Abstract: Methods and apparatuses relating to processing neural networks are described. In one embodiment, an apparatus to process a neural network includes a plurality of fully connected layer chips coupled by an interconnect; a plurality of convolutional layer chips each coupled by an interconnect to a respective fully connected layer chip of the plurality of fully connected layer chips and each of the plurality of fully connected layer chips and the plurality of convolutional layer chips including an interconnect to couple each of a forward propagation compute intensive tile, a back propagation compute intensive tile, and a weight gradient compute intensive tile of a column of compute intensive tiles between a first memory intensive tile and a second memory intensive tile.

Abstract: Methods and apparatuses relating to processing neural networks are described. In one embodiment, an apparatus to process a neural network includes a plurality of fully connected layer chips coupled by an interconnect; a plurality of convolutional layer chips each coupled by an interconnect to a respective fully connected layer chip of the plurality of fully connected layer chips and each of the plurality of fully connected layer chips and the plurality of convolutional layer chips including an interconnect to couple each of a forward propagation compute intensive tile, a back propagation compute intensive tile, and a weight gradient compute intensive tile of a column of compute intensive tiles between a first memory intensive tile and a second memory intensive tile.

Abstract: Processors may be enhanced by embedding programmable logic devices, such as field-programmable gate arrays. For instance, an application-specific integrated circuit device may include main fixed function circuitry operable to perform a main fixed function of the application-specific integrated circuit device. The application-specific integrated circuit also includes a support processor that performs operations outside of the main fixed function of the application-specific integrated circuit device, wherein the support processor comprises an embedded programmable fabric to provide programmable flexibility to application-specific integrated circuit device.

Abstract: The present disclosure is directed to 3-D stacked architecture for Programmable Fabrics and Central Processing Units (CPUs). The 3-D stacked orientation enables reconfigurability of the fabric, and allows the fabric to function using coarse-grained and fine-grained acceleration for offloading CPU processing. Additionally, the programmable fabric may be able to function to interface with multiple other compute chiplet components in the 3-D stacked orientation. This enables multiple compute components to communicate without the need for offloading the data communications between the compute chiplets.

Abstract: Methods and apparatuses relating to processing neural networks are described. In one embodiment, an apparatus to process a neural network includes a plurality of fully connected layer chips coupled by an interconnect; a plurality of convolutional layer chips each coupled by an interconnect to a respective fully connected layer chip of the plurality of fully connected layer chips and each of the plurality of fully connected layer chips and the plurality of convolutional layer chips including an interconnect to couple each of a forward propagation compute intensive tile, a back propagation compute intensive tile, and a weight gradient compute intensive tile of a column of compute intensive tiles between a first memory intensive tile and a second memory intensive tile.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: Embodiments of the invention describe a cache coherency protocol that eliminates the need for ordering between message classes and also eliminates home tracker preallocation. Embodiments of the invention describe a less complex conflict detection and resolution mechanism (at the home agent) without any performance degradation in form of bandwidth or latency compared to prior art solutions. Embodiments of the invention describe a home agent that may receive request messages, e.g., data ownership request messages and data request messages, which include issuance data indicating an order of the respective message issued. Said home agent may determine whether an early or late conflict exists based, at least in part, on a received conflict response message and the issuance data of a most recent completed transaction.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: Embodiments of the invention describe a cache coherency protocol that eliminates the need for ordering between message classes and also eliminates home tracker preallocation. Embodiments of the invention describe a less complex conflict detection and resolution mechanism (at the home agent) without any performance degradation in form of bandwidth or latency compared to prior art solutions. Embodiments of the invention describe a home agent that may receive request messages, e.g., data ownership request messages and data request messages, which include issuance data indicating an order of the respective message issued. Said home agent may determine whether an early or late conflict exists based, at least in part, on a received conflict response message and the issuance data of a most recent completed transaction.

Abstract: Systems and methods of operating a computing system may involve identifying a plurality of state values, wherein each state value corresponds to a computing thread associated with a processor. An average value can be determined for the plurality of state values, wherein a determination may be made as to whether to grant a frequency boost request based at least in part on the average value.

Abstract: Methods and apparatus relating to memory mirroring and migration at a Home Agent (HA) are described. In one embodiment, a home agent may mirror its data at a slave agent. In some embodiments, a bit in a directory may indicate status of cache lines. Other embodiments are also disclosed.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent and if a directory entry associated with the request is in an unknown state, an invalidating snoop message is sent to at least one other caching agent to invalidate information in a cache location of the other caching agent corresponding to the location of the read request, to enable setting of the directory entry into a known state. Other embodiments are described and claimed.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: A method comprises detecting a defective area in a Dynamic Random Access Memory (DRAM). The method further comprises establishing a redundant memory buffer at a per-memory module level. The method still further comprises loading the redundant memory buffer with a copy of data from the defective area. The method additionally comprises substituting data from the redundant memory buffer for data stored in the defective area upon a memory access to the defective area.

Abstract: Systems and methods of operating a computing system may involve identifying a plurality of state values, wherein each state value corresponds to a computing thread associated with a processor. An average value can be determined for the plurality of state values, wherein a determination may be made as to whether to grant a frequency boost request based at least in part on the average value.

Abstract: Methods and apparatus relating to home agent data and memory management are described. In one embodiment, a scrubber logic corrects an error at a location in a memory corresponding to a target address by writing back the corrected version of data to the target location. In an embodiment, a map out logic maps out an index or way of a directory cache in response to a number of errors, corresponding to the directory cache, exceeding a threshold value. Other embodiments are also disclosed.

Abstract: In one embodiment, a method includes receiving a read request from a first caching agent, determining whether a directory entry associated with the memory location indicates that the information is not present in a remote caching agent, and if so, transmitting the information from the memory location to the first caching agent before snoop processing with respect to the read request is completed. Other embodiments are described and claimed.

Abstract: A method is provided for managing defects in a semiconductor memory system having a plurality of addressable locations. In the method, a first plurality of the addressable locations is allocated as in-use locations, and a second plurality of the addressable locations is allocated as spare locations. A plurality of sets of the in-use locations, wherein each of the sets is associated with a memory defect, is determined. At least one of the sets includes a different number of in-use locations than another of the sets. Each of the sets of the in-use locations is associated with at least one corresponding set of the spare locations. Each of a plurality of data requests that is associated with one of the sets of the in-use locations is directed to the at least one corresponding set of the spare locations.