Abstract: A programmable hardware architecture for machine learning (ML) is proposed, which includes at least a host, a memory, a core, a data streaming engine, a instruction-streaming engine, and an interference engine. The core interprets a plurality of ML commands for a ML operation and/or data received from the host and coordinate activities of the engines based on the data in the received ML commands. The instruction-streaming engine translates the ML commands received from the core and provide a set of programming instructions to the data streaming engine and the inference engines based on the translated parameters. The data steaming engine sends one or more data streams to the inference engine in response to the received programming instructions. The inference engine then processes the data streams received from the data stream engine according to the programming instructions received from the instruction-streaming engine.

Abstract: An apparatus comprises processing circuitry to perform data processing in response to instructions; and a branch predictor to predict a branch outcome for a given branch instruction as one of taken and not-taken, based on branch prediction state information indexed based on at least one property of the given branch instruction. In a static branch prediction mode of operation, the branch predictor predicts the branch outcome based on static values of the branch prediction state information set independent of actual branch outcomes of branch instructions which are executed by the processing circuitry while in the static branch prediction mode. The static values of the branch prediction state information are programmable.

Abstract: An indication of a capacity of a CMB elasticity buffer and an indication of a throughput of one or more memory components associated with the CMB elasticity buffer can be received. An amount of time for data at the CMB elasticity buffer to be transmitted to one or more memory components can be determined based on the capacity of the CMB elasticity buffer and the throughput of the one or more memory components. Write data can be transmitted from a host system to the CMB elasticity buffer based on the determined amount of time for data at the CMB elasticity buffer to be transmitted to the one or more memory components.

Abstract: Saving and restoring machine state between multiple executions of an instruction. A determination is made that processing of an operation of an instruction executing on a processor has been interrupted prior to completion. Based on determining that the processing of the operation has been interrupted, current metadata of the processor is extracted. The metadata is stored in a location associated with the instruction and used to re-execute the instruction to resume forward processing of the instruction from where it was interrupted.

Abstract: Techniques are disclosed for data manipulation within a reconfigurable computing environment for data flow graph computation using exceptions. Processing elements are configured within a reconfigurable fabric to implement a data flow graph. The processing elements are loaded with process agents. Valid data is executed by a first process agent on a first processing element, where the first process agent corresponds to a starting node of the data flow graph. A second processing element detects that an error exception has occurred, where a second process agent is running on the second processing element. A done signal to a third process agent is withheld by the second process agent, where the third process agent is running on a third processing element. The second process agent raises an interrupt request, where the interrupt request is based on the detecting that an error exception has occurred.

Abstract: Apparatuses, storage media and methods associated with data transfer, are disclosed herein. In some embodiments, an apparatus for computing comprises: a commit generator and a media write generator. The commit generator is arranged to generate commit indicators correspondingly associated with media slices of a storage media to respectively denote to a storage media controller of the storage media whether to proceed with writing the media slices into the storage media. The media write generator is arranged provide data chunks of the media slices to be written into the storage media, and the associated commit indicators to the storage media controller. A size of each data chunk is smaller than a size of each media slice. Other embodiments are also described and claimed.

Abstract: A first operation identifier is assigned to a current operation directed to a memory component, the first operation identifier having a first entry in a first data structure that associates the first operation identifier with a first buffer identifier. It is determined whether the current operation collides with a prior operation assigned a second operation identifier, the second operation identifier having a second entry in the first data structure that associates the second operation identifier with a second buffer identifier. A latest flag is updated to indicate that the first entry is a latest operation directed to an address (1) in response to determining that the current operation collides with the prior operation and that the current and prior operations are read operations, or (2) in response to determining to determining that the current operation does not collide with a prior operation.

Abstract: According to one or more embodiments, an example computer-implemented method for executing one or more out-of-order instructions by a processing unit, includes decoding an instruction to be executed, and based on a determination that the instruction is a store instruction, identifying a split load-hit-store (LHS) table for the store instruction, wherein a LHS table of the processing unit includes multiple split LHS tables. Identifying the split LHS table includes determining, for the store instruction, a first split LHS table by performing a mod operation using one or more operands from the store instruction, and adding one or more parameters of the store instruction in the first split LHS table by generating an ITAG for the store instruction. The method further includes dispatching the store instruction for execution to an issue queue with the ITAG.

Abstract: Technologies for adaptive processing of multiple buffers is disclosed. A compute device may establish a buffer queue to which applications can submit buffers to be processed, such as by hashing the submitted buffers. The compute device monitors the buffer queue and determines an efficient way of processing the buffer queue based on the number of buffers present. The compute device may process the buffers serially with a single processor core of the compute device or may process the buffers in parallel with single-instruction, multiple data (SIMD) instructions. The compute device may determine which method to use based on a comparison of the throughput of serially processing the buffers as compared to parallel processing the buffers, which may depend on the number of buffers in the buffer queue.

Abstract: A digital signal processor having at least one streaming address generator, each with dedicated hardware, for generating addresses for writing multi-dimensional streaming data that comprises a plurality of elements. Each at least one streaming address generator is configured to generate a plurality of offsets to address the streaming data, and each of the plurality of offsets corresponds to a respective one of the plurality of elements. The address of each of the plurality of elements is the respective one of the plurality of offsets combined with a base address.

Abstract: A data storage device can employ a front end bus to optimize data storage performance. A first controller may be connected to a first memory via a first bus and to a second memory via a second bus with the first bus and first memory housed within an internal cavity of an enclosure while the second bus is exposed to an exterior surface of the housing and the second memory is separated from the internal cavity. The first controller can be configured to substitute the second memory for the first memory in response to a front end controller identifying a type of data storage of the second memory.

Abstract: Methods and apparatuses relating to a hashing accelerator having a frequency scaled message scheduler data path circuit are described.

Abstract: A processing unit to support inference acceleration for machine learning (ML) comprises an inline post processing unit configured to accept and maintain one or more lookup tables for performing each of one or more non-linear mathematical operations. The inline post processing unit is further configured to accept data from a set of registers maintaining output from a processing block instead of streaming the data from an on-chip memory (OCM), perform the one or more non-linear mathematical operations on elements of the data from the processing block via their corresponding lookup tables, and stream post processing result of the one or more non-linear mathematical operations back to the OCM after the one or more non-linear mathematical operations are complete.

Abstract: Systems and methods for providing audio streams over peripheral component interconnect (PCI) express (PCIE) links are disclosed. In particular, exemplary aspects of the present disclosure are used to calculate an uplink timing requirement and adjust a margin time before a modem encodes audio data so that the encoding is done before data is transmitted to an external network. Further aspects of the present disclosure allow a first integrated circuit (IC) to synchronize its clock with that of the modem.

Abstract: A method and apparatus that provide a solid state drive controller configured to analyze input/output commands from a host computing device to determine whether the commands include a vector or non-vector command, and is configured to generate a plurality of non-vector commands based on the physical addresses contained in the vector command.

Abstract: Techniques are disclosed relating to speculative execution of store instructions. In various embodiments, an integrated circuit includes an execution pipeline having a load store circuit. The load store circuit is configured to receive a first store instruction executable to store a first value in a memory accessible to the integrated circuit. Prior to the first store instruction committing, the load store circuit stores the first value in a store buffer. In response to the first store instruction committing, the load store circuit stores, in the store buffer, an indication that the first store instruction has committed. In various embodiments, the integrated circuit reads the stored indication to determine whether the first store instruction has committed and, responsive to the read indication, provides the first value for storage in the memory.

Abstract: The present disclosure generally relates to systems and methods for executing commands that are larger than the maximum data transfer size (MDTS) on a data storage device. The data storage device advertises to the host device a MDTS that is higher than the actual MDTS that the data storage device can actually handle. If the data storage device receives a command that is equal to or less than the actual MDTS, then the data storage device processes the command normally. If the data storage device receives a command that is greater than the actual MDTS, but with less than or equal to the advertised MDTS, the data storage device splits the command into a plurality of dummy commands for processing. Once all of the dummy commands have processed, the data storage device delivers a completion message to the host device.

Abstract: A method of activating scheduling instructions within a parallel processing unit is described. The method comprises decoding, in an instruction decoder, an instruction in a scheduled task in an active state and checking, by an instruction controller, if a swap flag is set in the decoded instruction. If the swap flag in the decoded instruction is set, a scheduler is triggered to de-activate the scheduled task by changing the scheduled task from the active state to a non-active state.

Abstract: A method of scheduling instructions within a parallel processing unit is described. The method comprises decoding, in an instruction decoder, an instruction in a scheduled task in an active state, and checking, by an instruction controller, if an ALU targeted by the decoded instruction is a primary instruction pipeline. If the targeted ALU is a primary instruction pipeline, a list associated with the primary instruction pipeline is checked to determine whether the scheduled task is already included in the list. If the scheduled task is already included in the list, the decoded instruction is sent to the primary instruction pipeline.

Abstract: Embodiments of this application disclose an instruction set processing method based on a chip architecture and apparatus, and a computer-readable storage medium. The method includes compiling a deep learning model based on the architecture of the chip, to obtain a deep learning instruction set corresponding to the chip; compressing the deep learning instruction set, to obtain a compressed instruction set; and storing the compressed instruction set in an instruction set buffer of the chip by writing in a register, the compressed instructions executing a task.