Abstract: A reconfigurable hardware accelerator for computers combines a high-speed dataflow processor, having programmable functional units rapidly reconfigured in a network of programmable switches, with a stream processor that may autonomously access memory in predefined access patterns after receiving simple stream instructions. The result is a compact, high-speed processor that may exploit parallelism associated with many application-specific programs susceptible to acceleration.

Abstract: A computer architecture employs multiple special-purpose processors having different affinities for program execution to execute substantial portions of general-purpose programs to provide improved performance with respect to a general-purpose processor executing the general-purpose program alone.

Abstract: A hardware accelerator for computers combines a stand-alone, high-speed, fixed program dataflow functional element with a stream processor, the latter of which may autonomously access memory in predefined access patterns after receiving simple stream instructions and provide them to the dataflow functional element. The result is a compact, high-speed processor that may exploit fixed program dataflow functional elements.

Abstract: A dataflow accelerator including a control/command core, a scratchpad and a coarse grain reconfigurable array (CGRA) according to an exemplary embodiment is disclosed. The scratchpad may include a write controller to transmit data to an input vector port interface and to receive data from the input vector port interface. The CGRA may receive data from the input vector port interface and includes a plurality of interconnects and a plurality of functional units.

Abstract: According to exemplary embodiments, a method, processor, and system for accelerating a recurrent neural network are presented. A method of accelerating a recurrent neural network may include distributing from a first master core to each of a plurality of processing cores a same relative one or more columns of weight matrix data for each of a plurality of gates in the neural network, broadcasting a current input vector from the first master core to each of the processing cores, and processing each column of weight matrix data in parallel, at each of the respective processing cores.

Abstract: A method for performing acceleration of simultaneous access to shared data may include providing a plurality of groups of cores and a plurality of shared memory structures, providing a pod comprising the plurality of groups of cores linked by a common broadcast channel, and coordinating each shared memory structure to provide a logically unified memory structure. Each memory structure may be associated with a group of cores, and each group of cores may include one or more cores. The common broadcast channel may be operatively coupled to each shared memory structure. The coordinating each shared memory structure may include identifying a simultaneous read-reuse load to a first shared memory structure, fetching data corresponding to the simultaneous read-reuse load, and forwarding the data to shared memory structures other than the first shared memory structure and to groups of cores other than a first group of cores via the broadcast channel.

Abstract: Aspects of the present invention provide a memory system comprising a plurality of stacked memory layers, each memory layer divided into memory sections, wherein each memory section connects to a neighboring memory section in an adjacent memory layer, and a logic layer stacked among the plurality of memory layers, the logic layer divided into logic sections, each logic section including a memory processing core, wherein each logic section connects to a neighboring memory section in an adjacent memory layer to form a memory vault of connected logic and memory sections, and wherein each logic section is configured to communicate directly or indirectly with a host processor. Accordingly, each memory processing core may be configured to respond to a procedure call from the host processor by processing data stored in its respective memory vault and providing a result to the host processor. As a result, increased performance may be provided.

Abstract: The amount of speed-up that can be obtained by optimizing the program to run on a different architecture is determined by static measurements of the program. Multiple such static measurements are processed by a machine learning system after being discretized to alter their accuracy vs precision. Static analysis requires less analysis overhead and permits analysis of program portions to optimize allocation of porting resources on a large program.

Abstract: According to exemplary embodiments, a method, processor, and system for accelerating a recurrent neural network are presented. A method of accelerating a recurrent neural network may include distributing from a first master core to each of a plurality of processing cores a same relative one or more columns of weight matrix data for each of a plurality of gates in the neural network, broadcasting a current input vector from the first master core to each of the processing cores, and processing each column of weight matrix data in parallel, at each of the respective processing cores.

Abstract: A method for performing acceleration of simultaneous access to shared data may include providing a plurality of groups of cores and a plurality of shared memory structures, providing a pod comprising the plurality of groups of cores linked by a common broadcast channel, and coordinating each shared memory structure to provide a logically unified memory structure. Each memory structure may be associated with a group of cores, and each group of cores may include one or more cores. The common broadcast channel may be operatively coupled to each shared memory structure. The coordinating each shared memory structure may include identifying a simultaneous read-reuse load to a first shared memory structure, fetching data corresponding to the simultaneous read-reuse load, and forwarding the data to shared memory structures other than the first shared memory structure and to groups of cores other than a first group of cores via the broadcast channel.

Abstract: A specialized memory access processor is placed between a main processor and accelerator hardware to handle memory access for the accelerator hardware. The architecture of the memory access processor is designed to allow lower energy memory accesses than can be obtained by the main processor in providing data to the hardware accelerator while providing the hardware accelerator with a sufficiently high bandwidth memory channel. In some embodiments, the main processor may enter a sleep state during accelerator calculations to substantially lower energy consumption.

Abstract: According to some embodiments, a dataflow accelerator comprises a control/command core, a scratchpad and a coarse grain reconfigurable array (CGRA). The scratchpad comprises a write controller to transmit data to an input vector port interface and to receive data from the input vector port interface. The CGRA receives data from the input vector port interface where the CGRA comprising a plurality of interconnects and a plurality of functional units.

Abstract: Aspects of the present invention provide a memory system comprising a plurality of stacked memory layers, each memory layer divided into memory sections, wherein each memory section connects to a neighboring memory section in an adjacent memory layer, and a logic layer stacked among the plurality of memory layers, the logic layer divided into logic sections, each logic section including a memory processing core, wherein each logic section connects to a neighboring memory section in an adjacent memory layer to form a memory vault of connected logic and memory sections, and wherein each logic section is configured to communicate directly or indirectly with a host processor. Accordingly, each memory processing core may be configured to respond to a procedure call from the host processor by processing data stored in its respective memory vault and providing a result to the host processor. As a result, increased performance may be provided.

Abstract: Aspects of the present invention provide a memory system comprising a plurality of stacked memory layers, each memory layer divided into memory sections, wherein each memory section connects to a neighboring memory section in an adjacent memory layer, and a logic layer stacked among the plurality of memory layers, the logic layer divided into logic sections, each logic section including a memory processing core, wherein each logic section connects to a neighboring memory section in an adjacent memory layer to form a memory vault of connected logic and memory sections, and wherein each logic section is configured to communicate directly or indirectly with a host processor. Accordingly, each memory processing core may be configured to respond to a procedure call from the host processor by processing data stored in its respective memory vault and providing a result to the host processor. As a result, increased performance may be provided.

Abstract: A dataflow computer processor is teamed with a general computer processor so that program portions of an application program particularly suited to dataflow execution may be transferred to the dataflow processor during portions of the execution of the application program by the general computer processor. During this time the general computer processor may be placed in partial shutdown for energy conservation.

Abstract: A hardware accelerator for computers combines a stand-alone, high-speed, fixed program dataflow functional element with a stream processor, the latter of which may autonomously access memory in predefined access patterns after receiving simple stream instructions and provide them to the dataflow functional element. The result is a compact, high-speed processor that may exploit fixed program dataflow functional elements.

Abstract: A reconfigurable hardware accelerator for computers combines a high-speed dataflow processor, having programmable functional units rapidly reconfigured in a network of programmable switches, with a stream processor that may autonomously access memory in predefined access patterns after receiving simple stream instructions. The result is a compact, high-speed processor that may exploit parallelism associated with many application-specific programs susceptible to acceleration.

Abstract: A computer architecture employs multiple special-purpose processors having different affinities for program execution to execute substantial portions of general-purpose programs to provide improved performance with respect to a general-purpose processor executing the general-purpose program alone.

Abstract: The amount of speed-up that can be obtained by optimizing the program to run on a different architecture is determined by static measurements of the program. Multiple such static measurements are processed by a machine learning system after being discretized to alter their accuracy vs precision. Static analysis requires less analysis overhead and permits analysis of program portions to optimize allocation of porting resources on a large program.

Abstract: A computer architecture allows for simplified exception handling by restarting the program after exceptions at the beginning of idempotent regions, the idempotent regions allowing re-execution without the need for restoring complex state information from checkpoints. Recovery from mis-speculation may be provided by a similar mechanism but using smaller idempotent regions reflecting a more frequent occurrence of mis-speculation. A compiler generating different idempotent regions for speculation and exception handling is also disclosed.