Stmicroelectronics S.r.l.

Stmicroelectronics International N.v.

Hewlett-packard Development Company, L.P.

Stmicroelectronics (grenoble 2) Sas

Stmicroelectronics (rousset) Sas

Giuseppe Desoli

Thomas Boesch

Surinder Pal Singh

Nitin Chawla

Anuj Grover

Carmine Cappetta

Manuj Ayodhyawasi

Paolo Faraboschi

Kedar Janardan Dhori

Promod Kumar

Tanmoy Roy

Paolo Zambotti

Saumya Suneja

Michele Rossi

Pasqualina Fragneto

Elio Guidetti

Gilles Pelissier

Vasanth Bala

Jean-Philippe Cousin

David Siorpaes

Beatrice Rossi

Fabio Giuseppe De Ambroggi

Giacomo Boracchi

Diego Carrera

Badr Bentaybi

Francesco Papariello

Joseph A Fisher

Wolfgang Johann Betz

Valentina Arrigoni

Ugo Maria Iannuzzi

Tommaso Majo

Antonio De Vita

Paolo Sergio Zambotti

A hardware accelerator includes a plurality of functional circuits, a stream switch, and a plurality of neural network processing cores coupled to the plurality of functional circuits via the stream switch to stream data to and from functional circuits of the plurality of functional circuits, wherein the neural network processing cores include at least one sender core having a buffer whose buffer content is sharable with at least one receiver core of the neural network processing cores via at least one of a dedicated output stream link or dedicated virtual channel on a pre-existing output stream link, and wherein the at least one of the dedicated output stream link or dedicated virtual channel is dedicated to sharing the buffer content via the stream switch.

Adaptive buffer sharing in multi-core reconfigurable streaming-based architectures

A system includes a host processor, a memory, a hardware accelerator and a configuration controller. The host processor, in operation, controls execution of a multi-stage processing task. The memory, in operation, stores data and configuration information. The hardware accelerator, in operation preforms operations associated with stages of the multi-stage processing task. The configuration controller is coupled to the host processor, the hardware accelerator, and the memory. The configuration controller executes a linked list of configuration operations, for example, under control of a finite state machine. The linked list consists of configuration operations selected from a defined set of configuration operations. Executing the linked list of configuration operations configures the plurality of configuration registers of the hardware accelerator to control operations of the hardware accelerator associated with a stage of the multi-stage processing task. The configuration controller may retrieve the linked list from the memory via a high-speed data bus.

Programmable hardware accelerator controller

A memory array arranged as a plurality of memory cells. The memory cells are configured to operate at a determined voltage. A memory management circuitry coupled to the plurality of memory cells tags a first set of the plurality of memory cells as low-voltage cells and tags a second set of the plurality of memory cells as high-voltage cells. A power source provides a low voltage to the first set of memory cells and provides a high voltage to the second set of memory cells based on the tags.

Tagged memory operated at lower VMIN in error tolerant system

A convolutional neural network includes convolution circuitry. The convolution circuitry performs convolution operations on input tensor values. The convolutional neural network includes requantization circuitry that requantizes convolution values output from the convolution circuitry.

Neural network hardware accelerator circuit with requantization circuits

Systems and devices are provided to enable granular control over a retention or active state of each of a plurality of memory circuits, such as a plurality of memory cell arrays, within a memory. Each respective memory array of the plurality of memory arrays is coupled to a respective ballast driver and a respective active memory signal switch for the respective memory array. One or more voltage regulators are coupled to a ballast driver gate node and to a bias node of at least one of the respective memory arrays. In operation, the respective active memory signal switch for a respective memory array causes the respective memory array to transition between an active state for the respective memory array and a retention state for the respective memory array.

Computing system power management device, system and method

An integrated circuit includes a reconfigurable stream switch and an arithmetic circuit. The stream switch, in operation, streams data. The arithmetic circuit has a plurality of inputs coupled to the reconfigurable stream switch. In operation, the arithmetic circuit generates an output according to AX+BY+C, where A, B and C are vector or scalar constants, and X and Y are data streams streamed to the arithmetic circuit through the reconfigurable stream switch.

Arithmetic unit for deep learning acceleration

Embodiments are directed towards a system on chip (SoC) that implements a deep convolutional network heterogeneous architecture. The SoC includes a system bus, a plurality of addressable memory arrays coupled to the system bus, at least one applications processor core coupled to the system bus, and a configurable accelerator framework coupled to the system bus. The configurable accelerator framework is an image and deep convolutional neural network (DCNN) co-processing system. The SoC also includes a plurality of digital signal processors (DSPs) coupled to the system bus, wherein the plurality of DSPs coordinate functionality with the configurable accelerator framework to execute the DCNN.

Deep convolutional network heterogeneous architecture

A convolutional accelerator framework (CAF) has a plurality of processing circuits including one or more convolution accelerators, a reconfigurable hardware buffer configurable to store data of a variable number of input data channels, and a stream switch coupled to the plurality of processing circuits. The reconfigurable hardware buffer has a memory and control circuitry. A number of the variable number of input data channels is associated with an execution epoch. The stream switch streams data of the variable number of input data channels between processing circuits of the plurality of processing circuits and the reconfigurable hardware buffer during processing of the execution epoch. The control circuitry of the reconfigurable hardware buffer configures the memory to store data of the variable number of input data channels, the configuring including allocating a portion of the memory to each of the variable number of input data channels.

Reconfigurable hardware buffer in a neural networks accelerator framework

Embodiments are directed towards a hardware accelerator engine that supports efficient mapping of convolutional stages of deep neural network algorithms. The hardware accelerator engine includes a plurality of convolution accelerators, and each one of the plurality of convolution accelerators includes a kernel buffer, a feature line buffer, and a plurality of multiply-accumulate (MAC) units. The MAC units are arranged to multiply and accumulate data received from both the kernel buffer and the feature line buffer. The hardware accelerator engine also includes at least one input bus coupled to an output bus port of a stream switch, at least one output bus coupled to an input bus port of the stream switch, or at least one input bus and at least one output bus hard wired to respective output bus and input bus ports of the stream switch.

Hardware accelerator engine

A device include on-board memory, an applications processor, a digital signal processor (DSP) cluster, a configurable accelerator framework (CAF), and a communication bus architecture. The communication bus communicatively couples the applications processor, the DSP cluster, and the CAF to the on-board memory. The CAF includes a reconfigurable stream switch and data volume sculpting circuitry, which has an input and an output coupled to the reconfigurable stream switch. The data volume sculpting circuitry receives a series of frames, each frame formed as a two dimensional (2D) data structure, and determines a first dimension and a second dimension of each frame of the series of frames. Based on the first and second dimensions, the data volume sculpting circuitry determines for each frame a position and a size of a region-of-interest to be extracted from the respective frame, and extracts from each frame, data in the frame that is within the region-of-interest.

Data volume sculptor for deep learning acceleration

Systems and devices are provided to increase computational and/or power efficiency for one or more neural networks via a computationally driven closed-loop dynamic clock control. A clock frequency control word is generated based on information indicative of a current frame execution rate of a processing task of the neural network and a reference clock signal. A clock generator generates the clock signal of neural network based on the clock frequency control word. A reference frequency may be used to generate the clock frequency control word, and the reference frequency may be based on information indicative of a sparsity of data of a training frame.

Variable clock adaptation in neural network processors

Embodiments of an electronic device include an integrated circuit, a reconfigurable stream switch formed in the integrated circuit along with a plurality of convolution accelerators and a decompression unit coupled to the reconfigurable stream switch. The decompression unit decompresses encoded kernel data in real time during operation of convolutional neural network.

Vector quantization decoding hardware unit for real-time dynamic decompression for parameters of neural networks

Techniques and systems are provided for implementing a convolutional neural network. One or more convolution accelerators are provided that each include a feature line buffer memory, a kernel buffer memory, and a plurality of multiply-accumulate (MAC) circuits arranged to multiply and accumulate data. In a first operational mode the convolutional accelerator stores feature data in the feature line buffer memory and stores kernel data in the kernel data buffer memory. In a second mode of operation, the convolutional accelerator stores kernel decompression tables in the feature line buffer memory.

Convolution acceleration with embedded vector decompression

Tagged memory operated at lower vmin in error tolerant system

A memory array arranged in multiple columns and rows. Computation circuits that each calculate a computation value from cell values in a corresponding column. A column multiplexer cycles through multiple data lines that each corresponds to a computation circuit. Cluster cycle management circuitry determines a number of multiplexer cycles based on a number of columns storing data of a compute cluster. A sensing circuit obtains the computation values from the computation circuits via the column multiplexer as the column multiplexer cycles through the data lines. The sensing circuit combines the obtained computation values over the determined number of multiplexer cycles. A first clock may initiate the multiplexer to cycle through its data lines for the determined number of multiplexer cycles, and a second clock may initiate each individual cycle. The multiplexer or additional circuitry may be utilized to modify the order in which data is written to the columns.

Growing community of inventors

San Fermo Della Battaglia, Italy

Giuseppe Desoli