Microsoft Technology Licensing, LLC

Larry Marvin Wall

Amol Ashok Ambardekar

Chad Balling McBride

Kent D Cedola

George Petre

Boris Bobrov

Benjamin Eliot Lundell

Kristof Roomp

Timothy Hume Heil

Aleksandar Tomic

Glen Buhlmann

Nicholas Duncan

Joseph Leon Corkery

Michael Ebstyne

Christopher Leonard Huybregts

Trebor Lee Connell

Hrishikesh Pathak

Damoun Houshangi

Alvin Wen Jie Chia

Chad Balling Mcbride

Neural processing elements are configured with a hardware AND gate configured to perform a logical AND operation between a sign extend signal and a most significant bit ('MSB') of an operand. The state of the sign extend signal can be based upon a type of a layer of a deep neural network (“DNN”) that generate the operand. If the sign extend signal is logical FALSE, no sign extension is performed. If the sign extend signal is logical TRUE, a concatenator concatenates the output of the hardware AND gate and the operand, thereby extending the operand from an N-bit unsigned binary value to an N+1 bit signed binary value. The neural processing element can also include another hardware AND gate and another concatenator for processing another operand similarly. The outputs of the concatenators for both operands are provided to a hardware binary multiplier.

Increased precision neural processing element

A deep neural network ('DNN') module compresses and decompresses neuron-generated activation data to reduce the utilization of memory bus bandwidth. The compression unit receives an uncompressed chunk of data generated by a neuron in the DNN module. The compression unit generates a mask portion and a data portion of a compressed output chunk. The mask portion encodes the presence and location of the zero and non-zero bytes in the uncompressed chunk of data. The data portion stores truncated non-zero bytes from the uncompressed chunk of data. A decompression unit receives a compressed chunk of data from memory in the DNN processor or memory of an application host. The decompression unit decompresses the compressed chunk of data using the mask portion and the data portion.

Neural network processor using compression and decompression of activation data to reduce memory bandwidth utilization

The performance of a neural network (NN) and/or deep neural network (DNN) can limited by the number of operations being performed as well as memory data management of a NN/DNN. Using vector quantization of neuron weight values, the processing of data by neurons can be optimize the number of operations as well as memory utilization to enhance the overall performance of a NN/DNN. Operatively, one or more contiguous segments of weight values can be converted into one or more vectors of arbitrary length and each of the one or more vectors can be assigned an index. The generated indexes can be stored in an exemplary vector quantization lookup table and retrieved by exemplary fast weight lookup hardware at run time on the fly as part of an exemplary data processing function of the NN as part of an inline de-quantization operation to obtain needed one or more neuron weight values.

Flexible hardware for high throughput vector dequantization with dynamic vector length and codebook size

Optimized memory usage and management is crucial to the overall performance of a neural network (NN) or deep neural network (DNN) computing environment. Using various characteristics of the input data dimension, an apportionment sequence is calculated for the input data to be processed by the NN or DNN that optimizes the efficient use of the local and external memory components. The apportionment sequence can describe how to parcel the input data (and its associated processing parameters&#x2014;e.g., processing weights) into one or more portions as well as how such portions of input data (and its associated processing parameters) are passed between the local memory, external memory, and processing unit components of the NN or DNN. Additionally, the apportionment sequence can include instructions to store generated output data in the local and/or external memory components so as to optimize the efficient use of the local and/or external memory components.

Dynamic sequencing of data partitions for optimizing memory utilization and performance of neural networks

Neural processing elements are configured with a hardware AND gate configured to perform a logical AND operation between a sign extend signal and a most significant bit ('MSB') of an operand. The state of the sign extend signal can be based upon a type of a layer of a deep neural network ('DNN') that generate the operand. If the sign extend signal is logical FALSE, no sign extension is performed. If the sign extend signal is logical TRUE, a concatenator concatenates the output of the hardware AND gate and the operand, thereby extending the operand from an N-bit unsigned binary value to an N+1 bit signed binary value. The neural processing element can also include another hardware AND gate and another concatenator for processing another operand similarly. The outputs of the concatenators for both operands are provided to a hardware binary multiplier.

An architecture is disclosed for an neural processing element having single instruction, multiple data ('SIMD') compute lanes. The neural processing element includes compute lanes having multipliers configured to multiply a binary operand with another binary operand to generate a binary output. The neural processing element also includes a single adder tree for summing the binary outputs of the hardware binary multipliers. The neural processing element also includes a storage element for storing a binary output of the single hardware binary adder tree.

Neural processing element with single instruction multiple data (SIMD) compute lanes

A computing system includes processor cores for executing applications that utilize functionality provided by a deep neural network ('DNN') processor. One of the cores operates as a resource and power management ('RPM') processor core. When the RPM processor receives a request to execute a DNN workload, it divides the DNN workload into workload fragments. The RPM processor then determines whether a workload fragment is to be statically allocated or dynamically allocated to a DNN processor. Once the RPM processor has selected a DNN processor, the RPM enqueues the workload fragment on a queue maintained by the selected DNN processor. The DNN processor dequeues workload fragments from its queue for execution. Once execution of a workload fragment has completed, the DNN processor generates an interrupt indicating that execution of the workload fragment has completed. The RPM processor can then notify the processor core that originally requested execution of the workload fragment.

Managing workloads of a deep neural network processor

Techniques to provide for improved (i.e., reduced) power consumption in an exemplary neural network (NN) and/or Deep Neural Network (DNN) environment using data management. Improved power consumption in the NN/DNN may be achieved by reducing a number of bit flips needed to process operands associated with one or more storages. Reducing the number bit flips associated with the NN/DNN may be achieved by multiplying an operand associated with a first storage with a plurality of individual operands associated with a plurality of kernels of the NN/DNN. The operand associated with the first storage may be neuron input data and the plurality of individual operands associated with the second storage may be weight values for multiplication with the neuron input data. The plurality of kernels may be arranged or sorted and subsequently processed in a manner that improves power consumption in the NN/DNN.

Reducing power consumption in a neural network environment using data management

A deep neural network (DNN) module is disclosed that can dynamically partition neuron workload to reduce power consumption. The DNN module includes neurons and a group partitioner and scheduler unit. The group partitioner and scheduler unit divides a workload for the neurons into partitions in order to maximize the number of neurons that can simultaneously process the workload. The group partitioner and scheduler unit then assigns a group of neurons to each of the partitions. The groups of neurons in the DNN module process the workload in their assigned partition to generate a partial output value. The neurons in each group can then sum their partial output values to generate a final output value for the workload. The neurons can be powered down once the groups of neurons have completed processing their assigned workload to reduce power consumption.

Dynamically partitioning workload in a deep neural network module to reduce power consumption

An exemplary artificial intelligence/machine learning hardware computing environment having an exemplary DNN module cooperating with one or more memory components can perform data sharing and distribution as well reuse of a buffer data to reduce the number of memory component read/writes thereby enhancing overall hardware performance and reducing power consumption. Illustratively, data from a cooperating memory component is read according to a selected operation of the exemplary hardware and written to corresponding other memory component for use by one or more processing elements (e.g., neurons). The data is read in such a manner to optimize the engagement of the one or more processing elements for each processing cycle as well as to reuse data previously stored in the one or more cooperating memory components. Operatively, the written data is copied to a shadow memory buffer prior to being consumed by the processing elements.

Enhancing processing performance of artificial intelligence/machine hardware by data sharing and distribution as well as reuse of data in neuron buffer/line buffer

A deep neural network ('DNN') module can determine whether processing of certain values in an input buffer or a weight buffer by neurons can be skipped. For example, the DNN module might determine whether neurons can skip the processing of values in entire columns of a neuron buffer. Processing of these values might be skipped if an entire column of an input buffer or a weight buffer are zeros, for example. The DNN module can also determine whether processing of single values in rows of the input buffer or the weight buffer can be skipped (e.g. if the values are zero). Neurons that complete their processing early as a result of skipping operations can assist other neurons with their processing. A combination operation can be performed following the completion of processing that transfers the results of the processing operations performed by a neuron to their correct owner.

Reducing power consumption in a neural network processor by skipping processing operations

A deep neural network (DNN) module utilizes parallel kernel and parallel input processing to decrease bandwidth utilization, reduce power consumption, improve neuron multiplier stability, and provide other technical benefits. Parallel kernel processing enables the DNN module to load input data only once for processing by multiple kernels. Parallel input processing enables the DNN module to load kernel data only once for processing with multiple input data. The DNN module can implement other power-saving techniques like clock-gating (i.e. removing the clock from) and power-gating (i.e. removing the power from) banks of accumulators based upon usage of the accumulators. For example, individual banks of accumulators can be power-gated when all accumulators in a bank are not in use, and do not store data for a future calculation. Banks of accumulators can also be clock-gated when all accumulators in a bank are not in use, but store data for a future calculation.

Power-efficient deep neural network module configured for parallel kernel and parallel input processing

The performance of a neural network (NN) and/or deep neural network (DNN) can be limited by the number of operations being performed as well as management of data among the various memory components of the NN/DNN. By inserting a selected padding in the input data to align the input data in memory, data read/writes can be optimized for processing by the NN/DNN thereby enhancing the overall performance of a NN/DNN. Operatively, an operations controller/iterator can generate one or more instructions that inserts the selected padding into the data. The data padding can be calculated using various characteristics of the input data as well as the NN/DNN as well as characteristics of the cooperating memory components. Padding on the output data can be utilized to support the data alignment at the memory components and the cooperating processing units of the NN/DNN.

Growing community of inventors

Seattle, WA, United States of America

Larry Marvin Wall