Abstract: A hierarchical network enables access for a stacked memory system including or more memory dies that each include multiple memory tiles. The processor die includes multiple processing tiles that are stacked with the one or more memory die. The memory tiles that are vertically aligned with a processing tile are directly coupled to the processing tile and comprise the local memory block for the processing tile. The hierarchical network provides access paths for each processing tile to access the processing tile's local memory block, the local memory block coupled to a different processing tile within the same processing die, memory tiles in a different die stack, and memory tiles in a different device. The ratio of memory bandwidth (byte) to floating-point operation (B:F) may improve 50× for accessing the local memory block compared with conventional memory. Additionally, the energy consumed to transfer each bit may be reduced by 10×.

Abstract: One embodiment sets forth a technique for performing matrix operations. The technique includes traversing a tree structure to access one or more non-empty regions within a matrix. The tree structure includes a first plurality of nodes and a second plurality of nodes corresponding to non-empty regions in the matrix. The first plurality of nodes includes a first node representing a first region and one or more second nodes that are children of the first node and represent second region(s) with an equal size formed within the first region. The second plurality of nodes include a third node representing a third region and one or more fourth nodes that are children of the third node and represent fourth region(s) with substantially equal numbers of non-zero matrix values formed within the third region. The technique also includes performing matrix operation(s) based on the non-empty region(s) to generate a matrix operation result.

Abstract: One embodiment sets forth a technique for performing matrix operations. The technique includes traversing a tree structure to access one or more non-empty regions within a matrix. The tree structure includes a first plurality of nodes and a second plurality of nodes corresponding to non-empty regions in the matrix. The first plurality of nodes includes a first node representing a first region and one or more second nodes that are children of the first node and represent second region(s) with an equal size formed within the first region. The second plurality of nodes include a third node representing a third region and one or more fourth nodes that are children of the third node and represent fourth region(s) with substantially equal numbers of non-zero matrix values formed within the third region. The technique also includes performing matrix operation(s) based on the non-empty region(s) to generate a matrix operation result.

Abstract: Embodiments of the present disclosure relate to application partitioning for locality in a stacked memory system. In an embodiment, one or more memory dies are stacked on the processor die. The processor die includes multiple processing tiles and each memory die includes multiple memory tiles. Vertically aligned memory tiles are directly coupled to and comprise the local memory block for a corresponding processing tile. An application program that operates on dense multi-dimensional arrays (matrices) may partition the dense arrays into sub-arrays associated with program tiles. Each program tile is executed by a processing tile using the processing tile's local memory block to process the associated sub-array. Data associated with each sub-array is stored in a local memory block and the processing tile corresponding to the local memory block executes the program tile to process the sub-array data.

Abstract: Embodiments of the present disclosure relate to memory page access instrumentation for generating a memory access profile. The memory access profile may be used to co-locate data near the processing unit that accesses the data, reducing memory access energy by minimizing distances to access data that is co-located with a different processing unit (i.e., remote data). Execution thread arrays and memory pages for execution of a program are partitioned across multiple processing units. The partitions are then each mapped to a specific processing unit to minimize inter-partition traffic given the processing unit physical topology.

Abstract: One embodiment sets forth a technique for performing one or more matrix multiplication operations based on a first matrix and a second matrix. The technique includes receiving data associated with the first matrix from a first traversal engine that accesses nonzero elements included in the first matrix via a first tree structure. The technique also includes performing one or more computations on the data associated with the first matrix and the data associated with the second matrix to produce a plurality of partial results. The technique further includes combining the plurality of partial results into one or more intermediate results and storing the one or more intermediate results in a first buffer memory.

Abstract: Embodiments of the present disclosure relate to application partitioning for locality in a stacked memory system. In an embodiment, one or more memory dies are stacked on the processor die. The processor die includes multiple processing tiles and each memory die includes multiple memory tiles. Vertically aligned memory tiles are directly coupled to and comprise the local memory block for a corresponding processing tile. An application program that operates on dense multi-dimensional arrays (matrices) may partition the dense arrays into sub-arrays associated with program tiles. Each program tile is executed by a processing tile using the processing tile's local memory block to process the associated sub-array. Data associated with each sub-array is stored in a local memory block and the processing tile corresponding to the local memory block executes the program tile to process the sub-array data.

Abstract: An initiating processing tile generates an offload request that may include a processing tile ID, source data needed for the computation, program counter, and destination location where the computation result is stored. The offload processing tile may execute the offloaded computation. Alternatively, the offload processing tile may deny the offload request based on congestion criteria. The congestion criteria may include a processing workload measure, whether a resource needed to perform the computation is available, and an offload request buffer fullness. In an embodiment, the denial message that is returned to the initiating processing tile may include the data needed to perform the computation (read from the local memory of the offload processing tile). Returning the data with the denial message results in the same inter-processing tile traffic that would occur if no attempt to offload the computation were initiated.

Abstract: Embodiments of the present disclosure relate to memory page access instrumentation for generating a memory access profile. The memory access profile may be used to co-locate data near the processing unit that accesses the data, reducing memory access energy by minimizing distances to access data that is co-located with a different processing unit (i.e., remote data). Execution thread arrays and memory pages for execution of a program are partitioned across multiple processing units. The partitions are then each mapped to a specific processing unit to minimize inter-partition traffic given the processing unit physical topology.

Abstract: A hierarchical network enables access for a stacked memory system including or more memory dies that each include multiple memory tiles. The processor die includes multiple processing tiles that are stacked with the one or more memory die. The memory tiles that are vertically aligned with a processing tile are directly coupled to the processing tile and comprise the local memory block for the processing tile. The hierarchical network provides access paths for each processing tile to access the processing tile's local memory block, the local memory block coupled to a different processing tile within the same processing die, memory tiles in a different die stack, and memory tiles in a different device. The ratio of memory bandwidth (byte) to floating-point operation (B:F) may improve 50× for accessing the local memory block compared with conventional memory. Additionally, the energy consumed to transfer each bit may be reduced by 10×.

Abstract: A combined on-package and off-package memory system uses a custom base-layer within which are fabricated one or more dedicated interfaces to off-package memories. An on-package processor and on-package memories are also directly coupled to the custom base-layer. The custom base-layer includes memory management logic between the processor and memories (both off and on package) to steer requests. The memories are exposed as a combined memory space having greater bandwidth and capacity compared with either the off-package memories or the on-package memories alone. The memory management logic services requests while maintaining quality of service (QoS) to satisfy bandwidth requirements for each allocation. An allocation may include any combination of the on and/or off package memories. The memory management logic also manages data migration between the on and off package memories.

Abstract: A hierarchical network enables access for a stacked memory system including or more memory dies that each include multiple memory tiles. The processor die includes multiple processing tiles that are stacked with the one or more memory die. The memory tiles that are vertically aligned with a processing tile are directly coupled to the processing tile and comprise the local memory block for the processing tile. The hierarchical network provides access paths for each processing tile to access the processing tile's local memory block, the local memory block coupled to a different processing tile within the same processing die, memory tiles in a different die stack, and memory tiles in a different device. The ratio of memory bandwidth (byte) to floating-point operation (B:F) may improve 50× for accessing the local memory block compared with conventional memory. Additionally, the energy consumed to transfer each bit may be reduced by 10×.

Abstract: Methods of operating a serial data bus divide series of data bits into sequences of one or more bits and encode the sequences as N-level symbols, which are then transmitted at multiple discrete voltage levels. These methods may be utilized to communicate over serial data lines to improve bandwidth and reduce crosstalk and other sources of noise.

Abstract: Embodiments include methods, systems and non-transitory computer-readable computer readable media including instructions for executing a prefetch kernel that includes memory accesses for prefetching data for a processing kernel into a memory, and, subsequent to executing at least a portion of the prefetch kernel, executing the processing kernel where the processing kernel includes accesses to data that is stored into the memory resulting from execution of the prefetch kernel.

Abstract: A combined on-package and off-package memory system uses a custom base-layer within which are fabricated one or more dedicated interfaces to off-package memories. An on-package processor and on-package memories are also directly coupled to the custom base-layer. The custom base-layer includes memory management logic between the processor and memories (both off and on package) to steer requests. The memories are exposed as a combined memory space having greater bandwidth and capacity compared with either the off-package memories or the on-package memories alone. The memory management logic services requests while maintaining quality of service (QoS) to satisfy bandwidth requirements for each allocation. An allocation may include any combination of the on and/or off package memories. The memory management logic also manages data migration between the on and off package memories.

Abstract: A combined on-package and off-package memory system uses a custom base-layer within which are fabricated one or more dedicated interfaces to off-package memories. An on-package processor and on-package memories are also directly coupled to the custom base-layer. The custom base-layer includes memory management logic between the processor and memories (both off and on package) to steer requests. The memories are exposed as a combined memory space having greater bandwidth and capacity compared with either the off-package memories or the on-package memories alone. The memory management logic services requests while maintaining quality of service (QoS) to satisfy bandwidth requirements for each allocation. An allocation may include any combination of the on and/or off package memories. The memory management logic also manages data migration between the on and off package memories.

Abstract: Embodiments of the present disclosure relate to application partitioning for locality in a stacked memory system. In an embodiment, one or more memory dies are stacked on the processor die. The processor die includes multiple processing tiles and each memory die includes multiple memory tiles. Vertically aligned memory tiles are directly coupled to and comprise the local memory block for a corresponding processing tile. An application program that operates on dense multi-dimensional arrays (matrices) may partition the dense arrays into sub-arrays associated with program tiles. Each program tile is executed by a processing tile using the processing tile's local memory block to process the associated sub-array. Data associated with each sub-array is stored in a local memory block and the processing tile corresponding to the local memory block executes the program tile to process the sub-array data.

Abstract: Embodiments of the present disclosure relate to memory stacked on processor for high bandwidth. Systems and methods are disclosed for providing a one-level memory for a processing system by stacking bulk memory on a processor die. In an embodiment, one or more memory dies are stacked on the processor die. The processor die includes multiple processing tiles, where each tile includes a processing unit, mapper, and tile network. Each memory die includes multiple memory tiles. The processing tile is coupled to each memory tile that is above or below the processing tile. The vertically aligned memory tiles comprise the local memory block for the processing tile. The ratio of memory bandwidth (byte) to floating-point operation (B:F) may improve 50× for accessing the local memory block compared with conventional memory. Additionally, the energy consumed to transfer each bit may be reduced by 10×.

Abstract: One embodiment sets forth a technique for generating a tree structure within a computer memory for storing sparse data. The technique includes dividing a matrix into a first plurality of equally sized regions. The technique also includes dividing at least one region in the first plurality of regions into a second plurality of regions, where the second plurality of regions includes a first region and one or more second regions that have a substantially equal number of nonzero matrix values and are formed within the first region. The technique further includes creating the tree structure within the computer memory by generating a first plurality of nodes representing the first plurality of regions, generating a second plurality of nodes representing the second plurality of regions, and grouping, under a first node representing the first region, one or more second nodes representing the one or more second regions.

Abstract: Embodiments include methods, systems and non-transitory computer-readable computer readable media including instructions for executing a prefetch kernel that includes memory accesses for prefetching data for a processing kernel into a memory, and, subsequent to executing at least a portion of the prefetch kernel, executing the processing kernel where the processing kernel includes accesses to data that is stored into the memory resulting from execution of the prefetch kernel.