Abstract: Embodiments relate to a neural processor circuit with scalable architecture for instantiating one or more neural networks. The neural processor circuit includes a data buffer coupled to a memory external to the neural processor circuit, and a plurality of neural engine circuits. To execute tasks that instantiate the neural networks, each neural engine circuit generates output data using input data and kernel coefficients. A neural processor circuit may include multiple neural engine circuits that are selectively activated or deactivated according to configuration data of the tasks. Furthermore, an electronic device may include multiple neural processor circuits that are selectively activated or deactivated to execute the tasks.

Abstract: Embodiments relate to a neural processor circuit with scalable architecture for instantiating one or more neural networks. The neural processor circuit includes a data buffer coupled to a memory external to the neural processor circuit, and a plurality of neural engine circuits. To execute tasks that instantiate the neural networks, each neural engine circuit generates output data using input data and kernel coefficients. A neural processor circuit may include multiple neural engine circuits that are selectively activated or deactivated according to configuration data of the tasks. Furthermore, an electronic device may include multiple neural processor circuits that are selectively activated or deactivated to execute the tasks.

Abstract: Embodiments relate to a neural processor circuit with scalable architecture for instantiating one or more neural networks. The neural processor circuit includes a data buffer coupled to a memory external to the neural processor circuit, and a plurality of neural engine circuits. To execute tasks that instantiate the neural networks, each neural engine circuit generates output data using input data and kernel coefficients. A neural processor circuit may include multiple neural engine circuits that are selectively activated or deactivated according to configuration data of the tasks. Furthermore, an electronic device may include multiple neural processor circuits that are selectively activated or deactivated to execute the tasks.

Abstract: Methods and apparatus for a knowledge-based deep learning refactoring model with tightly integrated functional nonparametric memory are disclosed. An example non-transitory computer readable medium comprises instructions that, when executed, cause a machine to at least estimate a first information extraction cost corresponding to retrieval of information from a local knowledge base, estimate a second information extraction cost corresponding retrieval of information from a remote knowledge base, select an information source based on the first and second estimated information extraction costs, query the selected information source, in response to determining that the selected information source was an external information source, store the queried information in the local knowledge base, organize the stored information in the local knowledge base, and return the queried information.

Abstract: Embodiments of the present disclosure are directed toward techniques and configurations for an optical accelerator including a photonics integrated circuit (PIC) for an optical neural network (ONN). In embodiments, an optical accelerator package includes the PIC and an electronics integrated circuit (EIC) that is heterogeneously integrated into the optical accelerator package to proximally provide pre- and post-processing of optical signal inputs and optical signal outputs provided to and received from an optical matrix multiplier of the PIC. In some embodiments, the EIC is a single EIC or discrete EICs to provide pre- and post-processing of the optical signal inputs and optical signal outputs including optical to electrical and electrical to optical transduction. Other embodiments may be described and/or claimed.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: In an embodiment, a matrix computation engine is configured to perform matrix computations (e.g. matrix multiplications). The matrix computation engine may perform numerous matrix computations in parallel, in an embodiment. More particularly, the matrix computation engine may be configured to perform numerous multiplication operations in parallel on input matrix elements, generating resulting matrix elements. In an embodiment, the matrix computation engine may be configured to accumulate results in a result memory, performing multiply-accumulate operations for each matrix element of each matrix.

Abstract: A communication system has special Agents in the subscriber terminals which detect the need of applications for data paths with QoS over the access network. The Agents have packet based control channels to a Remote Resource Manager installed outside the network typically as a web server and use the control channels for sending bandwidth allocation requests to the Remote Resource Manager which stores all bandwidth relevant information for a subscriber and delivers bandwidth and QoS class back to the Agents which adjust packet rate and packet QoS class marking accordingly. A Self-Sustaining Scheduler placed in the bottlenecks of the data path guarantees given delay times per QoS class and keeps packet drop rate below given limits if the Remote Resource Manager assigns bandwidth appropriately.

Abstract: An idle mode system has a clock gating circuit, a bus interface unit, memory interfaces and an interrupt and idle control unit. The clock gating circuit receives a first clock and designated idle-acknowledge signals. The clock gating circuit produces a second clock signal based on the first clock signal when fewer than all designated idle-acknowledge signals are received. The clock gating circuit produces no second clock signal when all designated idle-acknowledge signals are received. The bus interface unit receives bus access requests and receives the first and second clock signals. When a bus access request is made, the bus interface unit de-asserts its idle-acknowledge signal and passes the bus access request. The memory interfaces operate on the second clock. One interface receives the bus access request from the bus interface unit, withdraws its idle-acknowledge signal, processes the bus access request, and re-asserts its idle-acknowledge signal upon completion.

Abstract: A method and system of controlling out of order execution pipelines using pipeline skew parameters is disclosed. The pipeline skew parameters track the relative position of a load/store instruction in a load/store pipeline and a simultaneously issued integer instruction in a variable length integer pipeline. The pipeline skew parameters are used to improve data hazard detection, pipeline stalling, and instruction cancellation.

Abstract: A thread selection unit for a block multi-threaded processor includes a priority thread selector and an execution thread selector. The priority thread selector uses a maxtime register for each active thread to limit the time an active thread can be the priority thread. The execution thread selector is configured to select the priority thread as the execution thread when the priority thread is unblocked. However, if the priority thread is blocked, the execution thread selector selects a non-priority thread as the execution thread.

Abstract: A multithreaded processor includes a thread ID for each set of fetched bits in an instruction fetch and issue unit. The thread ID attaches to the instructions and operands of the set of fetched bits. Pipeline stages in the multithreaded processor stores the thread ID associated with each operand or instruction in the pipeline stage. The thread ID are used to maintain data coherency and to generate program traces that include thread information for the instructions executed by the multithreaded processor.

Abstract: A multithreaded processor includes a thread ID for each set of fetched bits in an instruction fetch and issue unit. The thread ID attaches to the instructions and operands of the set of fetched bits. Pipeline stages in the multithreaded processor stores the thread ID associated with each operand or instruction in the pipeline stage. The thread ID are used to maintain data coherency and to generate program traces that include thread information for the instructions executed by the multithreaded processor.

Abstract: A microprocessor system includes an address generator, an address selector, and memory system having multiple memory towers, which can be independently addressed. The address generator simultaneously generates a first memory address and a second memory address that is 1 row greater than the first memory address. The address selector determines whether the row portion of the first memory address or the second memory address is used for each memory tower. Because each tower can be addressed independently, a single memory access can be used to access data spanning multiple rows of the memory system.

Abstract: A variable length instruction pipeline includes optional expansion stages that can be included in the variable length instruction pipeline to avoid pipeline stalls. The expansion stages are removed from the variable length instruction pipeline when not needed to reduce the length of the pipeline, which reduces latency and other problems associated with long pipelines. For example, in one embodiment of the present invention, a variable length instruction pipeline includes a first pipeline stage, a first expansion stage, and a second pipeline stage. The second pipeline stage is configured to selectively receive instructions from the first pipeline stage or the first expansion stage if the first expansion stage holds an instruction.

Abstract: A real-time, multi-threaded embedded system includes rules for handling traps and interrupts to avoid problems such as priority inversion and re-entrancy. By defining a global interrupt priority value for all active threads and only accepting interrupts having a priority higher than the interrupt priority value, priority inversion can be avoided. Switching to the same thread before any interrupt servicing, and disabling interrupts and thread switching during interrupt servicing can simplify the interrupt handling logic. By storing trap background data for traps and servicing traps only in their originating threads, trap traceability can be preserved. By disabling interrupts and thread switching during trap servicing, unintended trap re-entrancy and servicing disruption can be prevented.