Abstract: Methods, systems, and apparatus for a matrix multiply unit implemented as a systolic array of cells are disclosed. Each cell of the matrix multiply includes: a weight matrix register configured to receive a weight input from either a transposed or a non-transposed weight shift register; a transposed weight shift register configured to receive a weight input from a horizontal direction to be stored in the weight matrix register; a non-transposed weight shift register configured to receive a weight input from a vertical direction to be stored in the weight matrix register; and a multiply unit that is coupled to the weight matrix register and configured to multiply the weight input of the weight matrix register with a vector data input in order to obtain a multiplication result.

Abstract: Methods, systems, and apparatus for a matrix multiply unit implemented as a systolic array of cells are disclosed. The matrix multiply unit may include cells arranged in columns of the systolic array. Two chains of weight shift registers per column of the systolic array are in the matrix multiply unit. Each weight shift register is connected to only one chain and each cell is connected to only one weight shift register. A weight matrix register per cell is configured to store a weight input received from a weight shift register. A multiply unit is coupled to the weight matrix register and configured to multiply the weight input of the weight matrix register with a vector data input in order to obtain a multiplication result.

Abstract: A memory system includes a plurality of memory nodes provided at different hierarchical levels of the memory system, each of the memory nodes including a corresponding memory storage and a cache. A memory node at a first of the different hierarchical levels is coupled to a processor with lower communication latency than a memory node at a second of the different hierarchical levels. The memory nodes are to cooperate to decide which of the memory nodes is to cache data of a given one of the memory nodes.

Abstract: A vector processing unit is described, and includes processor units that each include multiple processing resources. The processor units are each configured to perform arithmetic operations associated with vectorized computations. The vector processing unit includes a vector memory in data communication with each of the processor units and their respective processing resources. The vector memory includes memory banks configured to store data used by each of the processor units to perform the arithmetic operations. The processor units and the vector memory are tightly coupled within an area of the vector processing unit such that data communications are exchanged at a high bandwidth based on the placement of respective processor units relative to one another, and based on the placement of the vector memory relative to each processor unit.

Abstract: Methods, systems, and apparatus, including instructions encoded on storage media, for performing reduction of gradient vectors and similarly structured data that are generated in parallel, for example, on nodes organized in a mesh or torus topology defined by connections in at least two dimension between the nodes. The methods provide parallel computation and communication between nodes in the topology.

Abstract: Methods, systems, and apparatus, including instructions encoded on storage media, for performing reduction of gradient vectors and similarly structured data that are generated in parallel, for example, on nodes organized in a mesh or torus topology defined by connections in at least two dimension between the nodes. The methods provide parallel computation and communication between nodes in the topology.

Abstract: A semiconductor package includes a substrate, an integrated circuit disposed on the substrate, a memory support disposed on the integrated circuit, stacked memory disposed on the memory support and in communication with the integrated circuit, and a lid connected to the substrate. The integrated circuit has a low power region and a high power region. The memory support is disposed on the low power region of the integrated circuit and is configured to allow a flow of fluid therethrough to conduct heat away from the low power region of the integrated circuit. The lid defines a first port, a second port, and a lid volume fluidly connecting the first port and the second port. The lid volume is configured to house the integrated circuit, the memory support, and the stacked memory, while directing the flow of fluid to flow over the integrated circuit, the memory support, and the stacked memory.

Abstract: A detector detects, using an error code, an error in data stored in a memory. The detector determines whether the error is uncorrectable using the error code. In response to determining that the error is uncorrectable, an error handler associated with an application is invoked to handle the error in the data by recovering the data to an application-wide consistent state.

Abstract: A circuit for performing neural network computations for a neural network comprising a plurality of neural network layers, the circuit comprising: a matrix computation unit configured to, for each of the plurality of neural network layers: receive a plurality of weight inputs and a plurality of activation inputs for the neural network layer, and generate a plurality of accumulated values based on the plurality of weight inputs and the plurality of activation inputs; and a vector computation unit communicatively coupled to the matrix computation unit and configured to, for each of the plurality of neural network layers: apply an activation function to each accumulated value generated by the matrix computation unit to generate a plurality of activated values for the neural network layer.

Abstract: A non-volatile multi-level cell (“MLC”) memory device is disclosed. The memory device has an array of non-volatile memory cells, an array of non-volatile memory cells, with each non-volatile memory cell storing multiple groups of bits. A row buffer in the memory device has multiple buffer portions, each buffer portion storing one or more bits from the memory cells and having different read and write latencies and energies.

Abstract: A disclosed example apparatus includes a row address register (412) to store a row address corresponding to a row (608) in a memory array (602). The example apparatus also includes a row decoder (604) coupled to the row address register to assert a signal on a wordline (704) of the row after the memory receives a column address. In addition, the example apparatus includes a column decoder (606) to selectively activate a portion of the row based on the column address and the signal asserted on the wordline.

Abstract: A system can include an optical multiplexer to combine a plurality of optical input signals having respective wavelengths into a wide-channel optical input signal that is provided to an input channel. The system also includes a photonic packet switch comprising a switch core and a plurality of ports defining a switch radix of the photonic packet switch. The input channel and an output channel can be associated with one of the plurality of ports. The photonic packet switch can process the wide-channel optical input signal and can generate a wide-channel optical output signal that is provided to the output channel. The system further includes an optical demultiplexer to separate the wide-channel optical output signal into a plurality of optical output signals having respective wavelengths. The optical multiplexer and the optical demultiplexer can collectively provide the system with a radix greater than the switch radix.

Abstract: According to an example, a resiliency group for a memory node in a memory network can provide error correction for a memory access in the memory node. The memory access may be received from a main memory controller of a processor connected to the memory network. The memory access may be executed by a memory controller of the memory node.

Abstract: A circuit for performing neural network computations for a neural network comprising a plurality of neural network layers, the circuit comprising: a matrix computation unit configured to, for each of the plurality of neural network layers: receive a plurality of weight inputs and a plurality of activation inputs for the neural network layer, and generate a plurality of accumulated values based on the plurality of weight inputs and the plurality of activation inputs; and a vector computation unit communicatively coupled to the matrix computation unit and configured to, for each of the plurality of neural network layers: apply an activation function to each accumulated value generated by the matrix computation unit to generate a plurality of activated values for the neural network layer.

Abstract: A semiconductor package includes a substrate, an integrated circuit disposed on the substrate, a memory support disposed on the integrated circuit, stacked memory disposed on the memory support and in communication with the integrated circuit, and a lid connected to the substrate. The integrated circuit has a low power region and a high power region. The memory support is disposed on the low power region of the integrated circuit and is configured to allow a flow of fluid therethrough to conduct heat away from the low power region of the integrated circuit. The lid defines a first port, a second port, and a lid volume fluidly connecting the first port and the second port. The lid volume is configured to house the integrated circuit, the memory support, and the stacked memory, while directing the flow of fluid to flow over the integrated circuit, the memory support, and the stacked memory.

Abstract: A circuit for performing neural network computations for a neural network comprising a plurality of neural network layers, the circuit comprising: a matrix computation unit configured to, for each of the plurality of neural network layers: receive a plurality of weight inputs and a plurality of activation inputs for the neural network layer, and generate a plurality of accumulated values based on the plurality of weight inputs and the plurality of activation inputs; and a vector computation unit communicatively coupled to the matrix computation unit and configured to, for each of the plurality of neural network layers: apply an activation function to each accumulated value generated by the matrix computation unit to generate a plurality of activated values for the neural network layer.

Abstract: A circuit for performing neural network computations for a neural network comprising a plurality of neural network layers, the circuit comprising: a matrix computation unit configured to, for each of the plurality of neural network layers: receive a plurality of weight inputs and a plurality of activation inputs for the neural network layer, and generate a plurality of accumulated values based on the plurality of weight inputs and the plurality of activation inputs; and a vector computation unit communicatively coupled to the matrix computation unit and configured to, for each of the plurality of neural network layers: apply an activation function to each accumulated value generated by the matrix computation unit to generate a plurality of activated values for the neural network layer.

Abstract: According to an example, a method for adaptive-granularity row buffer (AG-RB) caching may include determining whether to cache data to a RB cache, and adjusting, by a processor or a memory side logic, an amount of the data to cache to the RB cache for different memory accesses, such as dynamic random-access memory (DRAM) accesses. According to another example, an AG-RB cache apparatus may include a 3D stacked DRAM including a plurality of DRAM dies including one or more DRAM banks, and a logic die including a RB cache. The AG-RB cache apparatus may further include a processor die including a memory controller including a predictor module to determine whether to cache data to the RB cache, and to adjust an amount of the data to cache to the RB cache for different DRAM accesses.

Abstract: An example system in accordance with an aspect of the present disclosure is to use local error detection (LED) and global error correction (GEC) information to check data fidelity and correct an error. The LED is to be calculated per cache line segment of data associated with a rank of a memory. Data fidelity may be checked in response to a memory read operation, based on the LED information, to identify a presence of an error and the location of the error among cache line segments of the rank. The cache line segment having the error may be corrected based on the GEC information, in response to identifying the error.

Abstract: A power management module can select one of a plurality of different operational modes for a hardware component in a computer system based on application performance and total computer system power consumption determined for each of the operational modes.