Nvidia Corporation

Alex Fit-Florea

Boris Ginsburg

Amir Gholaminejad

Slawomir Kierat

Pooya Davoodi

Hao Fang Wu

Ujval J Kapasi

Yilin Zhang

Dilip Sequeira

Michael Houston

Hao Wu

Marcin Andrychowicz

Szymon Migacz

Zacky Zhou

Maxim Milakov

Sergei Nikolaev

Ahmad Kiswani

A method, computer readable medium, and system are disclosed for rounding floating point values. Dynamic directional rounding is a rounding technique for floating point operations. A floating point operation (addition, subtraction, multiplication, etc.) is performed on an operand to compute a floating point result. A sign (positive or negative) of the operand is identified. In one embodiment, the sign determines a direction in which the floating point result is rounded (towards negative or positive infinity). When used for updating parameters of a neural network during backpropagation, dynamic directional rounding ensures that rounding is performed in the direction of the gradient.

Dynamic directional rounding

Aspects of the present invention are directed to computer-implemented techniques for improving the training of artificial neural networks using a reduced precision (e.g., float16) data format. Embodiments of the present invention rescale tensor values prior to performing matrix operations (such as matrix multiplication or matrix addition) to prevent overflow and underflow. To preserve accuracy throughout the performance of the matrix operations, the scale factors are defined using a novel data format to represent tensors, wherein a matrix is represented by the tuple X, where X=(a, v[.]), wherein a is a float scale factor and v[.] are scaled values stored in the float16 format. The value of any element X[i] according to this data format would be equal to a*v[i].

Tensor processing using low precision format

Aspects of the present invention are directed to computer-implemented techniques for performing data compression and conversion between data formats of varying degrees of precision, and more particularly for improving the inferencing (application) of artificial neural networks using a reduced precision (e.g., INT8) data format. Embodiments of the present invention generate candidate conversions of data output, then employ a relative measure of quality to identify the candidate conversion with the greatest accuracy (i.e., least divergence from the original higher precision values). The representation can be then be used during inference to perform computations on the resulting output data.

Automated methods for conversions to a lower precision data format

A floating-point unit and a method of identifying exception cases in a floating-point unit. In one embodiment, the floating-point unit includes: (1) a floating-point computation circuit having a normal path and an exception path and operable to execute an operation on an operand and (2) a decision circuit associated with the normal path and the exception path and configured to employ a flush-to-zero mode of the floating-point unit to determine which one of the normal path and the exception path is appropriate for carrying out the operation on the operand.

Circuit and method for identifying exception cases in a floating-point unit and graphics processing unit employing the same

Growing community of inventors

Belmont, CA, United States of America

Alex Fit-Florea