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 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.

Examples herein describe techniques for adapting a multiplier array (e.g., a systolic array implemented in a processing core) to perform different dot products. The processing core can include data selection logic that enables different configurations of the multiplier array in the core. For example, the data selection logic can enable different configurations of the multiplier array while using the same underlying hardware. That is, the multiplier array is fixed hardware but the data selection can transmit data into the matrix multiplier such that it is configured to perform different length dot products, perform more dot products in parallel, or change its output precision. In this manner, the same underlying hardware (i.e., the multiplier array) can be reconfigured for different dot products which can result in much more efficient use of the hardware.

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.

An embodiment of an apparatus comprises one or more fractional width fused multiply-accumulate (FMA) circuits configured as a shared Wallace tree, and circuitry coupled to the one or more fractional width FMA circuits to provide one or more fractional width FMA operations through the one or more fractional width FMA circuits. Other embodiments are disclosed and claimed.

An accelerator device includes a first processing unit to access a structure of a graph dataset, and a second processing unit coupled with the first processing unit to perform computations based on data values in the graph dataset.

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.

The present disclosure relates to an accelerator for systolic array-friendly data placement. The accelerator may include: a systolic array comprising a plurality of operation units, wherein the systolic array is configured to receive staged input data and perform operations using the staged input to generate staged output data, the staged output data comprising a number of segments; a controller configured to execute one or more instructions to generate a pattern generation signal; a data mask generator; and a memory configured to store the staged output data using the generated masks. The data mask generator may include circuitry configured to: receive the pattern generation signal from the controller, and, based on the received signal, generate a mask corresponding to each segment of the staged output data.

A direct memory access (DMA) engine can be used to multicast data from system memory to a target memory for loading into an array. The DMA engine may include a controller that is configured to receive a data transfer request, and generate a set of write operations for the output interface. The set of write operations can include, for each of multiple partitions of the target memory, a write operation to write usable data from the multicast data to an address offset in the corresponding partition, and an additional write operation to write filler data from the multicast data to a null device address.

In one example, a neural network accelerator executes instructions to: load a first weight data element of an array of weight data elements from a memory into a systolic array; extract, from the instructions, information indicating a first number of input data elements to be obtained from a first address of the memory and a second number of input data elements to be skipped between adjacent input data elements to be obtained, the first address being based on first coordinates of the first weight data element, and the first and second numbers being based on a stride of a convolution operation; based on the information, obtain first input data elements from the first address of the memory; and control the systolic array to perform first computations based on the first weight data element and the first input data elements to generate first output data elements of an output data array.