Abstract: Apparatus and methods are disclosed for performing matrix operations, including operations suited to neural network and other machine learning accelerators and applications, using dual exponent formats. Disclosed matrix formats include single exponent bounding box floating-point (SE-BBFP) and dual exponent bounding box floating-point (DE-BBFP) formats. Shared exponents for each element are determined for each element based on whether the element is used as a row of matrix tile or a column of a matrix file, for example, for a dot product operation. Computing systems suitable for employing such neural networks include computers having general-purpose processors, neural network accelerators, or reconfigure both logic devices, such as Field programmable gate arrays (FPGA). Certain techniques disclosed herein can provide improved system performance while reducing memory and network bandwidth used.

Abstract: An integrated circuit includes a pipeline of compare logic stages. The pipeline, at successive pipeline stages, determines whether each of a set of input symbols meets a corresponding programmable criteria. The compare logic stages each compare the set of input symbols to a respective programmable value. The compare logic stages also each provide, to a respective successive compare logic stage, a corresponding plurality of indicators of whether respective ones of the set of input symbols met the corresponding programmable criteria for that compare logic stage. The corresponding programmable criteria are configurable to be based at least in part on the corresponding plurality of indicators from a respective previous compare logic stage.

Abstract: Each processor of the SIMD array performs the computations for a respective neuron of a neural network. As part of this computation, each processor of the SIMD array multiplies an input to a weight and accumulates the result for its assigned neuron each (MAC) instruction cycle. A table in a first memory is used to store which input is fed to each processor of the SIMD array. A crossbar is used to route a specific input to each processor each MAC cycle. A second memory is used to provide the appropriate weight to each processor that corresponds the input being routed to that processor.

Abstract: Each processor of the SIMD array performs the computations for a respective neuron of a neural network. As part of this computation, each processor of the SIMD array multiplies an input to a weight and accumulates the result for its assigned neuron each (MAC) instruction cycle. A table in a first memory is used to store which input is fed to each processor of the SIMD array. A crossbar is used to route a specific input to each processor each MAC cycle. A second memory is used to provide the appropriate weight to each processor that corresponds the input being routed to that processor.

Abstract: An integrated circuit includes a pipeline of compare logic stages. The pipeline, at successive pipeline stages, determines whether each of a set of input symbols meets a corresponding programmable criteria. The compare logic stages each compare the set of input symbols to a respective programmable value. The compare logic stages also each provide, to a respective successive compare logic stage, a corresponding plurality of indicators of whether respective ones of the set of input symbols met the corresponding programmable criteria for that compare logic stage. The corresponding programmable criteria are configurable to be based at least in part on the corresponding plurality of indicators from a respective previous compare logic stage.

Abstract: An integrated circuit includes a pipeline of compare logic stages. The pipeline, at successive pipeline stages, determines whether each of a set of input symbols meets a corresponding programmable criteria. The compare logic stages each compare the set of input symbols to a respective programmable value. The compare logic stages also each provide, to a respective successive compare logic stage, a corresponding plurality of indicators of whether respective ones of the set of input symbols met the corresponding programmable criteria for that compare logic stage. The corresponding programmable criteria are configurable to be based at least in part on the corresponding plurality of indicators from a respective previous compare logic stage.

Abstract: An integrated circuit includes a pipeline of compare logic stages. The pipeline, at successive pipeline stages, determines whether each of a set of input symbols meets a corresponding programmable criteria. The compare logic stages each compare the set of input symbols to a respective programmable value. The compare logic stages also each provide, to a respective successive compare logic stage, a corresponding plurality of indicators of whether respective ones of the set of input symbols met the corresponding programmable criteria for that compare logic stage. The corresponding programmable criteria are configurable to be based at least in part on the corresponding plurality of indicators from a respective previous compare logic stage.

Abstract: Each processor of the SIMD array performs the computations for a respective neuron of a neural network. As part of this computation, each processor of the SIMD array multiplies an input to a weight and accumulates the result for its assigned neuron each (MAC) instruction cycle. A table in a first memory is used to store which input is fed to each processor of the SIMD array. A crossbar is used to route a specific input to each processor each MAC cycle. A second memory is used to provide the appropriate weight to each processor that corresponds the input being routed to that processor.

Abstract: A method of repeating a non-voice signal such as a dual tone, multiple frequency (DTMF) signal, by inserting a delay sequence of data values into an output data sequence of data values, a portion of the output data sequence following the delay sequence being the same as a corresponding portion of an input sequence of decoded data obtained from a low bit-rate speech decoder. The input sequence has at least one distorted non-voice sequence. The method provides for inserting a substantially undistorted non-voice sequence into the output sequence, the undistorted sequence being at least of substantially the same length as the distorted sequence, a portion of the output sequence following the undistorted sequence being the same as a corresponding portion of the input sequence, and the output sequence being substantially free of the distorted non-voice signal.

Abstract: In an ADPCM (Adaptive Differential PCM) system, in which the signal is commonly coded in C.sub.i, Q.sub.n, and .sigma. parameters, a lower sampling rate which normally causes distortion is made possible by deriving additional parameters A.sub.k, B.sub.k as a function of the error (distortion) between the original signal S.sub.n and the sampled signal Y.sub.n. The A.sub.k, B.sub.k coefficients control a distortion filter at the receiver.