Abstract: 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 management of data among the various memory components of the NN/DNN. Using virtualized hardware iterators, data for processing by the NN/DNN can be traversed and configured to optimize the number of operations as well as memory utilization to enhance the overall performance of a NN/DNN. Operatively, an iterator controller can generate instructions for execution by the NN/DNN representative of one more desired iterator operation types and to perform one or more iterator operations. Data can be iterated according to a selected iterator operation and communicated to one or more neuron processors of the NN/DD for processing and output to a destination memory. The iterator operations can be applied to various volumes of data (e.g., blobs) in parallel or multiple slices of the same volume.

Abstract: 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 flyas 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.

Abstract: The performance of a neural network (NN) can be limited by the number of operations being performed. Using a line buffer that is directed to shift a memory block by a selected shift stride for cooperating neurons, data that is operatively residing memory and which would require multiple write cycles into a cooperating line buffer can be processed as in a single line buffer write cycle thereby enhancing the performance of a NN/DNN. A controller and/or iterator can generate one or more instructions having the memory block shifting values for communication to the line buffer. The shifting values can be calculated using various characteristics of the input data as well as the NN/DNN inclusive of the data dimensions. The line buffer can read data for processing, shift the data of the memory block and write the data in the line buffer for subsequent processing.

Abstract: 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.

Abstract: 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.

Abstract: 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.

Abstract: 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 management of data among the various memory components of the NN/DNN. Using virtualized hardware iterators, data for processing by the NN/DNN can be traversed and configured to optimize the number of operations as well as memory utilization to enhance the overall performance of a NN/DNN. Operatively, an iterator controller can generate instructions for execution by the NN/DNN representative of one more desired iterator operation types and to perform one or more iterator operations. Data can be iterated according to a selected iterator operation and communicated to one or more neuron processors of the NN/DD for processing and output to a destination memory. The iterator operations can be applied to various volumes of data (e.g., blobs) in parallel or multiple slices of the same volume.

Abstract: 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 management of data among the various memory components of the NN/DNN. Using virtualized hardware iterators, data for processing by the NN/DNN can be traversed and configured to optimize the number of operations as well as memory utilization to enhance the overall performance of a NN/DNN. Operatively, an iterator controller can generate instructions for execution by the NN/DNN representative of one more desired iterator operation types and to perform one or more iterator operations. Data can be iterated according to a selected iterator operation and communicated to one or more neuron processors of the NN/DD for processing and output to a destination memory. The iterator operations can be applied to various volumes of data (e.g., blobs) in parallel or multiple slices of the same volume.

Abstract: An exemplary computing environment having a DNN module can maintain one or more bandwidth throttling mechanisms. Illustratively, a first throttling mechanism can specify the number of cycles to wait between transactions on a cooperating fabric component (e.g., data bus). Illustratively, a second throttling mechanism can be a transaction count limiter that operatively sets a threshold of a number of transactions to be processed during a given transaction sequence and limits the number of transactions such as multiple transactions in flight to not exceed the set threshold. In an illustrative operation, in executing these two exemplary calculated throttling parameters, the average bandwidth usage and the peak bandwidth usage can be limited. Operatively, with this fabric bandwidth control, the processing units of the DNN are optimized to process data across each transaction cycle resulting in enhanced processing and lower power consumption.

Abstract: An exemplary computing environment having a DNN module can maintain one or more bandwidth throttling mechanisms. Illustratively, a first throttling mechanism can specify the number of cycles to wait between transactions on a cooperating fabric component (e.g., data bus). Illustratively, a second throttling mechanism can be a transaction count limiter that operatively sets a threshold of a number of transactions to be processed during a given transaction sequence and limits the number of transactions such as multiple transactions in flight to not exceed the set threshold. In an illustrative operation, in executing these two exemplary calculated throttling parameters, the average bandwidth usage and the peak bandwidth usage can be limited. Operatively, with this fabric bandwidth control, the processing units of the DNN are optimized to process data across each transaction cycle resulting in enhanced processing and lower power consumption.

Abstract: Hardware accelerated synchronization of data movement across multiple direct memory access (DMA) engines is provided using techniques in which the order of descriptor processing is guaranteed for scenarios involving a single CPU and multiple DMA engines as well as those involving multiple CPUs and multiple DMA engines.

Abstract: A direct memory access (DMA) engine may be responsible to enable and control DMA data flow within a computing system. The DMA engine moves blocks of data, associated with descriptors in a plurality of queues, from a source to a destination memory location or address, autonomously from control by a computer system's processor. Based on analysis of the data blocks linked to the descriptors in the queues, the DMA engine and its associated DMA fragmenter ensure that data blocks stored linked to descriptors in the queues do not remain idle for an exorbitant period of time. The DMA fragmenter may divide large data blocks into smaller data blocks to ensure that the processing of large data blocks does not preclude the timely processing of smaller data blocks associated with one or more descriptors in the queues. The data blocks stored may be two-dimensional data blocks.

Abstract: Hardware accelerated synchronization of data movement across multiple direct memory access (DMA) engines is provided using techniques in which the order of descriptor processing is guaranteed for scenarios involving a single CPU and multiple DMA engines as well as those involving multiple CPUs and multiple DMA engines.

Abstract: A deep neural network (DNN) processor is configured to execute layer descriptors in layer descriptor lists. The descriptors define instructions for performing a forward pass of a DNN by the DNN processor. The layer descriptors can also be utilized to manage the flow of descriptors through the DNN module. For example, layer descriptors can define dependencies upon other descriptors. Descriptors defining a dependency will not execute until the descriptors upon which they are dependent have completed. Layer descriptors can also define a “fence,” or barrier, function that can be used to prevent the processing of upstream layer descriptors until the processing of all downstream layer descriptors is complete. The fence bit guarantees that there are no other layer descriptors in the DNN processing pipeline before the layer descriptor that has the fence to be asserted is processed.

Abstract: An exemplary computing environment having a DNN module can maintain one or more bandwidth throttling mechanisms. Illustratively, a first throttling mechanism can specify the number of cycles to wait between transactions on a cooperating fabric component (e.g., data bus). Illustratively, a second throttling mechanism can be a transaction count limiter that operatively sets a threshold of a number of transactions to be processed during a given transaction sequence and limits the number of transactions such as multiple transactions in flight to not exceed the set threshold. In an illustrative operation, in executing these two exemplary calculated throttling parameters, the average bandwidth usage and the peak bandwidth usage can be limited. Operatively, with this fabric bandwidth control, the processing units of the DNN are optimized to process data across each transaction cycle resulting in enhanced processing and lower power consumption.

Abstract: A direct memory access (DMA) engine may be responsible to enable and control DMA data flow within a computing system. The DMA engine moves blocks of data, associated with descriptors in a plurality of queues, from a source to a destination memory location or address, autonomously from control by a computer system's processor. Based on analysis of the data blocks linked to the descriptors in the queues, the DMA engine and its associated DMA fragmenter ensure that data blocks stored linked to descriptors in the queues do not remain idle for an exorbitant period of time. The DMA fragmenter may divide large data blocks into smaller data blocks to ensure that the processing of large data blocks does not preclude the timely processing of smaller data blocks associated with one or more descriptors in the queues. The data blocks stored may be two-dimensional data blocks.

Abstract: 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—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.

Abstract: 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.

Abstract: 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 flyas 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.

Abstract: 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.