Memory devices, systems and methods are described, such as those including a dynamically configurable channel depth. Devices, systems and methods are described that adjust channel depth based on hardware and/or software requirements. One such device provides for virtual memory operations where a channel depth is adjusted for the same physical memory region responsive to requirements of different memory processes.

A storage system caches logical-to-physical address table entries read in volatile memory. The logical-to-physical address table entries are stored in codewords. The storage system can vary a number or size of an entry in a codeword. Additionally or alternatively, each codeword can store both complete and partial logical-to-physical address table entries. In one example, a codeword having 62 bytes of data and two bytes of error correction code stores 15 complete logical-to-physical address table entries and one partial logical-to-physical address table entry, where the remainder of the partial entry is stored in another codeword. This configuration strikes a good balance between storage space efficiency and random-access write performance.

Disclosed are systems and methods that determine whether instances of data (e.g., forward activations, backward derivatives of activations) that are used to train deep neural networks are to be stored on-chip or off-chip. The disclosed systems and methods are also used to prune the data (discard or delete selected instances of data). A system includes a hierarchical arrangement of on-chip and off-chip memories, and also includes a hierarchical arrangement of data selector devices that are used to decide whether to discard data and where in the system the data is to be discarded.

A network interface device has a data source, a data sink and an interconnect configured to receive data from the data source and to output data to the data sink. The interconnect has a memory having memory cells. Each memory cell has a width which matches a bus segment width. The memory is configured to receive a first write output with a width corresponding to the bus segment width. The write output comprises first data to be written to a first memory cell of the memory, the first data being from the data source.

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.

Apparatuses and methods can be related to supplementing AI processing in memory. An accelerator and/or a host can perform AI processing. Some of the operations comprising the AI processing can be performed by a memory device instead of by an accelerator and/or a host. The memory device can perform AI processing in conjunction with the host and/or accelerator.

Described are memory modules that include address-buffer components and data-buffer components that together support wide- and narrow-data modes. The address-buffer component manages communication between a memory controller and two sets of memory components. In the wide-data mode, the address-buffer enables memory components in each set and instructs the data-buffer components to communicate full-width read and write data by combining data from or to from both sets for each memory access. In the narrow-data mode, the address-buffer enables memory components in just one of the two sets and instructs the data-buffer components to half-width read and write data with one set per memory access.

An apparatus and a method are disclosed. In the apparatus, a memory management unit includes: a first cache unit, adapted to store a plurality of first source operands and one first write address; a second cache unit, adapted to store at least one pair of a second source operand and a second destination address; a write cache module, adapted to discriminate between destination addresses of a plurality of store instructions, so as to store, in the first cache unit, a plurality of source operands corresponding to consecutive destination addresses, and to store, in the second cache unit, non-consecutive destination addresses and source operands corresponding to the non-consecutive destination addresses, where the first write address is an initial address of the consecutive destination addresses; and a bus transmission module, adapted to transmit the plurality of first source operands and the first write address in the first cache unit to a memory through a bus in a write burst transmission mode. In embodiments of the present invention, a burst transmission mode is established between a processor and the bus. This can help reduce occupation of bus address bandwidth and accelerate write efficiency of the memory.

The present disclosure provides an interconnect for a non-uniform memory architecture platform to provide remote access where data can dynamically and adaptively be compressed and decompressed at the interconnect link. A requesting interconnect link can add a delay to before transmitting requested data onto an interconnect bus, compress the data before transmission, or packetize and compress data before transmission. Likewise, a remote interconnect link can decompress request data.

A computer comprising a plurality of processor devices connected in a ring, wherein each of the processor devices is connected to each of two neighbouring ones of the processor devices by a respective physical inter-processor link. Each of a set of external memory device stores a local portion of the externally stored dataset. Each processor device executes instructions to: determine that a synchronisation point has been reached by the plurality of processor devices; responsive to the determination, access from its connected external memory device its local portion of the externally stored dataset stored; record a copy of its local portion of the externally stored dataset in its local memory; transmit its local portion of the externally stored dataset to at least one of its connected neighbouring processing devices; and receive an incoming portion of the externally stored dataset from at least one of its connected neighbouring processing devices.