Cerebras Systems Inc.

Intel Corporation

Mosys, Inc.

Institute for the Development of Emerging Architectures, L.l.c.

Michael J Morrison

Gary R Lauterbach

Sean Lie

Michael Edwin James

Srikanth Arekapudi

Jay B Patel

Michael J Miller

Alan Hersh Karp

Nazar A Zaidi

Jerome C Huck

Glenn T Colon-Bonet

Roger A Golliver

Gautam B Doshi

Peter Markstein

Philip Arnold Ferolito

Bharat N Zaveri

Guillermo Juan Rozas

Elango Ganesan

Sivakumar Makineni

Josef R Call

Ricardo Ramirez

David N Harris

Dipak Sikdar

Deepak J Aatresh

Ghassan Khadder

Man Kit Tang

Stanislav Funiak

Toru M Kuzuhara

Guatam Bhawandas Doshi

Andrew Paul Kelm

Thomas Grutkowski

Vladimir Kibardin

Vijay Anand Reddy Korthikanti

Michael Miller

Toru Kuzuhara

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a compute element and a routing element. Each compute element has memory. Each router enables communication via wavelets with at least nearest neighbors in a 2D mesh. Routing is controlled by respective virtual channel specifiers in each wavelet and routing configuration information in each router. A compute element conditionally selects for task initiation a previously received wavelet specifying a particular one of the virtual channels. The conditional selecting excludes the previously received wavelet for selection until at least block/unblock state maintained for the particular virtual channel is in an unblock state. The compute element executes block/unblock instructions to modify the block/unblock state.

Task synchronization for accelerated deep learning

Techniques in wavelet filtering for advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets in accordance with virtual channel specifiers. Each processing element is enabled to perform local filtering of wavelets received at the processing element, selectively, conditionally, and/or optionally discarding zero or more of the received wavelets, thereby preventing further processing of the discarded wavelets. The wavelet filtering is performed by one or more configurable wavelet filters operable in various modes, such as counter, sparse, and range modes.

Advanced wavelet filtering for accelerated deep learning

Techniques in placement of compute and memory for accelerated deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element to execute programmed instructions using the data and a router to route the wavelets. The routing is in accordance with virtual channel specifiers of the wavelets and controlled by routing configuration information of the router. A software stack determines placement of compute resources and memory resources based on a description of a neural network. The determined placement is used to configure the routers including usage of the respective colors. The determined placement is used to configure the compute elements including the respective programmed instructions each is configured to execute.

Placement of compute and memory for accelerated deep learning

Techniques in dynamic routing for advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element comprises a compute element enabled to execute programmed instructions using the data and a router enabled to route the wavelets via static routing, dynamic routing, or both. The routing is in accordance with a respective virtual channel specifier of each of the wavelets and controlled by routing configuration information of the router. The static techniques enable statically specifiable neuron connections. The dynamic techniques enable information from the wavelets to alter the routing configuration information during neural network processing.

Dynamic routing for accelerated deep learning

Basic wavelet filtering for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency, such as accuracy of learning, accuracy of prediction, speed of learning, performance of learning, and energy efficiency of learning. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Each compute element has processing resources and memory resources. Each router enables communication via wavelets with at least nearest neighbors in a 2D mesh. Stochastic gradient descent, mini-batch gradient descent, and continuous propagation gradient descent are techniques usable to train weights of a neural network modeled by the processing elements. Reverse checkpoint is usable to reduce memory usage during the training.

Accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a compute element and a routing element. Each router enables communication via wavelets with at least nearest neighbors in a 2D mesh. Routing is controlled by virtual channel specifiers in each wavelet and routing configuration information in each router. Execution of an activate instruction or completion of a fabric vector operation activates one of the virtual channels. A virtual channel is selected from a pool comprising previously activated virtual channels and virtual channels associated with previously received wavelets. A task corresponding to the selected virtual channel is activated by executing instructions corresponding to the selected virtual channel.

Task activating for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements performs flow based computations on wavelets of data. Each processing element has a compute element and a routing element. Each compute element has memory. Each router enables communication via wavelets with nearest neighbors in a 2D mesh. A compute element receives a wavelet. If a control specifier of the wavelet is a first value, then instructions are read from the memory of the compute element in accordance with an index specifier of the wavelet. If the control specifier is a second value, then instructions are read from the memory of the compute element in accordance with a virtual channel specifier of the wavelet. Then the compute element initiates execution of the instructions.

Control wavelet for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Instructions executed by the compute element include operand specifiers, some specifying a data structure register storing a data structure descriptor describing an operand as a fabric vector or a memory vector. The data structure descriptor further describes the memory vector as one of a one-dimensional vector, a four-dimensional vector, or a circular buffer vector. Optionally, the data structure descriptor specifies an extended data structure register storing an extended data structure descriptor. The extended data structure descriptor specifies parameters relating to a four-dimensional vector or a circular buffer vector.

Data structure descriptors for deep learning acceleration

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Each compute element has memory. At least a first single neuron is implemented using resources of a plurality of the array of processing elements. At least a portion of a second neuron is implemented using resources of one or more of the plurality of processing elements. In some usage scenarios, the foregoing neuron implementation enables greater performance by enabling a single neuron to use the computational resources of multiple processing elements and/or computational load balancing across the processing elements while maintaining locality of incoming activations for the processing elements.

Neuron smearing for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of compute elements and routers performs flow-based computations on wavelets of data. Some instructions are performed in iterations, such as one iteration per element of a fabric vector or FIFO. When sources for an iteration of an instruction are unavailable, and/or there is insufficient space to store results of the iteration, indicators associated with operands of the instruction are checked to determine whether other work can be performed. In some scenarios, other work cannot be performed and processing stalls. Alternatively, information about the instruction is saved, the other work is performed, and sometime after the sources become available and/or sufficient space to store the results becomes available, the iteration is performed using the saved information.

Microthreading for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, and energy efficiency. An array of processing elements comprising a portion of a neural network accelerator performs flow-based computations on wavelets of data. Each processing element has a respective compute element and a respective routing element. Each compute element has a respective floating-point unit enabled to perform stochastic rounding, thus in some circumstances enabling reducing systematic bias in long dependency chains of floating-point computations. The long dependency chains of floating-point computations are performed, e.g., to train a neural network or to perform inference with respect to a trained neural network.

Floating-point unit stochastic rounding for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of cost, accuracy, performance, and energy efficiency. The deep learning accelerator is implemented at least in part via wafer-scale integration. The wafer comprises a plurality of processor elements, each augmented with redundancy-enabling couplings. The redundancy-enabling couplings enable using redundant ones of the processor elements to replace defective ones of the processor elements. Defect information gathered at wafer test and/or in-situ, such as in a datacenter, is used to determine configuration information for the redundancy-enabling couplings.

Processor element redundancy for accelerated deep learning

Techniques in advanced deep learning provide improvements in one or more of accuracy, performance, energy efficiency, and cost. In a first embodiment, a scaled array of processing elements is implementable with varying dimensions of the processing elements to enable varying price/performance systems. In a second embodiment, an array of clusters communicates via high-speed serial channels. The array and the channels are implemented on a Printed Circuit Board (PCB). Each cluster comprises respective processing and memory elements. Each cluster is implemented via a plurality of 3D-stacked dice, 2.5D-stacked dice, or both in a Ball Grid Array (BGA). A processing portion of the cluster is implemented via one or more Processing Element (PE) dice of the stacked dice. A memory portion of the cluster is implemented via one or more High Bandwidth Memory (HBM) dice of the stacked dice.

Scaled compute fabric for accelerated deep learning

Growing community of inventors

Santa Clara, CA, United States of America

Michael J Morrison