Google Inc.

Intel Corporation

Other

David Patterson

Jason Rupert Redgrave

Albert Meixner

Ofer Shacham

Daniel Finchelstein

William R Mark

Qiuling Zhu

Donald C Stark

Neeti Desai

Joshua L Coates

Patrick E Bozeman

Edward Y Chang

David L Green

Thomas Svensson

Taylor S Gautier

Edward T Chang

F Alan Jones

Georgina L Russell

Michael Gonzalez

Rob Eaton

Daniel Frederic Finchelstein

Aspects of the disclosure are directed to a tree-based network architecture for serving and/or training machine learning models. The architecture includes one or more multi-chip packages having a plurality of compute-memory stacks connected via an input/output (I/O) die. The I/O die includes an aggregator to aggregate computations from the compute-memory stacks. The architecture can further include a plurality of the multi-chip packages connected on a server via a server level aggregator and a plurality of the servers connected on a rack via a rack level aggregator for further aggregation of the computations from the compute-memory stacks. The tree-based network architecture allows for fewer hops, resulting in lower latency and savings in bandwidth when serving and/or training machine learning models.

Tree-based network architecture for accelerating machine learning collective operations

A method is described that includes executing a convolutional neural network layer on an image processor having an array of execution lanes and a two-dimensional shift register. The two-dimensional shift register provides local respective register space for the execution lanes. The executing of the convolutional neural network includes loading a plane of image data of a three-dimensional block of image data into the two-dimensional shift register. The executing of the convolutional neural network also includes performing a two-dimensional convolution of the plane of image data with an array of coefficient values by sequentially: concurrently multiplying within the execution lanes respective pixel and coefficient values to produce an array of partial products; concurrently summing within the execution lanes the partial products with respective accumulations of partial products being kept within the two dimensional register for different stencils within the image data; and, effecting alignment of values for the two-dimensional convolution within the execution lanes by shifting content within the two-dimensional shift register array.

Convolutional neural network on programmable two dimensional image processor

A vehicle control system for controlling braking in a vehicle may include a braking system operably coupled to one or more wheels of the vehicle to provide brake inputs to the one or more wheels responsive to a torque request generated based on pedal position. The system may further include a pedal position sensor operably coupled to a brake pedal to determine the pedal position responsive to actuation of the brake pedal by a driver of the vehicle, a pedal feel simulator operably coupled to the brake pedal to provide tactile feedback to the driver via a pedal force applied to the brake pedal, an accelerometer for determining a rate of velocity reduction of the vehicle during the actuation of the brake pedal, and a feedback augmenter operably coupled to the pedal feel simulator to provide a pedal force offset to increase the pedal force provided to the brake pedal based on the rate of velocity reduction.

System and method for providing a brake fade indication

A method is described that includes, on an image processor having a two dimensional execution lane array and a two dimensional shift register array, repeatedly shifting first content of multiple rows or columns of the two dimensional shift register array and repeatedly executing at least one instruction between shifts that operates on the shifted first content and/or second content that is resident in respective locations of the two dimensional shift register array that the shifted first content has been shifted into.

Block operations for an image processor having a two-dimensional execution lane array and a two-dimensional shift register

An apparatus is described. The apparatus includes an execution lane array coupled to a two dimensional shift register array structure. Locations in the execution lane array are coupled to same locations in the two-dimensional shift register array structure such that different execution lanes have different dedicated registers.

Two dimensional shift array for image processor

A method is described that includes executing a convolutional neural network layer on an image processor having an array of execution lanes and a two-dimensional shift register. The executing of the convolutional neural network includes loading a plane of image data of a three-dimensional block of image data into the two-dimensional shift register. The executing of the convolutional neural network also includes performing a two-dimensional convolution of the plane of image data with an array of coefficient values by sequentially: concurrently multiplying within the execution lanes respective pixel and coefficient values to produce an array of partial products; concurrently summing within the execution lanes the partial products with respective accumulations of partial products being kept within the two dimensional register for different stencils within the image data; and, effecting alignment of values for the two-dimensional convolution within the execution lanes by shifting content within the two-dimensional shift register array.

An apparatus is described. The apparatus includes an image processing unit. The image processing unit includes a plurality of stencil processor circuits each comprising an array of execution unit lanes coupled to a two-dimensional shift register array structure to simultaneously process multiple overlapping stencils through execution of program code. The image processing unit includes a plurality of sheet generators respectively coupled between the plurality of stencil processors and the network. The sheet generators are to parse input line groups of image data into input sheets of image data for processing by the stencil processors, and, to form output line groups of image data from output sheets of image data received from the stencil processors. The image processing unit includes a plurality of line buffer units coupled to the network to pass line groups in a direction from producing stencil processors to consuming stencil processors to implement an overall program flow.

Architecture for high performance, power efficient, programmable image processing

A method is described that includes instantiating, within an application software development environment, a virtual processor having an instruction set architecture and memory model that contemplate first and second regions of reserved memory. The first reserved region is to keep data of an input image array. The second reserved region is to keep data of an output image array. The method also includes simulating execution of a memory load instruction of the instruction set architecture by automatically targeting the first reserved region and identifying desired input data with first and second coordinates relative to the virtual processor's position within an orthogonal coordinate system and expressed in the instruction format of the memory load instruction.

Virtual image processor instruction set architecture (ISA) and memory model and exemplary target hardware having a two-dimensional shift array structure

Growing community of inventors

Kensington, CA, United States of America

David Patterson