Abstract: A multi-protocol storage device avoids entering a thermal shutdown mode by switching between protocols. The storage device communicates with a host in a first mode using a first protocol. The storage device receives a temperature request from the host, monitors its temperature, and transmits a response to the host when the temperature of the storage device meets a predefined temperature that is below a thermal shutdown threshold. The storage device receives a thermal throttling instruction from the host and switches to a second mode to communicate with the host using a second protocol that uses less resources than the first protocol. The storage device performs thermal throttling until the temperature of the storage device reaches a normal temperature zone. When the temperature of the storage device returns to the normal temperature zone, the storage device returns to the first mode.

Abstract: Methods and apparatus are provided for arbitrating access to, and usage of, various device resources of a data storage device (DSD) configured for Machine Learning with Low-Power. The data storage device may include a TinyML controller with an artificial intelligence (AI) accelerator integrated with a data storage controller on a system-on-a-chip (SoC). The device resources may be, e.g., storage resources such as random access memory (RAM) devices, non-volatile memory (NVM) arrays, and latches formed on NVM dies of the NVM arrays. The resource arbitration may be based, for example, on parameters pertaining to ML operations performed by an ML controller that includes the AI accelerator, such as a turnaround time of an ML epoch or a stage-wise execution time. The resource arbitration is configured to provide for the efficient interleaving of the ML/AI operations performed by the ML controller and data storage operations performed by the data storage controller.

Abstract: Methods and apparatus are provided for arbitrating access to, and usage of, various device resources of a data storage device (DSD) configured for Machine Learning with Low-Power. The data storage device may include a TinyML controller with an artificial intelligence (AI) accelerator integrated with a data storage controller on a system-on-a-chip (SoC). The device resources may be, e.g., storage resources such as random access memory (RAM) devices, non-volatile memory (NVM) arrays, and latches formed on NVM dies of the NVM arrays. The resource arbitration may be based, for example, on parameters pertaining to ML operations performed by an ML controller that includes the AI accelerator, such as a turnaround time of an ML epoch or a stage-wise execution time. The resource arbitration is configured to provide for the efficient interleaving of the ML/AI operations performed by the ML controller and data storage operations performed by the data storage controller.

Abstract: In one embodiment, a memory system is provided comprising a memory die and a controller. The memory die comprises a non-volatile memory, a data latch, and an on-chip randomizer. The controller is configured to send a command to the memory die to cause the on-chip randomizer to store random data in the data latch and send data to the memory die to overwrite some, but not all, of the random data in the data latch, wherein the memory die is configured to transfer the data and random data stored in the data latch to the non-volatile memory. Other embodiments are provided.

Abstract: In one embodiment, a memory system is provided comprising a memory die and a controller. The memory die comprises a non-volatile memory, a data latch, and an on-chip randomizer. The controller is configured to send a command to the memory die to cause the on-chip randomizer to store random data in the data latch and send data to the memory die to overwrite some, but not all, of the random data in the data latch, wherein the memory die is configured to transfer the data and random data stored in the data latch to the non-volatile memory. Other embodiments are provided.