International Business Machines Corporation

Jun Sawada

Andrew S Cassidy

Dharmendra S Modha

John Vernon Arthur

Paul A Merolla

Bryan L Jackson

Pallab Datta

Rodrigo Alvarez-Icaza Rivera

Myron Dale Flickner

Steven Kyle Esser

Rathinakumar Appuswamy

Brian Taba

Hartmut Penner

Jennifer Klamo

Filipp A Akopyan

Subramanian Srikanteswara Iyer

Hung Cai Ngo

John Edward Barth, Jr

Benjamin G Shaw

Rodrigo Alvarez-Icaza

Wendy Ann Belluomini

Tapan K Nayak

Carlos O Otero

Viresh Paruthi

Yutaka Nakamura

Peter Anthony Sandon

Chen Guo

Michael Vincent DeBole

Rodrigo Alvarez Icaza Rivera

Michael V DeBole

Simon J Hollis

Keith R Walker

Jason Raymond Baumgartner

Yen-Fu Chen

Robert K Montoye

Hari Mony

Kevin John Nowka

Bruce Martin Fleischer

John Hans Handy-Bosma

Mei Yang Selvage

Andrew Kenneth Martin

Michael L Case

Chandler Todd McDowell

H Peter Hofstee

John E Barth

Steve Kyle Esser

Jeffrey L McKinstry

Filipp Akopyan

Steve Esser

Jeffrey Mckinstry

Andrew Stephen Cassidy

Brian Seisho Taba

Carlos Ortega Otero

Networks of distributed neural cores are provided with hierarchical parallelism. In various embodiments, a plurality of neural cores is provided. Each of the plurality of neural cores comprises a plurality of vector compute units configured to operate in parallel. Each of the plurality of neural cores is configured to compute in parallel output activations by applying its plurality of vector compute units to input activations. Each of the plurality of neural cores is assigned a subset of output activations of a layer of a neural network for computation. Upon receipt of a subset of input activations of the layer of the neural network, each of the plurality of neural cores computes a partial sum for each of its assigned output activations, and computes its assigned output activations from at least the computed partial sums.

Hierarchical parallelism in a network of distributed neural network cores

A neural inference chip includes a global weight memory; a neural core; and a network connecting the global weight memory to the at least one neural core. The neural core comprises a local weight memory. The local weight memory comprises a plurality of memory banks. Each of the plurality of memory banks is uniquely addressable by at least one index. The neural inference chip is adapted to store in the global weight memory a compressed weight block comprising at least one compressed weight matrix. The neural inference chip is adapted to transmit the compressed weight block from the global weight memory to the core via the network. The core is adapted to decode the at least one compressed weight matrix into a decoded weight matrix and store the decoded weight matrix in its local weight memory. The at core is adapted to apply the decoded weight matrix to a plurality of input activations to produce a plurality of output activations.

Compressed weight distribution in networks of neural processors

Conflict-free, stall-free, broadcast networks on neural inference chips are provided. In various embodiments, a neural inference chip comprises a plurality of network nodes and a network on chip interconnecting the plurality of network nodes. The network comprises at least one pair of directional paths. The paths of each pair have opposite directions and a common end. The network is configured to accept data at any of the plurality of nodes. The network is configured to propagate data along a first of the pair of directional paths from a source node to the common end of the pair of directional paths and along a second of the pair of directional paths from the common end of the pair of directional paths to one or more destination node.

Conflict-free, stall-free, broadcast network on chip

Multi-agent instruction execution engines for neural inference processing are provided. In various embodiments, a neural core is provided. The neural core includes an instruction memory. The instruction memory comprises a plurality of instruction streams, each instruction stream associated with one of a plurality of agents. The instruction memory further comprises a plurality of shared functional units. The neural core is adapted to concurrently execute the plurality of instruction streams on the plurality of associated agents. The execution includes maintaining a separate program counter for each of the plurality of agents, determining a plurality of operations from the instructions of each instruction stream, and directing the operations to the shared functional units. The instructions of each instruction stream are statically scheduled prior to runtime to ensure their execution is conflict free.

Multi-agent instruction execution engine for neural inference processing

A neural inference chip is provided, including at least one neural inference core. The at least one neural inference core is adapted to apply a plurality of synaptic weights to a plurality of input activations to produce a plurality of intermediate outputs. The at least one neural inference core comprises a plurality of activation units configured to receive the plurality of intermediate outputs and produce a plurality of activations. Each of the plurality of activation units is configured to apply a configurable activation function to its input. The configurable activation function has at least a re-ranging term and a scaling term, the re-ranging term determining the range of the activations and the scaling term determining the scale of the activations. Each of the plurality of activations units is configured to obtain the re-ranging term and the scaling term from one or more look up tables.

Efficient method for VLSI implementation of useful neural network activation functions

Mapping of neural network layers to physical neural cores is provided. In various embodiments, a neural network description describing a plurality of neural network layers is read. Each of the plurality of neural network layers has an associated weight tensor, input tensor, and output tensor. A plurality of precedence relationships among the plurality of neural network layers is determined. The weight tensor, input tensor, and output tensor of each of the plurality of neural network layers are mapped onto an array of neural cores.

Scheduler for mapping neural networks onto an array of neural cores in an inference processing unit

Systems for neural network computation are provided. A neural network processor comprises a plurality of neural cores. The neural network processor has one or more processor precisions per activation. The processor is configured to accept data having a processor feature dimension. A transformation circuit is coupled to the neural network processor, and is adapted to: receive an input data tensor having an input precision per channel at one or more features; transform the input data tensor from the input precision to the processor precision; divide the input data into a plurality of blocks, each block conforming to one of the processor feature dimensions; provide each of the plurality of blocks to one of the plurality of neural cores. The neural network processor is adapted to compute, by the plurality of neural cores, output of one or more neural network layers.

Data representation for dynamic precision in neural network cores

Networks for distributing parameters and data to neural network compute cores. In various embodiments, a neural inference chip comprises a plurality of neural cores and at least one network interconnecting the plurality of neural cores. Each of the plurality of neural cores is adapted to apply a plurality of synaptic weights to a plurality of input activations to produce a plurality of output activations. The at least one network is adapted to simultaneously deliver synaptic weights and/or input activations to the plurality of neural cores.

Networks for distributing parameters and data to neural network compute cores

Defect resistant designs for location-sensitive neural network processor arrays are provided. In various embodiments, plurality of neural network processor cores are arrayed in a grid. The grid has a plurality of rows and a plurality of columns. A network interconnects at least those of the plurality of neural network processor cores that are adjacent within the grid. The network is adapted to bypass a defective core of the plurality of neural network processor cores by providing a connection between two non-adjacent rows or columns of the grid, and transparently routing messages between the two non-adjacent rows or columns, past the defective core.

Defect resistant designs for location-sensitive neural network processor arrays

Hardware neural network processors, are provided. A neural core includes a weight memory, an activation memory, a vector-matrix multiplier, and a vector processor. The vector-matrix multiplier is adapted to receive a weight matrix from the weight memory, receive an activation vector from the activation memory, and compute a vector-matrix multiplication of the weight matrix and the activation vector. The vector processor is adapted to receive one or more input vector from one or more vector source and perform one or more vector functions on the one or more input vector to yield an output vector. In some embodiments a programmable controller is adapted to configure and operate the neural core.

Runtime reconfigurable neural network processor core

Neural network processing hardware using parallel computational architectures with reconfigurable core-level and vector-level parallelism is provided. In various embodiments, a neural network model memory is adapted to store a neural network model comprising a plurality of layers. Each layer has at least one dimension and comprises a plurality of synaptic weights. A plurality of neural cores is provided. Each neural core includes a computation unit and an activation memory. The computation unit is adapted to apply a plurality of synaptic weights to a plurality of input activations to produce a plurality of output activations. The computation unit has a plurality of vector units. The activation memory is adapted to store the input activations and the output activations. The system is adapted to partition the plurality of cores into a plurality of partitions based on dimensions of the layer and the vector units.

Parallel computational architecture with reconfigurable core-level and vector-level parallelism

Instruction distribution in an array of neural network cores is provided. In various embodiments, a neural inference chip is initialized with core microcode. The chip comprises a plurality of neural cores. The core microcode is executable by the neural cores to execute a tensor operation of a neural network. The core microcode is distributed to the plurality of neural cores via an on-chip network. The core microcode is executed synchronously by the plurality of neural cores to compute a neural network layer.

Instruction distribution in an array of neural network cores

Neural inference chips are provided. A neural core of the neural inference chip comprises a vector-matrix multiplier; a vector processor; and an activation unit operatively coupled to the vector processor. The vector-matrix multiplier, vector processor, and/or activation unit is adapted to operate at variable precision.

Flexible precision neural inference processing unit

Neural inference chips for computing neural activations are provided. In various embodiments, a neural inference chip comprises at least one neural core, a memory array, an instruction buffer, and an instruction memory. The instruction buffer has a position corresponding to each of a plurality of elements of the memory array. The instruction memory provides at least one instruction to the instruction buffer. The instruction buffer advances the at least one instruction between positions in the instruction buffer. The instruction buffer provides the at least one instruction to at least one of the plurality of elements of the memory array from its associated position in the instruction buffer when the memory of the at least one of the plurality of elements contains data associated with the at least one instruction. Each element of the memory array provides a data block from its memory to its horizontal buffer in response to the arrival of an associated instruction from the instruction buffer. The horizontal buffer of each element of the memory array provides a data block to the horizontal buffer of another of the elements of the memory array or to the at least one neural core.

Neural network weight distribution from a grid of memory elements

Growing community of inventors

Austin, TX, United States of America

Jun Sawada