Abstract: A system is described that performs training operations for a neural network, the system including an analog circuit element functional block with an array of analog circuit elements, and a controller. The controller monitors error values computed using an output from each of one or more initial iterations of a neural network training operation, the one or more initial iterations being performed using neural network data acquired from the memory. When one or more error values are less than a threshold, the controller uses the neural network data from the memory to configure the analog circuit element functional block to perform remaining iterations of the neural network training operation. The controller then causes the analog circuit element functional block to perform the remaining iterations.

Abstract: Selecting an error correction code type for a memory device includes: selecting, by the memory device in dependence upon predefined selection criteria, one of a plurality of error correction code types and carrying out memory access requests utilizing the selected error correction code type.

Abstract: A memory module includes one or more programmable ECC engines that may be programed by a host processing element with a particular ECC implementation. As used herein, the term “ECC implementation” refers to ECC functionality for performing error detection and subsequent processing, for example using the results of the error detection to perform error correction and to encode corrupted data that cannot be corrected, etc. The approach allows an SoC designer or company to program and reprogram ECC engines in memory modules in a secure manner without having to disclose the particular ECC implementations used by the ECC engines to memory vendors or third parties.

Abstract: A processing system includes a processing device coupled to a memory configured to check for and correct faults in requested data. In response to correcting the faults of the requested data, the memory sends the corrected data and unused check bits to the processing device as a plurality of fetch returns. The memory also sends a parity fetch based on the corrected data and one or more operations to the processing device. After receiving the plurality of fetch returns and the unused check bits, the processing device checks each fetch return for faults based on the unused check bits. In response to determining that a fetch return includes a fault, the processing device erases the fetch return and reconstructs the fetch return based on one or more other received fetch returns and the parity fetch.

Abstract: A system comprising an electronic device that includes a processor is described. During operation, the processor acquires a full version of a neural network, the neural network including internal elements for processing instances of input image data having a set of color channels. The processor then generates, from the neural network, a set of sub-networks, each sub-network being a separate copy of the neural network with the internal elements for processing at least one of the color channels in instances of input image data removed, so that each sub-network is configured for processing a different set of one or more color channels in instances of input image data. The processor next provides the sub-networks for processing instances of input image data—and may itself use the sub-networks for processing instances of input image data.

Abstract: Error handling for resilient software includes: receiving data indicating a region of resilient memory; detecting an error associated with a region of memory; and preventing raising an exception for the error in response to the region of memory falling within the region of resilient memory by preventing the region of memory as being identified as including the error.

Abstract: Systems, apparatuses, and methods for managing buffers in a neural network implementation with heterogeneous memory are disclosed. A system includes a neural network coupled to a first memory and a second memory. The first memory is a relatively low-capacity, high-bandwidth memory while the second memory is a relatively high-capacity, low-bandwidth memory. During a forward propagation pass of the neural network, a run-time manager monitors the usage of the buffers for the various layers of the neural network. During a backward propagation pass of the neural network, the run-time manager determines how to move the buffers between the first and second memories based on the monitored buffer usage during the forward propagation pass. As a result, the run-time manager is able to reduce memory access latency for the layers of the neural network during the backward propagation pass.

Abstract: A memory system uses error detection codes to detect when errors have occurred in a region of memory. A count of the number of errors is kept and a notification is output in response to the number of errors satisfying a threshold value. The notification is an indication to a host (e.g., a program accessing or managing a machine learning system) that the threshold number of errors have been detected in the region of memory. As long as the number of errors that have been detected in the region of memory remains under the threshold number no notification need be output to the host.

Abstract: A system comprising an electronic device that includes a processor is described. During operation, the processor acquires a full version of a neural network, the neural network including internal elements for processing instances of input image data having a set of color channels. The processor then generates, from the neural network, a set of sub-networks, each sub-network being a separate copy of the neural network with the internal elements for processing at least one of the color channels in instances of input image data removed, so that each sub-network is configured for processing a different set of one or more color channels in instances of input image data. The processor next provides the sub-networks for processing instances of input image data—and may itself use the sub-networks for processing instances of input image data.

Abstract: A system and method for protecting memory instructions against faults are described. The system and method include converting the slave instructions to dummy operations, modifying memory arbiter to issue up to N master and N slave global/shared memory instructions per cycle, sending master memory requests to memory system, using slave requests for error checking, entering master requests to the GM/LM FIFO, storing slave requests in a register, and comparing the entered master requests with the stored slave requests.

Abstract: A method and apparatus for mitigating row hammer attacks is provided. A row hammer alert is generated by a component of a memory architecture controlling operation of a memory device. The component may be a memory controller, coherency logic, or data fabric. The component obtains a physical address of an aggressor row that caused the alert and obtains an identifier of an execution context corresponding to the physical address. The component generates an error message for a processing device, the error message including the identifier of the execution context. The processing device retrieves the error message when performing a context switch. The processing device then generates an event received by the operating system. The operating system then takes action to reduce row hammer by the execution context, such as ending, restarting, or throttling the execution context.

Abstract: A method and system for memory attack mitigation in a memory device includes receiving, at a memory controller, an allocation of a page in memory. One or more device controllers detects an aggressor-victim set within the memory. Based upon the detection, an address of the allocated page is identified for further action.

Abstract: Methods and processing devices are provided for error protection to support instruction replay for executing idempotent instructions at a processing in memory PIM device. The processing apparatus includes a PIM device configured to execute an idempotent instruction. The processing apparatus also includes a processor, in communication with the PIM device, configured to issue the idempotent instruction to the PIM device for execution at the PIM device and reissue the idempotent instruction to the PIM device when one of execution of the idempotent instruction at the PIM device results in an error and a predetermined latency period expires from when the idempotent instruction is issued.

Abstract: Error checking data used in offloaded operations is disclosed. A remote execution device receives a request from a host to store a data block in a memory region. The data block includes data and host-generated error checking information for the data. The remote execution device updates the data block by overwriting the host-generated error checking information with locally generated error checking information for the data. The data block is then stored in the memory region.

Abstract: Techniques for performing in-memory matrix multiplication, taking into account temperature variations in the memory, are disclosed. In one example, the matrix multiplication memory uses ohmic multiplication and current summing to perform the dot products involved in matrix multiplication. One downside to this analog form of multiplication is that temperature affects the accuracy of the results. Thus techniques are provided herein to compensate for the effects of temperature increases on the accuracy of in-memory matrix multiplications. According to the techniques, portions of input matrices are classified as effective or ineffective. Effective portions are mapped to low temperature regions of the in-memory matrix multiplier and ineffective portions are mapped to high temperature regions of the in-memory matrix multiplier. The matrix multiplication is then performed.

Abstract: Error checking data used in offloaded operations is disclosed. A remote execution device receives a request from a host to store a data block in a memory region. The data block includes data and host-generated error checking information for the data. The remote execution device updates the data block by overwriting the host-generated error checking information with locally generated error checking information for the data. The data block is then stored in the memory region.

Abstract: Providing host-based error detection capabilities in a remote execution device is disclosed. A remote execution device performs a host-offloaded operation that modifies a block of data stored in memory. Metadata is generated locally for the modified of block of data such that the local metadata generation emulates host-based metadata generation. Stored metadata for the block of data is updated with the locally generated metadata for the modified portion of the block of data. When the host performs an integrity check on the modified block of data using the updated metadata, the host does not distinguish between metadata generated by the host and metadata generated in the remote execution device.

Abstract: Providing host-based error detection capabilities in a remote execution device is disclosed. A remote execution device performs a host-offloaded operation that modifies a block of data stored in memory. Metadata is generated locally for the modified of block of data such that the local metadata generation emulates host-based metadata generation. Stored metadata for the block of data is updated with the locally generated metadata for the modified portion of the block of data. When the host performs an integrity check on the modified block of data using the updated metadata, the host does not distinguish between metadata generated by the host and metadata generated in the remote execution device.

Abstract: Methods and processing devices are provided for error protection to support instruction replay for executing idempotent instructions at a processing in memory PIM device. The processing apparatus includes a PIM device configured to execute an idempotent instruction. The processing apparatus also includes a processor, in communication with the PIM device, configured to issue the idempotent instruction to the PIM device for execution at the PIM device and reissue the idempotent instruction to the PIM device when one of execution of the idempotent instruction at the PIM device results in an error and a predetermined latency period expires from when the idempotent instruction is issued.

Abstract: A method and system for memory attack mitigation in a memory device includes receiving, at a memory controller, an allocation of a page in memory. One or more device controllers detects an aggressor-victim set within the memory. Based upon the detection, an address of the allocated page is identified for further action.