Abstract: Lossy tensor compression and decompression circuits compress and decompress tensor elements based on the values of neighboring tensor elements. The lossy compression circuit scales each decompressed tensor element of a tile by a scaling factor that is based on the maximum value that can be represented by the number of bits used to represent a compressed tensor element, and the greatest value and least value of the tensor elements of the tile. The lossy decompression circuit performs the inverse of the lossy compression. The compression circuit and decompression circuit have parallel multiplier circuits and parallel adder circuits to perform the lossy compression and lossy decompression, respectively.

Abstract: Embodiments herein describe an FFT that can bypass one or more stages when processing smaller frames. For example, when all the stages in the FFT are active, the FFT can process up to a maximum supported point size. However, the particular application may only every send smaller sized frames to the FFT. Instead of unnecessarily passing these frames through the beginning stages of the FFT (which adds latency and consumes power), the embodiments herein can bypass the unneeded stages which reduces the maximum point size the FFT can process but saves power and reduces latency. For example, the FFT can have selection circuitry (e.g., multiplexers) disposed between each stage that permits the input data to either bypass the previous stage(s) or the subsequent stage(s), depending on the architecture of the FFT. The bypassed stages can then be deactivated to conserve power.

Abstract: Apparatus and associated methods relate to providing a modified CORDIC approach and implementing the modified CORDIC approach in SoftMax calculation to reduce usage of hardware resources.

Abstract: Modifying a flow table for a network accelerator can include, in response to determining that a flow table of a network accelerator does not include any rules corresponding to first packet data of a first flow received from a network, forwarding the first packet data to a host computer. Subsequent to the flow table being updated to include a new rule for the first flow, for second packet data of the first flow received from the network, the second packet data can be processed using the new rule. The second packet data can be queued. In response to receiving the first packet data from the host computer, the first packet data can be processed using the new rule. The processed packet data can be forwarded to a destination port followed by the queued second packet data.

Abstract: An example integrated circuit includes an array of circuit tiles; interconnect coupling the circuit tiles in the array, the interconnect including interconnect tiles each having a plurality of connections that include at least a connection to a respective one of the circuit tiles and a connection to at least one other interconnect tile; and a plurality of local crossbars in each of the interconnect tiles, the plurality of local crossbars coupled to form a non-blocking crossbar, each of the plurality of local crossbars including handshaking circuitry for asynchronous communication.

Abstract: A processing circuit includes a random access memory (RAM) configured to look up a first next state based on a first address simultaneously with looking up a second next state based on a second address. The first address is formed of a first current state and an input data and the second address is formed of a second current state and the input data. The processing circuit includes a state control circuit that receives the first and second next states, the first current state, and the second current state, and a first-in-first-out (FIFO) memory that stores selected ones of the first and second next states, the first current state, and the second current state. The processing circuit includes a multiplexer configured to selectively pass two states from the FIFO memory or two states from the state control circuit as a third current state and a fourth current state.

Abstract: An integrated circuit includes a high-speed interface configured to communicate with a host system for debugging and a debug hub coupled to the high-speed interface. The debug hub is configured to receive a debug command from the host system as memory mapped data. The integrated circuit also includes a plurality of debug cores coupled to the debug hub. Each debug core is coupled to the debug hub by channels. The debug hub is configured to translate the debug command to a data stream and provide the data stream to a target debug core of the plurality of debug cores based on an address specified by the debug command.

Abstract: An example hardware accelerator for a computer system includes a programmable device and further includes kernel logic configured in a programmable fabric of the programmable device, and an intellectual property (IP) checker circuit in the kernel logic. The IP checker circuit is configured to obtain a device identifier (ID) of the programmable device and a signed whitelist, the signed whitelist including a list of device IDs and a signature, verify the signature of the signed whitelist, compare the device ID against the list of device IDs, and selectively assert or deassert an enable of the kernel logic in response to presence or absence, respectively, of the device ID in the list of device IDs and verification of the signature.

Abstract: Simulation of a circuit design using accelerated models can include determining, using computer hardware, that a design unit of a circuit design specified in a hardware description language is a prime block and determining, using the computer hardware, an output vector corresponding to an output of the prime block. Using the computer hardware, contents of the prime block can be replaced with an accelerated simulation model specified in a high level language, wherein the accelerated simulation model can determine a value for the output of the prime block as a function of values of one or more inputs of the prime block using the output vector. Using the computer hardware, the circuit design can be elaborated and compiled into object code that is executable to simulate the circuit design.

Abstract: An integrated circuit includes a plurality of data processing engines (DPEs) DPEs. Each DPE may include a core configured to perform computations. A first DPE of the plurality of DPEs includes a first core coupled to an input cascade connection of the first core. The input cascade connection is directly coupled to a plurality of source cores of the plurality of DPEs. The input cascade connection includes a plurality of inputs, wherein each of the plurality of inputs is connected to a cascade output of a different one of the plurality of source cores. The input cascade connection is programmable to enable a selected one of the plurality of inputs.

Abstract: A network interface device has an input configured to receive data from a network. The data is for one of a plurality of different applications. The network interface device also has at least one processor configured to determine which of a plurality of available different caches in a host system the data is to be injected by accessing to a receive queue comprising at least one descriptor indicating a cache location in one of said plurality of caches to which data is to be injected, wherein said at least one descriptor, which indicates the cache location, has an effect on subsequent descriptors of said receive queue until a next descriptor indicates another cache location. The at least one processor is also configured to cause the data to be injected to the cache location in the host system.

Abstract: Embodiments herein describe a hardware based scrubbing scheme where correction logic is integrated with memory elements such that scrubbing is performed by hardware. The correction logic reads the data words stored in the memory element during idle cycles. If a correctable error is detected, the correction logic can then use a subsequent idle cycle to perform a write to correct the error (i.e., replace the corrupted data stored in the memory element with corrected data). By using built-in or integrated correction logic, the embodiments herein do not add extra work for the processor, or can work with applications that do not include a processor. Further, because the correction logic scrubs the memory during idle cycles, correcting bit errors does not have a negative impact on the performance of the memory element. Memory scrubbing can delay the degradation of data error, extending the integrity of the data in the memory.

Abstract: Mining proxy acceleration may include receiving, within a mining proxy, packetized data from a mining pool server and determining, using the mining proxy, whether the packetized data qualifies for broadcast processing. In response to determining that the packetized data qualifies for broadcast processing, the packetized data can be modified using the mining proxy to generate broadcast data. The broadcast data can be broadcast, using the mining proxy, to a plurality of miners subscribed to the mining proxy.

Abstract: Some examples described herein provide for an on-die virtual probe in an integrated circuit structure for measurement of voltages. In an example, an integrated circuit comprises a voltage-controlled frequency oscillator circuitry and a processor circuitry. The voltage-controlled frequency oscillator circuitry comprises a plurality of circuitry components and is configured to generate a signal having a frequency related to a supply voltage. The voltage-controlled frequency oscillator circuitry is disposed at a location of the integrated circuit proximal to the supply voltage being monitored. The processor circuitry is configured to identify a relationship between the frequency of the signal and the supply voltage. The processor circuitry is also configured to determine a value of the supply voltage associated with the signal based on the identified relationship. The processor circuitry further monitors on-die transient voltages at the location of the integrated circuit based on the value of the supply voltage.

Abstract: Implementing a hardware description language (HDL) memory includes determining, using computer hardware, a width and a depth of the HDL memory specified as an HDL module for implementation in an integrated circuit (IC), partitioning, using the computer hardware, the HDL memory into a plurality of super slices corresponding to columns and the plurality of super slices into a plurality of super tiles arranged in rows. A heterogeneous memory array may be generated, using the computer hardware. The heterogeneous memory array is formed of different types of memory primitives of the IC. Input and output circuitry configured to access the heterogeneous memory array can be generated using the computer hardware.

Abstract: Systems and methods for designing an information processing system are described. In one embodiment, a design space is partitioned into a plurality of independent partitions based on a defined set of rules. A unique processing core is assigned to each partition. A plurality of starting points is generated for each partition, where each starting point is associated with a machine learning algorithm. The starting points for each partition may include a performance driven seed and an area-driven seed. A set of feasible designs associated with the information processing system are determined.

Abstract: In disclosed approaches of neural network processing, a host computer system copies an input data matrix from host memory to a shared memory for performing neural network operations of a first layer of a neural network by a neural network accelerator. The host instructs the neural network accelerator to perform neural network operations of each layer of the neural network beginning with the input data matrix. The neural network accelerator performs neural network operations of each layer in response to the instruction from the host. The host waits until the neural network accelerator signals completion of performing neural network operations of layer i before instructing the neural network accelerator to commence performing neural network operations of layer i+1, for i?1. The host instructs the neural network accelerator to use a results data matrix in the shared memory from layer i as an input data matrix for layer i+1 for i?1.

Abstract: A circuit arrangement includes an array of MAC circuits, wherein each MAC circuit includes a cache configured for storage of a plurality of kernels. The MAC circuits are configured to receive a first set of data elements of an IFM at a first rate. The MAC circuits are configured to perform first MAC operations on the first set of the data elements and a first one of the kernels associated with a first OFM depth index during a first MAC cycle, wherein a rate of MAC cycles is faster than the first rate. The MAC circuits are configured to perform second MAC operations on the first set of the data elements and a second one of the kernels associated with a second OFM depth index during a second MAC cycle that consecutively follows the first MAC cycle.

Abstract: Disclosed circuits and methods involve a first register configured to store of a first convolutional neural network (CNN) instruction during processing of the first CNN instruction and a second register configured to store a second CNN instruction during processing of the second CNN instruction. Each of a plurality of address generation circuits is configured to generate one or more addresses in response to an input CNN instruction. Control circuitry is configured to select one of the first CNN instruction or the second CNN instruction as input to the address generation circuits.

Abstract: Some examples described herein relate to multi-chip devices. In an example, a multi-chip device includes first and second chips. The first chip includes a power supply circuit and a logic circuit. The first and second chips are coupled together. The second chip is configured to receive power from the power supply circuit. The second chip includes a programmable circuit, a pull-up circuit, and a detector circuit. The detector circuit is configured to detect a presence of a power voltage on the second chip and responsively output a presence signal. The power voltage on the second chip is based on the power from the power supply circuit. The logic circuit is configured to generate a pull-up signal based on the presence signal. The pull-up circuit is configured to receive the pull-up signal and configured to pull up a voltage of a node of the programmable circuit responsive to the pull-up signal.