Chevron U.s.a. Inc.

Nvidia Corporation

Thor Mikal Johnsen

Oliver Pell

Tamas Nemeth

Wei Liu

Kevin John Davies

Alexander Loddoch

Sharan Chetlur

Lei Zhuo

Peeter Akerberg

Oliver Charlesworth

Simon Layton

Natalia Gimelshein

Speed of training a neural network is improved by updating the weights of the neural network in parallel. In at least one embodiment, after back propagation, gradients are distributed to a plurality of processors, each of which calculate a portion of the updated weights of the neural network.

Distributed weight update for backpropagation of a neural network

A method is described for seismic modeling implemented on a computer system including multiple computer nodes, which includes receiving, at a computer processor, a 4D speed of sound in water library and a marine seismic survey geometry; creating, via the computer processor, a full survey 3D Earth model for each seismic shot time based on the 4D speed of sound in water library to generate a set of dynamic 3D Earth models; simulating 4D seismic shots for the set of dynamic 3D Earth models by simulating multiple shots on each node to generate a set of 4D seismic shot gathers; and outputting the 4D seismic shot gathers.

System and method for marine seismic modeling

Embodiments of a method for implementing a finite difference time domain modeling with multiple APCs are disclosed herein. The disclosed methods and systems overcome the memory capacity limitation of APCs by having each APC perform multiple timesteps on a small piece of the computational domain or data volume in a APC queued manner. The cost of transferring data between host and compute accelerator can then be amortized over multiple timesteps, greatly reducing the amount of PCI bandwidth required to sustain high propagation speeds. The APC queued nature of the algorithm achieves linear scaling of PCI throughput with increasing number of APCs, allowing the algorithm to scale up to many dozens of APCs in some embodiments.

System and method of implementing finite difference time domain models with multiple accelerated processing components (APCS)

Improved finite-difference staggered grid wave propagation systems and methods. One illustrative computer-based wave field simulation method includes: applying at least one signal to a grid of model cells forming a model space, each model cell having stress values associated with stress nodes and velocity values associated with velocity nodes staggered from the stress nodes; and propagating the at least one signal as a wave into the model space by alternately updating the stress values and the velocity values to obtain a time-dependent wave field associated with the at least one signal. The stress value updating includes, for each model cell: determining spatial derivatives of the velocity values for the model cell; interpolating the spatial derivatives to multiple stress nodes within the model cell; and, for each stress node within the model cell, combining the spatial derivatives associated with that stress node to update at least one stress value associated with that stress node.

Re-ordered interpolation and convolution for faster staggered-grid processing

A finite difference wavefield modeling framework decouples the tasks of physical modeling and hardware-software optimization through the use of a platform-agnostic intermediate representation in the form of a dataflow graph. In at least some embodiments a wavefield simulation method includes specifying a kernel of operations to be applied at each point in a space and representing the kernel as a platform-agnostic dataflow graph. For each of multiple implementation platforms, the method further includes: deriving from the platform-agnostic graph a customized graph for the implementation platform; translating the customized graph into configuration information for the implementation platform; supplying the configuration information to the implementation platform to obtain a wavefield simulator; and employing the wavefield simulator to repeatedly apply said kernel of operations to each point in the space with specified input or output signals.

Growing community of inventors

Houston, TX, United States of America

Thor Mikal Johnsen