Google Inc.

Andrew Everett Phelps

Norman Paul Jouppi

Gregory Michael Thorson

Thomas James Norrie

Jonathan Alexander Ross

Olivier Temam

Christopher Aaron Clark

Dong Hyuk Woo

Kyle Nesbit

Reginald Clifford Young

Dan Luu

William Lacy

Matthew Leever Hedlund

Gurushankar Rajamani

Sushma Honnavara-Prasad

Vinayak Anand Gokhale

Hsin-Jung Yang

Pareesa Ameneh Golnari

Oliver Temam

Methods, systems, and apparatus including a special purpose hardware chip for training neural networks are described. The special-purpose hardware chip may include a scalar processor configured to control computational operation of the special-purpose hardware chip. The chip may also include a vector processor configured to have a 2-dimensional array of vector processing units which all execute the same instruction in a single instruction, multiple-data manner and communicate with each other through load and store instructions of the vector processor. The chip may additionally include a matrix multiply unit that is coupled to the vector processor configured to multiply at least one two-dimensional matrix with a second one-dimensional vector or two-dimensional matrix in order to obtain a multiplication result.

Special purpose neural network training chip

A vector reduction circuit configured to reduce an input vector of elements comprises a plurality of cells, wherein each of the plurality of cells other than a designated first cell that receives a designated first element of the input vector is configured to receive a particular element of the input vector, receive, from another of the one or more cells, a temporary reduction element, perform a reduction operation using the particular element and the temporary reduction element, and provide, as a new temporary reduction element, a result of performing the reduction operation using the particular element and the temporary reduction element. The vector reduction circuit also comprises an output circuit configured to provide, for output as a reduction of the input vector, a new temporary reduction element corresponding to a result of performing the reduction operation using a last element of the input vector.

Vector reduction processor

A vector processing unit is described, and includes processor units that each include multiple processing resources. The processor units are each configured to perform arithmetic operations associated with vectorized computations. The vector processing unit includes a vector memory in data communication with each of the processor units and their respective processing resources. The vector memory includes memory banks configured to store data used by each of the processor units to perform the arithmetic operations. The processor units and the vector memory are tightly coupled within an area of the vector processing unit such that data communications are exchanged at a high bandwidth based on the placement of respective processor units relative to one another, and based on the placement of the vector memory relative to each processor unit.

Vector processing unit

Methods, systems, and apparatus, including computer-readable media, are described for a hardware circuit configured to implement a neural network. The circuit includes a first memory, respective first and second processor cores, and a shared memory. The first memory provides data for performing computations to generate an output for a neural network layer. Each of the first and second cores include a vector memory for storing vector values derived from the data provided by the first memory. The shared memory is disposed generally intermediate the first memory and at least one core and includes: i) a direct memory access (DMA) data path configured to route data between the shared memory and the respective vector memories of the first and second cores and ii) a load-store data path configured to route data between the shared memory and respective vector registers of the first and second cores.

Shared scratchpad memory with parallel load-store

A circuit comprises an input register configured to receive an input vector of elements, a control register configured to receive a control vector of elements, wherein each element of the control vector corresponds to a respective element of the input vector, and wherein each element specifies a permutation of a corresponding element of the input vector, and a permute execution circuit configured to generate an output vector of elements corresponding to a permutation of the input vector. Generating each element of the output vector comprises accessing, at the input register, a particular element of the input vector, accessing, at the control register, a particular element of the control vector corresponding to the particular element of the input vector, and outputting the particular element of the input vector as an element at a particular position of the output vector that is selected based on the particular element of the control vector.

Permuting in a matrix-vector processor

In a system including vector registers storing right-hand side data and left-hand side data, first and second matrix staging registers, and a systolic array of processing cells for conducting matrix multiplication operations using the right-hand side data and left-hand side data, one or more processors load the right-hand side data from the vector registers to the first matrix staging register based on an instruction indicating whether to transpose the right-hand side data, load the left-hand side data from the vector registers into the second matrix staging register based on another instruction indicating whether to transpose the left-hand side data, load the right-hand side data from the first matrix staging register into the systolic array, and, in a cycle of the matrix multiplication operation, pass one or more columns of the left-hand side data from the second matrix staging register to a column of the systolic array.

Transposition of left-hand-side operand to systolic matrix multiplication

A distributed storage system including memory hosts and at least one curator in communication with the memory hosts. Each memory host has memory, and the curator manages striping of data across the memory hosts. In response to a memory access request by a client in communication with the memory hosts and the curator, the curator provides the client a file descriptor mapping data stripes and data stripe replications of a file on the memory hosts for remote direct memory access of the file on the memory hosts.

Single-sided distributed storage system

Methods, systems, and apparatus for a matrix multiply unit implemented as a systolic array of cells are disclosed. The matrix multiply unit may include cells arranged in columns of the systolic array. Two chains of weight shift registers per column of the systolic array are in the matrix multiply unit. Each weight shift register is connected to only one chain and each cell is connected to only one weight shift register. A weight matrix register per cell is configured to store a weight input received from a weight shift register. A multiply unit is coupled to the weight matrix register and configured to multiply the weight input of the weight matrix register with a vector data input in order to obtain a multiplication result.

Low latency matrix multiply unit

Methods, systems, and apparatus for performing a matrix multiplication using a hardware circuit are described. An example method begins by obtaining an input activation value and a weight input value in a first floating point format. The input activation value and the weight input value are multiplied to generate a product value in a second floating point format that has higher precision than the first floating point format. A partial sum value is obtained in a third floating point format that has a higher precision than the first floating point format. The partial sum value and the product value are combined to generate an updated partial sum value that has the third floating point format.

Performing matrix multiplication in hardware

Methods, systems, and apparatus, including computer-readable media, are described for performing vector reductions using a shared scratchpad memory of a hardware circuit having processor cores that communicate with the shared memory. For each of the processor cores, a respective vector of values is generated based on computations performed at the processor core. The shared memory receives the respective vectors of values from respective resources of the processor cores using a direct memory access (DMA) data path of the shared memory. The shared memory performs an accumulation operation on the respective vectors of values using an operator unit coupled to the shared memory. The operator unit is configured to accumulate values based on arithmetic operations encoded at the operator unit. A result vector is generated based on performing the accumulation operation using the respective vectors of values.

Vector reductions using shared scratchpad memory

Methods, systems, and apparatus for a matrix multiply unit implemented as a systolic array of cells are disclosed. Each cell of the matrix multiply includes: a weight matrix register configured to receive a weight input from either a transposed or a non-transposed weight shift register; a transposed weight shift register configured to receive a weight input from a horizontal direction to be stored in the weight matrix register; a non-transposed weight shift register configured to receive a weight input from a vertical direction to be stored in the weight matrix register; and a multiply unit that is coupled to the weight matrix register and configured to multiply the weight input of the weight matrix register with a vector data input in order to obtain a multiplication result.

Growing community of inventors

Middleton, WI, United States of America

Andrew Everett Phelps