Abstract: In training a deep neural network using reduced precision, gradient computation operates on larger values without affecting the rest of the training procedure. One technique trains the deep neural network to develop loss, scales the loss, computes gradients at a reduced precision, and reduces the magnitude of the computed gradients to compensate for scaling of the loss. In one example non-limiting arrangement, the training forward pass scales a loss value by some factor S and the weight update reduces the weight gradient contribution by 1/S. Several techniques can be used for selecting scaling factor S and adjusting the weight update.

Abstract: In training a deep neural network using reduced precision, gradient computation operates on larger values without affecting the rest of the training procedure. One technique trains the deep neural network to develop loss, scales the loss, computes gradients at a reduced precision, and reduces the magnitude of the computed gradients to compensate for scaling of the loss. In one example non-limiting arrangement, the training forward pass scales a loss value by some factor S and the weight update reduces the weight gradient contribution by 1/S. Several techniques can be used for selecting scaling factor S and adjusting the weight update.

Abstract: Apparatuses, systems, and techniques to transform data sets, such as matrices representing layers of neural networks, to increase sparsity and/or other characteristics of said data sets to improve performance in computations, such as neural network computations. In at least one embodiment, one or more subsets of data in one or more sets of data are rearranged as part of a process to increase sparsity in said one or more sets of data to satisfy one or more one or more structural sparsity constraints.

Abstract: A method, computer readable medium, and system are disclosed for rounding numerical values. A set of bits from an input value is identified as a rounding value. A second set of bits representing a second value is extracted from the input value and added with the rounding value to produce a sum. The sum is truncated to produce the rounded output value. Thus, the present invention provides a stochastic rounding technique that rounds up an input value as a function of a second value and a rounding value, both of which were obtained from the input value. When the second value and rounding value are obtained from consistent bit locations of the input value, the resulting output value is deterministic. Stochastic rounding, which is deterministic, is advantageously applicable in deep learning applications.

Abstract: A method, computer readable medium, and system are disclosed for rounding numerical values. A set of bits from an input value is identified as a rounding value. A second set of bits representing a second value is extracted from the input value and added with the rounding value to produce a sum. The sum is truncated to produce the rounded output value. Thus, the present invention provides a stochastic rounding technique that rounds up an input value as a function of a second value and a rounding value, both of which were obtained from the input value. When the second value and rounding value are obtained from consistent bit locations of the input value, the resulting output value is deterministic. Stochastic rounding, which is deterministic, is advantageously applicable in deep learning applications.

Abstract: A compiler and a method of compiling code that reduces memory bandwidth when processing code on a computer are provided herein. In one embodiment, the method includes: (1) automatically identifying a sequence of operations for fusing, wherein the sequence of operations correspond to instructions from a source code, (2) determining subdivisions of a final output of the sequence of operations, (3) determining input data and intermediate operations needed to obtain a final subdivision output for each of the subdivisions and (4) automatically generating code to fuse the sequence of operations employing the subdivisions, wherein the automatically identifying and the automatically generating are performed by a processor.

Abstract: In training a deep neural network using reduced precision, gradient computation operates on larger values without affecting the rest of the training procedure. One technique trains the deep neural network to develop loss, scales the loss, computes gradients at a reduced precision, and reduces the magnitude of the computed gradients to compensate for scaling of the loss. In one example non-limiting arrangement, the training forward pass scales a loss value by some factor S and the weight update reduces the weight gradient contribution by 1/S. Several techniques can be used for selecting scaling factor S and adjusting the weight update.

Abstract: A compiler and a method of compiling code that reduces memory bandwidth when processing code on a computer are provided herein. In one embodiment, the method includes: (1) automatically identifying a sequence of operations for fusing, wherein the sequence of operations correspond to instructions from a source code, (2) determining subdivisions of a final output of the sequence of operations, (3) determining input data and intermediate operations needed to obtain a final subdivision output for each of the subdivisions and (4) automatically generating code to fuse the sequence of operations employing the subdivisions, wherein the automatically identifying and the automatically generating are performed by a processor.