A memory expansion device operable with a host computer system (host) comprises a non-volatile memory (NVM) subsystem, cache memory, and control logic configurable to receive a submission from the host including a read command and specifying a payload in the NVM subsystem and demand data in the payload. The control logic is configured to request ownership of a set of cache lines corresponding to the payload, to indicate completion of the submission after acquiring ownership of the cache lines, and to load the payload to the cache memory. The set of cache lines correspond to a set of cache lines in a coherent destination memory space accessible by the host. The control logic is further configured to, after indicating completion of the submission and in response to a request from the host to read demand data in the payload, return the demand data after determining that the demand data is in the cache memory.

Embodiments of the disclosure provide methods and systems for memory management. The method can include: receiving a request for allocating target node data to a memory space, wherein the memory space includes a buffer and an external memory and the target node data comprises property data and structural data and represents a target node of a graph having a plurality of nodes and edges; determining a node degree associated with the target node data; allocating the target node data to the memory space based on the determined node degree.

A method of performing a copy-on-write on a shared memory page is carried out by a device communicating with a processor via a coherence interconnect. The method includes: adding a page table entry so that a request to read a first cache line of the shared memory page includes a cache-line address of the shared memory page and a request to write to a second cache line of the shared memory page includes a cache-line address of a new memory page; in response to the request to write to the second cache line, storing new data of the second cache line in a second memory and associating the second cache-line address with the new data stored in the second memory; and in response to a request to read the second cache line, reading the new data of the second cache line from the second memory.

Techniques for providing an overlap data buffer to store portions of tiles between passes of chained layers of a neural network are described. One accelerator circuit includes one or more processing units to execute instructions corresponding to the chained layers in multiple passes. In a first pass, the processing unit(s) receives a first input tile of an input feature map from a primary buffer and performs a first operation on the first input tile to obtain a first output tile. The processing unit stores the first output tile in the primary buffer and identifies a portion of the first output tile as corresponding to overlap data between tiles of the input feature map. The processing unit stores the portion in a secondary buffer. In a second pass, the processing unit retrieves the portion to avoid fetching the portion that overlaps and computing the overlap data again.

Embodiments are generally directed to methods and apparatuses for dynamically changing data priority in a cache. An embodiment of an apparatus comprising: a priority controller to: receive a memory access request to request data; and set a priority flag for the memory access request based on an accumulated access amount of data stored in a memory block to be accessed by the memory access request to dynamically change a priority level of the requested data.

In one embodiment, a cache coherent system includes one or more agents (e.g., coherent agents) that may cache data used by the system. The system may include a point of coherency in a memory controller in the system, and thus the agents may transmit read requests to the memory controller to coherently read data. The point of coherency may determine if the data is cached in another agent, and may transmit a copy back request to the other agent if the other agent has modified the data. The system may include an interconnect between the agents and the memory controller. At a point on the interconnect at which traffic from the agents converges, a copy back response may be converted to a fill for the requesting agent.

Systems and methods are provided for using a distributed cache architecture with different methods to load balance requests depending upon whether a requested data item is a freely-requested item (e.g., a “hot key”). The cache may be implemented as a consistent hash ring, and most keys may be assigned to particular node based on a consistent hash. For hot key requests, the requests may be distributed among a subset of nodes rather than being assigned to a specific node using consistent hashing. When a witness service is used to ensure that cached data is fresh, verification requests for data regarding hot keys may be batched to avoid overloading the witness service with hot key requests.

A system comprises a processor coupled to a plurality of memory units. Each of the plurality of memory units includes a request processing unit and a plurality of memory banks. Each request processing unit includes a plurality of decomposition units and a crossbar switch, the crossbar switch communicatively connecting each of the plurality of decomposition units to each of the plurality of memory banks. The processor includes a plurality of processing elements and a communication network communicatively connecting the plurality of processing elements to the plurality of memory units. At least a first processing element of the plurality of processing elements includes a control logic unit and a matrix compute engine. The control logic unit is configured to access the plurality of memory units using a dynamically programmable distribution scheme.

A system comprises a processor coupled to a plurality of memory units. Each of the plurality of memory units includes a request processing unit and a plurality of memory banks. Each request processing unit includes a plurality of decomposition units and a crossbar switch, the crossbar switch communicatively connecting each of the plurality of decomposition units to each of the plurality of memory banks. The processor includes a plurality of processing elements and a communication network communicatively connecting the plurality of processing elements to the plurality of memory units. At least a first processing element of the plurality of processing elements includes a control logic unit and a matrix compute engine. The control logic unit is configured to access the plurality of memory units using a dynamically programmable distribution scheme.

An artificial intelligence chip and a data operation method are provided. The artificial intelligence chip receives a command carrying first data and address information and includes a chip memory, a computing processor, a base address register, and an extended address processor. The base address register is configured to access an extended address space in the chip memory. The extended address processor receives the command. The extended address processor determines an operation mode of the first data according to the address information. When the address information points to a first section of the extended address space, the extended address processor performs a first operation on the first data. When the address information points to a section other than the first section of the extended address space, the extended address processor notifies the computing processor of the operation mode and the computing processor performs a second operation on the first data.