round and post-normalize circuit to produce the sum and difference result outputs.