Physics simulation on machine-learning accelerated hardware platforms

Patent metadata

Field	Value
Publication number	US-11550971-B1
Application number	US-201916251708-A
Country	US
Kind code	B1
Filing date	Jan 18, 2019
Priority date	Jan 18, 2019
Publication date	Jan 10, 2023
Grant date	Jan 10, 2023

Abstract

At least one machine-accessible storage medium that provides instructions that, when executed by a machine, will cause the machine to perform operations. The operations comprise configuring a simulated environment to be representative of a physical device based, at least in part, on an initial description of the physical device that described structural parameters of the physical device. The operations further comprise performing a physics simulation with an artificial intelligence (“AI”) accelerator. The AI accelerator includes a matrix multiply unit for computing convolution operations via a plurality of multiply-accumulate units. The operations further comprise computing a field response in response of the physical device in response to an excitation source within the simulated environment when performing the physics simulation. The field response is computed, at least in part, with the convolution operations to perform spatial differencing.

First claim

What is claimed is:

1. At least one non-transitory machine-accessible storage medium that provides instructions that, when executed by a machine, will cause the machine to perform operations comprising:
configuring a simulated environment to be representative of a physical device based, at least in part, on an initial description of the physical device that describes structural parameters of the physical device;
performing a physics simulation using an artificial intelligence (“AI”) accelerator, wherein the AI accelerator includes a matrix multiply unit for computing convolution operations via a plurality of multiply-accumulate (“MAC”) units; and
computing a field response of the physical device in response to an excitation source within the simulated environment when performing the physics simulation, wherein the field response is computed, at least in part, with the convolution operations to perform spatial differencing.

2. The at least one non-transitory machine-accessible storage medium of claim 1 , which provides additional instructions that, when executed by the machine, will cause the machine to perform further operations comprising:
computing a loss metric based, at least in part, on a comparison between a performance metric of the physical device determined from the field response and a target performance metric;
performing a second physics simulation of the physical device using the AI accelerator to compute a loss response of the physical device to an adjoint source for determining an influence of changes in the structural parameters on the loss metric, and wherein the loss response is determined, at least in part, with the convolution operations to perform the spatial differencing; and
generating a revised description of the physical device by updating the structural parameters to reduce the loss metric.

3. The at least one non-transitory machine-accessible storage medium of claim 2 , which provides additional instructions that, when executed by the machine, will cause the machine to perform further operations comprising:
iteratively performing cycles with the AI accelerator, each of the cycles including successively performing the physics simulation, performing the second physics simulation, and generating the revised description of the physical device, wherein the cycles iteratively reduce the loss metric until the loss metric substantially converges such that a difference between the performance metric and the target performance metric is within a threshold range.

4. The at least one non-transitory machine-accessible storage medium of claim 1 , wherein the physical device is described by a plurality of voxels within the simulated environment, wherein the spatial differencing is performed, in part, by defining one or more kernels of the convolution operations that results in calculating differences between field values of neighboring voxels included in the plurality of voxels to approximate one or more spatial derivatives for computing the field response.

5. The at least one non-transitory machine-accessible storage medium of claim 4 , wherein the differences between the field values of the neighboring voxels are taken along at least one of a first dimension, a second dimension, or a third dimension, wherein spatial dimensionality of the physical device is one-dimensional, two-dimensional, or three-dimensional, and wherein the spatial dimensionality is defined by at least one of the first dimension, the second dimension, or the third dimension.

6. The at least one non-transitory machine-accessible storage medium of claim 5 , which provides additional instructions that, when executed by the machine, will cause the machine to perform further operations comprising:
splitting at least one of the first dimension, the second dimension, or the third dimension into multiple virtual dimensions to describe the physical device with the spatial dimensionality redefined for performing the convolution operations with the AI accelerator; and
mapping at least one of the multiple virtual dimensions to a batch dimension or feature dimension of the convolution operations.

7. The at least one non-transitory machine-accessible storage medium of claim 6 , wherein a first virtual dimension, included in the multiple virtual dimensions, associated with the second dimension is mapped to the batch dimension of the convolution operations, and wherein a second virtual dimension, included in the multiple virtual dimensions, associated with third dimension is mapped to the feature dimension of the convolution operations.

8. The at least one non-transitory machine-accessible storage medium of claim 6 , wherein the splitting into the multiple virtual dimensions is based, at least in part, on an arrangement of the plurality of MAC units of the matrix multiply unit.

9. The at least one non-transitory machine-accessible storage medium of claim 6 , wherein a size of at least one of the multiple virtual dimensions is matched to a memory subsystem of the AI accelerator.

10. The at least one non-transitory machine-accessible storage medium of claim 4 , wherein at least one of the one or more kernels is defined such that the spatial differencing is third or higher order accurate to approximate the one or more spatial derivatives.

11. The at least one non-transitory machine-accessible storage medium of claim 1 , which provides additional instructions that, when executed by the machine, will cause the machine to perform further operations comprising:
dividing the simulated environment including the physical device into a plurality of subdomains, wherein the matrix multiply unit (MXU) is a first MXU included in a plurality of MXUs that are interconnected such that the AI accelerator is a distributed system, and wherein each of the plurality of subdomains are respectively assigned to a corresponding one of the plurality of MXUs.

12. The at least one non-transitory machine-accessible storage medium of claim 11 , wherein the physical device is described by a plurality of voxels within the simulated environment, wherein the plurality of voxels is associated with field values that collectively represent the field response of the physical device, wherein each of the plurality of subdomains includes a respective portion of the plurality of voxels, and wherein the at least one non-transitory machine-accessible storage medium provides additional instructions that, when executed by the machine, will cause the machine to perform further operations comprising:
communicating the field values of a first portion of the plurality of voxels that interface with adjacent subdomains to corresponding MXUs included in the plurality of MXUs associated with the adjacent subdomains for computing the field response of the physical device.

13. The at least one non-transitory machine-accessible storage medium of claim 12 , wherein the plurality of subdomains overlap one another such that a second portion of the plurality of voxels is included in two or more subdomains included in the plurality of subdomains.

14. The at least one non-transitory machine-accessible storage medium of claim 12 , wherein a rate of the communicating is based, at least in part, on a degree of overlap between the plurality of subdomains.

15. The at least one non-transitory machine-accessible storage medium of claim 12 , wherein the spatial differencing is performed, in part, by defining one or more kernels of the convolution operations that results in calculating differences between the field values of the plurality of voxels describing the physical device within the simulated environment to approxima

Assignees

• X Dev Llc

Inventors

• Lu Jesse
• Adolf Brian
• Schubert Martin

Classifications

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06F30/20 Primary

  Design optimisation, verification or simulation (optimisation, verification or simulation of circuit designs G06F30/30) · CPC title

• G06F7/5443

  Sum of products (for applications thereof, see the relevant places, e.g. G06F17/10, H03H17/00) · CPC title

• G06F17/16

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title

• G06F30/27

  using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model · CPC title