89 results about "High Bandwidth Memory" patented technology

A memory system is divided into memory subsystems. Each subsystem includes a slave controller. Each slave controller is coupled to a serial link. A master controller is coupled to the slave controllers via the serial links, and the master controller is capable of initiating a memory access to a memory subsystem by communicating with the slave controller via the serial link. Each memory subsystem includes memory arrays coupled to the slave controller. Each memory array includes memory channels. Memory accesses to a memory array on a memory subsystem are interleaved by the slave controller between the memory channels, and memory accesses to a memory subsystem are striped by the slave controller between the memory arrays on the memory subsystem. Memory accesses are striped between memory subsystems by the master controller. The master controller and slave controllers communicate by sending link packets and protocol packets over the serial links.

A specialized memory access processor is placed between a main processor and accelerator hardware to handle memory access for the accelerator hardware. The architecture of the memory access processor is designed to allow lower energy memory accesses than can be obtained by the main processor in providing data to the hardware accelerator while providing the hardware accelerator with a sufficiently high bandwidth memory channel. In some embodiments, the main processor may enter a sleep state during accelerator calculations to substantially lower energy consumption.

A short latency and high bandwidth memory includes a systolic memory that is sub-divided into a plurality of memory arrays, including banks and pipelines that access these banks. Shorter latency and faster performance is achieved with this memory, because each bank is smaller in size and is accessed more rapidly. A high throughput rate is accomplished because of the pipelining. Memory is accessed at the pipeline frequency with the proposed read and write mechanism. Design complexity is reduced because each bank within the memory is the same and repeated. The memory array size is re-configured and organized to fit within desired size and area parameters.

Disclosed herein are embodiments of an asynchronous memory device that use internal delay elements to enable memory access pipelining. In one embodiment, the delay elements are responsive to an input load control signal, and are calibrated with reference to periodically received timing pulses. Different numbers of the delay elements are configured to produce different asynchronous delays and to strobe sequential pipeline elements of the memory device.

The present invention relates to a system according to claim 1, where the pixel buffer cache comprises at least one row descriptor for tracking and monitoring the activities of read and write requests of a particular tile. A system for providing a high bandwidth memory access to a graphics processor comprising: (a) a frame buffer for storing at least one frame, where said frame is stored in a tiled manner; (b) a memory controller for controlling said frame buffer; (c) a pixel buffer cache for storing multiple sections of at least one memory row of said frame buffer, and for processing requests to access pixels of said frame buffer; (d) a graphics accelerator having an interface to said pixel buffer cache for processing a group of related pixels; and (e) a CPU for processing graphic commands and controlling said graphics accelerator and said pixel buffer cache.

According to some embodiments, a high bandwidth memory device includes a base die and a plurality of memory dies stacked on the base die and electrically connected to the base die through a plurality of through substrate vias. The base die includes a plurality of first input buffers configured to receive channel clock signals, channel command/addresses, and channel data from a plurality of first bumps connected to the outside of the base die, a plurality of second input buffers configured to receive test clock signals, test command/addresses, and test data from a plurality of second bumps connected to the outside of the base die, a monitoring unit, a plurality of first output buffers connected to the monitoring unit and configured to output monitored data from the monitoring unit to the plurality of second bumps, and a plurality of paths from the plurality of first input buffers to the monitoring unit. The plurality of second bumps are connected to receive test clock signals, test command/addresses, and test data from the outside of the base die during a first operation mode, and to receive monitored data from the plurality of first output buffers during a second operation mode.

This invention describes an improved high bandwidth chip-to-chip interface for memory devices, which is capable of operating at higher speeds, while maintaining error free data transmission, consuming lower power, and supporting more load. Accordingly, the invention provides a memory subsystem comprising at least two semiconductor devices; a main bus containing a plurality of bus lines for carrying substantially all data and command information needed by the devices, the semiconductor devices including at least one memory device connected in parallel to the bus; the bus lines including respective row command lines and column command lines; a clock generator for coupling to a clock line, the devices including clock inputs for coupling to the clock line; and the devices including programmable delay elements coupled to the clock inputs to delay the clock edges for setting an input data sampling time of the memory device.

A method of coordinating memory commands in a high-bandwidth memory HBM+ system, the method including sending a host memory controller command from a host memory controller to a memory, receiving the host memory controller command at a coordinating memory controller, forwarding the host memory controller command from the coordinating memory controller to the memory, and scheduling, by the coordinating memory controller, a coordinating memory controller command based on the host memory controller command.

A memory subsystem is provided, including a memory controller integrated circuit (IC), a memory bus and a memory IC, all which use fewer signals than common DDR type memory of the same peak bandwidth. Using no more than 22 switching signals, the subsystem can transfer data over 3000 Megabytes/second across the bus interconnecting the ICs. Signal count reduction is attained by time-multiplexing address/control commands onto at least some of the same signals used for data transfer. A single bus signal is used to initiate bus operation, and once in operation the single signal can transfer addressing and control information to the memory IC concurrent with data transfer via a serial protocol based on 16 bit samples of this single bus signal. Bus bandwidth can be scaled by adding additional data and data strobe IO signals. These additional data bus signals might be used only for data and data mask transport. The physical layout of one version of the memory IC dispatches switching signal terminals adjacent to one short edge of the memory die to minimize the die area overhead for controller IC memory interface circuitry when used in a stacked die multi-chip package with said memory controller IC. The memory IC interface signal placement and signal count minimize signal length and circuitry for the memory bus signals.

According to some example embodiments of the present disclosure, in a method for a memory lookup mechanism in a high-bandwidth memory system, the method includes: using a memory die to conduct a multiplication operation using a lookup table (LUT) methodology by accessing a LUT, which includes floating point operation results, stored on the memory die; sending, by the memory die, a result of the multiplication operation to a logic die including a processor and a buffer; and conducting, by the logic die, a matrix multiplication operation using computation units.

A data placement and migration method for mixed deployment scenarios of heterogeneous memory and multi-type applications is composed of an application static analysis method, a data static placement method and a runtime dynamic migration mechanism. The method is applicable to heterogeneous memory systems consisting of high bandwidth memory and double rate memory and mixed deployment scenarios of asingle delay-dependent application and multiple batch applications, from two angles of static data placement tool and dynamic migration runtime system, the application is divided into logical data object sets by using the dependencies between data objects. Based on the different characteristics of the two memories, different data placement strategies are developed for different types of logical data objects, and dynamic monitoring and real-time migration are carried out during the running process of the application, so as to make full use of various memory resources of heterogeneous memory systems and meet the performance requirements of delay-dependent applications and batch applications respectively.

System and method for performing a high-bandwidth memory copy. Memory transfer instructions allow for copying of data from a first memory location to a second memory location without the use of load and store word instructions thereby achieving a high-bandwidth copy. In one embodiment, the method includes the steps of (1) decoding a destination address from a first memory transfer instruction, (2) storing the destination address in a register in the bus interface unit, (3) decoding a source address from a second memory transfer instruction, and (4) copying the contents of a memory location specified by the source memory address to a memory location specified by the contents of the register. Other methods and a microprocessor system are also presented.

The invention provides a high-bandwidth memory-based neural network calculation apparatus and method. The neural network calculation apparatus comprises at least one high-bandwidth memory and a neuralnetwork accelerator, wherein each high-bandwidth memory comprises multiple memories accumulated in a stacked manner; and the neural network accelerator is electrically connected with the high-bandwidth memory, performs data exchange with the high-bandwidth memory, and executes neural network calculation. According to the apparatus and the method, the storage bandwidth can be greatly increased; the high-bandwidth memory serves as a memory of the neural network calculation apparatus, so that the data exchange of input data and calculation parameters can be performed between a buffer and the memory more quickly, and the IO time is greatly shortened; the high-bandwidth memory adopts a stacked structure and does not occupy a transverse plane space, so that the area of the neural network calculation apparatus can be greatly reduced and can be reduced to about 5% of the area in the prior art; and the power consumption of the neural network calculation apparatus is reduced.

Providing scalable dynamic random access memory (DRAM) cache management using tag directory caches is provided. In one aspect, a DRAM cache management circuit is provided to manage access to a DRAM cache in a high-bandwidth memory. The DRAM cache management circuit comprises a tag directory cache and a tag directory cache directory. The tag directory cache stores tags of frequently accessed cache lines in the DRAM cache, while the tag directory cache directory stores tags for the tag directory cache. The DRAM cache management circuit uses the tag directory cache and the tag directory cache directory to determine whether data associated with a memory address is cached in the DRAM cache of the high-bandwidth memory. Based on the tag directory cache and the tag directory cache directory, the DRAM cache management circuit may determine whether a memory operation can be performed using the DRAM cache and/or a system memory DRAM.

The invention discloses a model training device, method and system and a computer readable storage medium, and the device comprises: a high-bandwidth storage group which is used for storing to-be-trained data; a global memory which is used for storing model parameters of the target model; a processor which is connected with the high-bandwidth memory bank and the global memory and is used for operating the to-be-trained data and the initial model parameters stored in the global memory based on a training algorithm of the target model to obtain target model parameters and storing the target model parameters into the global memory, wherein the processor is established based on the network-on-chip. In the present application, the storage capacity of the high-bandwidth memory is large, so thatmore to-be-trained data can be stored, and the read-write frequency of the high-bandwidth memory is reduced; the processor is built based on the network-on-chip, so that the processor can transmit corresponding data in the processor by means of the network-on-chip, the read-write frequency of the data to be transmitted is reduced, and compared with the prior art, the training efficiency of the model can be improved.