51results about "Systolic arrays" patented technology

Systems and methods for automated systolic array design from a high-level program are disclosed. One implementation of a systolic array design supporting a convolutional neural network includes a two-dimensional array of reconfigurable processing elements arranged in rows and columns. Each processing element has an associated SIMD vector and is connected through a local connection to at least one other processing element. An input feature map buffer having a double buffer is configured to store input feature maps, and an interconnect system is configured to pass data to neighboring processing elements in accordance with a processing element scheduler. A CNN computation is mapped onto the two-dimensional array of reconfigurable processing elements using an automated system configured to determine suitable reconfigurable processing element parameters.

A processor for use in a router, the processor having a systolic array pipeline for processing data packets to determine to which output port of the router the data packet should be routed. In one embodiment, the systolic array pipeline includes a plurality of programmable functional units and register files arranged sequentially as stages, for processing packet contexts (which contain the packet's destination address) to perform operations, under programmatic control, to determine the destination port of the router for the packet. A single stage of the systolic array may contain a register file and one or more functional units such as adders, shifters, logical units, etc., for performing, in one example, very long instruction word (vliw) operations. The processor may also include a forwarding table memory, on-chip, for storing routing information, and a cross bar selectively connecting the stages of the systolic array with the forwarding table memory.

A data stream processing unit (DPU) and method for use are provided. A DPU includes a number of processing elements arranged in a sequence, and each datum in the data stream visits each processing element in sequence. Each processing element has a memory circuit, data and metadata input and output channels, and a computing circuit. The metadata input represents a partial computational state that is associated with each datum as it passes through the DPU. The computing circuit for each processing element operates on the data and metadata inputs as a function of its position in the sequence, producing an altered partial computational state that accompanies the datum. Each computing circuit may be modeled, for example, as a finite state machine, and the collection of processing elements cooperate to perform the computation. The computing circuits may be collectively programmed to perform any desired computation.

Systems, apparatuses, methods, and software for processing data in pipeline architectures are provided herein. In one example, a pipeline architecture is presented. The pipeline architecture includes a plurality of processing stages, linked in series, that iteratively process data as the data propagates through the plurality of processing stages. The pipeline architecture includes at least one other processing stage linked in series with and preceded by the plurality of processing stages and configured to iteratively process the data a number of times based at least on an iteration count comprising how many times the data was iteratively processed as the data propagated through the plurality of processing stages.

A processor having a systolic array that can perform operations efficiently is provided. The processor includes multiple processing cores aligned in a matrix, and each of the processing cores includes an arithmetic unit array including multiple arithmetic units that can form a systolic array. Each of the processing cores includes a first memory that stores first data, a second memory that stores second data, a first multiplexer that connects a first input for receiving the first data at the arithmetic unit array to an output of the first memory in the processing core or an output of the arithmetic unit array in an adjacent processing core, and a second multiplexer that connects a second input for receiving the second data at the arithmetic unit array to an output of the second memory in the processing core or an output of the arithmetic unit array in an adjacent processing core.

A mechanism synchronizes instruction code executing on a processor of a processing engine in an intermediate network station. The processing engine is configured as a systolic array having a plurality of processors arrayed as rows and columns. The mechanism comprises a boundary (temporal) synchronization mechanism for cycle-based synchronization within a processor of the array. The synchronization mechanism is generally implemented using specialized synchronization micro operation codes (“opcodes”).

A linear systolic array is added to the lower side of a trapezoid systolic array created by combining a triangular systolic array and a square systolic array. In order to make the connection among the cells fixed, the intermediate result output from each row of the trapezoid systolic array to a lower row is shifted in phase with respect to the intermediate result of the complex MFA algorithm, the phase shift is absorbed by the next row, and the phase shift in the intermediate result output from the last row of the trapezoid systolic array is corrected by the linear systolic array. Each cell is implemented by a CORDIC circuit that processes vector angle computation, vector rotation, division, and multiply-and-accumulate with a constant delay.

In an array processing section, using data strings entered from input ports, a plurality of data processor elements execute predetermined operations while transferring data to each other, and output data strings of results of the operations from a plurality of output ports. A first data string converter converts data strings stored in a plurality of data storages of a data storage group into a placement suitable for the operations in the array processing section, and enters the converted data strings into the input ports of the array processing section. A second data string converter converts the data strings output from output ports of the array processing section into a placement to be stored in the plurality of data storages of the data storage group.

A calculation circuit may include a plurality of calculator groups constituting a systolic array composed of a plurality of rows and columns, wherein calculator groups included in each of the rows propagate a data value set through a single data path corresponding to the row in a data propagation direction, and propagate a plurality of drain value sets through a plurality of drain paths corresponding to the row in a drain propagation direction, and wherein a calculator group of the calculator groups included in each of the rows comprises a plurality of MAC (Multiplier-Accumulator) circuits, and the MAC circuits generate drain values respectively included in the drain value sets at the same time. The calculator groups included in each column may further propagate a weight value set corresponding to the column through a plurality of weight data paths corresponding to the column.

A tensor computation dataflow accelerator semiconductor circuit is disclosed. The data flow accelerator includes a DRAM bank and a peripheral array of multiply-and-add units disposed adjacent to the DRAM bank. The peripheral array of multiply-and-add units are configured to form a pipelined dataflow chain in which partial output data from one multiply-and-add unit from among the array of multiply-and-add units is fed into another multiply-and-add unit from among the array of multiply-and-add units for data accumulation. Near-DRAM-processing dataflow (NDP-DF) accelerator unit dies may be stacked atop a base die. The base die may be disposed on a passive silicon interposer adjacent to a processor or a controller. The NDP-DF accelerator units may process partial matrix output data in parallel. The partial matrix output data may be propagated in a forward or backward direction. The tensor computation dataflow accelerator may perform a partial matrix transposition.

Aspects of the disclosure are directed to a computation unit implementing a systolic array and configured for detecting errors while processing data on the systolic array. Checksum circuit in communication with a systolic array is configured to compute checksums and perform error detection while the systolic array processes input data. Instead of pre-generating checksums in input matrices, input matrices can be directly fed into the systolic array through the checksum circuit. On the output side, the checksum circuit can generate and compare checksums with checksums in an output matrix generated by the systolic array. Error checking the operations to generate the output matrix can be performed without delaying the operations of the systolic array, and without preprocessing the input matrices.

A data bus includes process elements and a linear main pipeline. Each process element is coupled to a linear pipeline having M stages arranged in series, each of the M stages including a buffer element configured to buffer a data bit sequence and to forward the buffered data bit sequence from a first of the buffer elements to a last of the buffer elements. The linear main pipeline includes N pipeline stage elements arranged in series. Each pipeline stage element is connected to the last buffer element of a respective linear pipeline and configured to read-out one or more of the buffered data bit sequences and to forward the read-out data bit sequences from one of N pipeline stag elements to a next of the N pipeline stage elements.

The invention provides a calculation optimization method and device for a shallow depth model based on a systolic array. By determining the standard input depth/standard output depth of the systolic array and the input depth and the output depth of the shallow depth model, determining the input splicing number and the output splicing number. And according to the input splicing number and the output splicing number value, splicing the input data and/or the corresponding output data so as to obtain the input data in batches and provide the input data to the systolic array in parallel, and/or caching a plurality of output data corresponding to the input data through an output buffer and then outputting the output data in batches, therefore, the migration efficiency of the input data and/or the output data can be improved, and the utilization efficiency of the systolic array can also be improved.

Disclosed embodiments relate to deep learning implementations using systolic arrays and fused operations. In one example, a processor includes fetch and decode circuitry to fetch and decode an instruction having fields to specify an opcode and locations of a destination and N source matrices, the opcode indicating the processor is to load the N source matrices from memory, perform N convolutions on the N source matrices to generate N feature maps, and store results of the N convolutions in registers to be passed to an activation layer, wherein the processor is to perform the N convolutions andthe activation layer with at most one memory load of each of the N source matrices. The processor further includes scheduling circuitry to schedule execution of the instruction and execution circuitry to execute the instruction as per the opcode.