Abstract: A stacked processor-plus-memory device includes a processing die with an array of processing elements of an artificial neural network. Each processing element multiplies a first operand—e.g. a weight—by a second operand to produce a partial result to a subsequent processing element. To prepare for these computations, a sequencer loads the weights into the processing elements as a sequence of operands that step through the processing elements, each operand stored in the corresponding processing element. The operands can be sequenced directly from memory to the processing elements or can be stored first in cache. The processing elements include streaming logic that disregards interruptions in the stream of operands.

Abstract: Row hammer is mitigated by issuing, to a memory device, mitigation operation (MOP) commands in order to cause the refresh of rows at a specified vicinity of a suspected aggressor row. These mitigation operation commands are each associated with respective row addresses that indicate the suspected aggressor row and an indicator of which neighbor row in the vicinity of the suspected aggressor row is to be refreshed. The mitigation operation commands are issued in response to a fixed number of activate commands. The suspected aggressor row is selected by randomly choosing, with equal probability, one of the N previous activate commands to supply its associated row address as the suspected aggressor row address. The neighbor row may be selected randomly with a probability that diminishes inversely with the distance between the suspected aggressor row and the neighbor row.

Abstract: Multiple independent point-to-point memory channels are operated, by at least one controller, in parallel to form a wider memory channel. The memory components on these point-to-point channels include the ability to connect to multiple (e.g., 2) instances of these independent memory channels. The controller operates multiple instances of the wider channels with the memory components configured in a clamshell mode. A single memory component is also operated in clamshell mode to provide error correction code information, independently of the other wider channels, to multiple instances of the wider memory channel.

Abstract: An integrated circuit (IC) memory device includes an array of storage cells configured into multiple banks. Interface circuitry receives refresh commands from a host memory controller to refresh the multiple banks for a first refresh mode. On-die refresh control circuitry selectively generates local refresh commands to refresh the multiple banks in cooperation with the host memory controller during a designated hidden refresh interval in a second refresh mode. Mode register circuitry stores a value indicating whether the on-die refresh control circuitry is enabled for use during the second refresh mode. The interface circuitry includes backchannel control circuitry to transmit a corrective action control signal during operation in the second refresh mode.

Abstract: A high-capacity cache memory is implemented by multiple heterogenous DRAM dies, including a dedicated tag-storage DRAM die architected for low-latency tag-address retrieval and thus rapid hit/miss determination, and one or more capacity-optimized cache-line DRAM dies that render a net cache-line storage capacity orders of magnitude beyond that of state-of-the art SRAM cache implementations. The tag-storage die serves double-duty in some implementations, yielding rapid tag hit/miss determination for cache-line read/write requests while also serving as a high-capacity snoop-filter in a memory-sharing multiprocessor environment.

Abstract: An interconnected stack of one or more Dynamic Random Access Memory (DRAM) die has a base logic die and one or more custom logic or processor die. The processor logic die snoops commands sent to and through the stack. In particular, the processor logic die may snoop mode setting commands (e.g., mode register set—MRS commands). At least one mode setting command that is ignored by the DRAM in the stack is used to communicate a command to the processor logic die. In response the processor logic die may prevent commands, addresses, and data from reaching the DRAM die(s). This enables the processor logic die to send commands/addresses and communicate data with the DRAM die(s). While being able to send commands/addresses and communicate data with the DRAM die(s), the processor logic die may execute software using the DRAM die(s) for program and/or data storage and retrieval.

Abstract: One or more neural network layers are implemented by respective sets of signed multiply-accumulate units that generate dual analog result signals indicative of positive and negative product accumulations, respectively. The two analog result signals and thus the positive and negative product accumulations are differentially combined to produce a merged analog output signal that constitutes the output of a neural node within the subject neural network layer.

Abstract: Multiple device stacks are interconnected in a ring topology. The inter-device stack communication may utilize a handshake protocol. This ring topology may include the host so that the host may initialize and load the device stacks with data and/or commands (e.g., software, algorithms, etc.). The inter-device stack interconnections may also be configured to include/remove the host and/or to implement varying numbers of separate ring topologies. By configuring the system with more than one ring topology, and assigning different problems to different rings, multiple, possibly unrelated, machine learning tasks may be performed in parallel by the device stack system.

Abstract: A DRAM includes at least four groups of memory cores and at least four memory access channel interfaces that, in a first mode, each respectively are to receive memory access commands, directed to a corresponding one of the groups of memory cores. One-half of the memory access channel interfaces are to, in a second mode, each respectively receive memory access commands, directed to a corresponding two of four of the groups of memory cores. The memory access channel interfaces to have electrical connection conductors that lie on opposing sides of at least one line of reflectional symmetry from a second one-half of the one-half of the at least four memory access channel interfaces.

Abstract: A system that calibrates timing relationships between signals involved in performing write operations is described. This system includes a memory controller which is coupled to a set of memory chips, wherein each memory chip includes a phase detector configured to calibrate a phase relationship between a data-strobe signal and a clock signal received at the memory chip from the memory controller during a write operation. Furthermore, the memory controller is configured to perform one or more write-read-validate operations to calibrate a clock-cycle relationship between the data-strobe signal and the clock signal, wherein the write-read-validate operations involve varying a delay on the data-strobe signal relative to the clock signal by a multiple of a clock period.

Abstract: Computing devices, methods, and systems for switch-based free memory tracking in data center environments are disclosed. An exemplary switch integrated circuit (IC), which is used in a switched fabric or a network, can include a processing device and a tracking structure that is distributed with at least a second switch IC. The tracking structure tracks free memory units that are accessible in a first set of nodes by the second switch IC. The processing device receives a request for a number of free memory units. The processing device forwards the request to a node in the first set of nodes that has at least the number of free memory units or forwards the request to the second switch IC that has at least the number of free memory units or responds to the request with a response that indicates that the request could not be fulfilled.

Abstract: An application-specific integrated circuit for an artificial neural network is integrated with a high-bandwidth memory. A processing die with tiled neural-network processing units is bonded to a stack of memory dies with memory banks laid out to establish relatively short connections to overlying processing units. The memory banks form vertical groups of banks for each overlying processing unit. A switch matrix on the processing die allows each processing unit to communicate with its vertical group of banks via a short, fast inter-die memory channel or with more remote groups of banks under neighboring processing units.

Abstract: Refresh management commands are issued to a memory device in order to cause the refresh of rows in the vicinity of one or more rows being “hammered.” These refresh management commands are each associated with respective row addresses that indicates the row(s) to be refreshed in order to mitigate the likelihood of data being corrupted by “row hammer.” In an embodiment, the refresh management commands are issued in response to a varying number of activate (ACT) commands having been issued since the last refresh management command. The row selected for a given refresh management command may be selected based on rows that have recently been activated. The selection may be based on “pools” of recently activated rows where these pools are of unequal size. The selection from a given pool may be based on algorithmic and/or random techniques.

Abstract: A DRAM includes at least four groups of memory cores and at least four memory access channel interfaces that, in a first mode, each respectively are to receive memory access commands, directed to a corresponding one of the groups of memory cores. One-half of the memory access channel interfaces are to, in a second mode, each respectively receive memory access commands, directed to a corresponding two of four of the groups of memory cores. The memory access channel interfaces to have electrical connection conductors that lie on opposing sides of at least one line of reflectional symmetry from a second one-half of the one-half of the at least four memory access channel interfaces.

Abstract: Multiple device stacks are interconnected in a ring topology. The inter-device stack communication may utilize a handshake protocol. This ring topology may include the host so that the host may initialize and load the device stacks with data and/or commands (e.g., software, algorithms, etc.). The inter-device stack interconnections may also be configured to include/remove the host and/or to implement varying numbers of separate ring topologies. By configuring the system with more than one ring topology, and assigning different problems to different rings, multiple, possibly unrelated, machine learning tasks may be performed in parallel by the device stack system.

Abstract: Multiple independent point-to-point memory channels are operated, by at least one controller, in parallel to form a wider memory channel. The memory components on these point-to-point channels include the ability to connect to multiple (e.g., 2) instances of these independent memory channels. The controller operates multiple instances of the wider channels with the memory components configured in a clamshell mode. A single memory component is also operated in clamshell mode to provide error correction code information, independently of the other wider channels, to multiple instances of the wider memory channel.

Abstract: A system that calibrates timing relationships between signals involved in performing write operations is described. This system includes a memory controller which is coupled to a set of memory chips, wherein each memory chip includes a phase detector configured to calibrate a phase relationship between a data-strobe signal and a clock signal received at the memory chip from the memory controller during a write operation. Furthermore, the memory controller is configured to perform one or more write-read-validate operations to calibrate a clock-cycle relationship between the data-strobe signal and the clock signal, wherein the write-read-validate operations involve varying a delay on the data-strobe signal relative to the clock signal by a multiple of a clock period.

Abstract: Processing elements include interfaces that allow direct access to memory banks on one or more DRAMs in an integrated circuit stack. These additional (e.g., per processing element) direct interfaces may allow the processing elements to have direct access to the data in the DRAM stack. Based on the size/type of operands being processed, and the memory bandwidth of the direct interfaces, rate calculation circuitry on the processor die determines the speed each processing element and/or processing nodes within each processing element are operated.

Abstract: Command/address and timing information is distributed to buffer integrated circuits on a module using multiple wavelengths of light modulated with the same information. Each individual wavelength of modulated light carrying command/address information is received by a corresponding single buffer device that deserializes the command/address information and communicates it electrically to memory devices(s). Likewise, each individual wavelength of modulated light carrying timing/synchronization/clock information is received by a corresponding single buffer device and used to synchronize accesses to the memory device(s). Thus, multiple buffer integrated circuits on a module each receive information from the CPU using different wavelengths of light transmitted on the same waveguide.

Abstract: Multiple independent point-to-point memory channels are operated, by at least one controller, in parallel to form a wider memory channel. The memory components on these point-to-point channels include the ability to connect to multiple (e.g., 2) instances of these independent memory channels. The controller operates multiple instances of the wider channels with the memory components configured in a clamshell mode. A single memory component is also operated in clamshell mode to provide error correction code information, independently of the other wider channels, to multiple instances of the wider memory channel.