Abstract: Systems and techniques for a memory-side cache directory-based request queue are described herein. A memory request is received at an interface of a memory device. One or more fields of the memory request are written into an entry of a directory data structure. The identifier of the entry is pushed onto a queue. To perform the memory request, the identifier is popped off of the queue and a field of the memory request is retrieved from the entry of the directory data structure using the identifier. Then, a process on the memory request can be performed using the field retrieved from the entry of the directory data structure.

Abstract: System and techniques for variable execution time atomic operations are described herein. When an atomic operation for a memory device is received, the run length of the operation is measured. If the run length is beyond a threshold, a cache line for the operation is locked while the operation runs. A result of the operation is queued until it can be written to the cache line. At that point, the cache line is unlocked.

Abstract: To achieve low-latency register error correction, a register can be read as part of an instruction when that instruction is the currently executing instruction in a processor. A correctable error in data produced from reading the register can be detected. In response to detecting the correctable error, the currently executing instruction in the processor can be changed into a register update instruction that is executed to overwrite the data in the register with corrected data. Then, the original (e.g., unchanged) instruction can be rescheduled.

Abstract: System and techniques for recall pending cache line eviction are described herein. A queue that includes a deferred memory request is kept for a cache line. Metadata for the queue is stored in a cache line tag. When a recall is needed, the metadata is written from the tag to a first recall storage, referenced by a memory request ID. After the recall request is transmitted, the memory request ID is written to a second recall storage referenced by the message ID of the recall request. Upon receipt of a response to the recall request, the queue for the cache line can be restored by using the message ID in the response to lookup the memory request ID from the second recall storage, then using the memory request ID to lookup the metadata from the first recall storage, and then writing the metadata into the tag for the cache line.

Abstract: System and techniques for host-preferred memory operation are described herein. At a memory-side cache of a memory device that includes accelerator hardware, a first memory operation can be received from a host. A determination that the first memory operation corresponds to a cache set based on an address of the first memory operation is made. A second memory operation can be received from the accelerator hardware. Another determination can be made that the second memory operation corresponds to the cache set. Here, the first memory operation can be enqueued in a host queue of the cache set and the second memory operation can be enqueued in an internal request queue of the cache set. The first memory operation and the second memory operation can be executed as each is dequeued.

Abstract: System and techniques for variable execution time atomic operations are described herein. When an atomic operation for a memory device is received, the run length of the operation is measured. If the run length is beyond a threshold, a cache line for the operation is locked while the operation runs. A result of the operation is queued until it can be written to the cache line. At that point, the cache line is unlocked.

Abstract: System and techniques for recall pending cache line eviction are described herein. A queue that includes a deferred memory request is kept for a cache line. Metadata for the queue is stored in a cache line tag. When a recall is needed, the metadata is written from the tag to a first recall storage, referenced by a memory request ID. After the recall request is transmitted, the memory request ID is written to a second recall storage referenced by the message ID of the recall request. Upon receipt of a response to the recall request, the queue for the cache line can be restored by using the message ID in the response to lookup the memory request ID from the second recall storage, then using the memory request ID to lookup the metadata from the first recall storage, and then writing the metadata into the tag for the cache line.

Abstract: System and techniques for memory side cache request handling are described herein. When a memory request is received, a cache set for the memory request is determined. Here, the cache set has multiple ways and each way corresponds to a cache line. It can be detected that a way of the multiple ways is not ready for the memory request. In this case, a representation of the memory request is stored in a queue of multiple queues based on an interface upon which the memory request was received and the present ways of the cache set. Entries from the multiple queues can be dequeued in a defined order to determine a next memory request to process. The defined order gives priority to memory requests for a present way and then for external over internal requests.

Abstract: System and techniques for evicting a cache line with pending control request are described herein. A memory request—that includes an address corresponding to a set of cache lines—can be received. A determination can be made that a cache line of the set of cache lines will be evicted to process the memory request. Another determination can be made that a control request has been made to a host from the memory device and that the control request pending when it is determined that the cache line will be evicted. Here, a counter corresponding to the set of cache lines can be incremented (e.g., by one) to track the pending control request in face of eviction. Then, the cache line can be evicted.

Abstract: System and techniques for silent cache line eviction are described herein. A memory device receives a memory operation from a host. The memory operation establishes data and metadata in a cache line of the memory device upon receipt. The metadata is stored in a memory element that corresponds to the cache line. Later, an eviction trigger to evict the cache line is identified. Then, in response to the eviction trigger, current metadata of the cache line is compared with the metadata in the memory element to determine whether the metadata has changed. the cache line can be evicted without writing to backing memory in response to the metadata being unchanged.

Abstract: System and techniques for synchronized request handling at a memory device are described herein. A request is received at the memory device. Here, the request indicates a memory address corresponding to a set of cache lines and a single cache line in the set of cache lines. The memory device maintains a deferred list for the set of cache lines and a set of lists with each member of the set of lists corresponding to one cache line in the set of cache lines. The memory device tests the deferred list to determine that the deferred list is not empty and places the request in the deferred list.

Abstract: Systems and techniques for a memory-side cache directory-based request queue are described herein. A memory request is received at an interface of a memory device. One or more fields of the memory request are written into an entry of a directory data structure. The identifier of the entry is pushed onto a queue. To perform the memory request, the identifier is popped off of the queue and a field of the memory request is retrieved from the entry of the directory data structure using the identifier. Then, a process on the memory request can be performed using the field retrieved from the entry of the directory data structure.

Abstract: System and techniques for memory side cache request handling are described herein. When a memory request is received, a cache set for the memory request is determined. Here, the cache set has multiple ways and each way corresponds to a cache line. It can be detected that a way of the multiple ways is not ready for the memory request. In this case, a representation of the memory request is stored in a queue of multiple queues based on an interface upon which the memory request was received and the present ways of the cache set. Entries from the multiple queues can be dequeued in a defined order to determine a next memory request to process. The defined order gives priority to memory requests for a present way and then for external over internal requests.

Abstract: Various examples are directed to systems and methods for executing a transaction between hardware compute elements of a computing system. A first hardware compute element may send a first write request to a second hardware compute element via a network structure. The first write request may comprise first source identifier data describing the first hardware compute element and first payload data describing a processing task requested by the first hardware compute element. The network structure may store first write request state data describing the first write request. Before the processing task is completed, the second hardware compute element may send a first write confirm message.

Abstract: A chiplet system can include a Serial Peripheral Interface (SPI) bus for communication. A primary device coupled to the SPI bus can generate read or write instructions for a secondary device. In response to instructions from the primary device, the secondary device can prepare a response. The response message can include a secondary device status field to indicate a readiness of the secondary device to provide a particular data payload to the primary device. Using deferrals that are communicated from the secondary device to the primary device can enable longer latency SPI operations to proceed without monopolizing the SPI bus.

Abstract: Devices and techniques for memory access bounds checking for a programmable atomic operator are described herein. A processor can execute a programmable atomic operator with a base memory address. The processor can obtain a memory interleave size indicator corresponding to the programmable atomic operator and calculate a contiguous memory address range from the base memory address and the memory interleave size. The processor can then detect that a memory request from the programmable atomic operator is outside the contiguous memory address range and deny the memory request when it is outside of the contiguous memory address range and allow the memory request otherwise.

Abstract: A chiplet system comprises an interposer including interconnect and multiple chiplets arranged on the interposer and interconnected using the interconnect of the interposer. The multiple chiplets include a throttle level bus source chiplet including a throttle level bus drive interface configured to place a throttle level value onto the throttle level bus, and one or more throttle level bus receiver chiplets operatively coupled to the throttle level bus. Each chiplet of the multiple chiplets includes throttling logic circuitry configured to set a throttle level of a chiplet according to the throttle level value.

Abstract: Devices and techniques for on-demand programmable atomic kernel loading are described herein. A programmable atomic unit (PAU) of a memory controller can receive an invocation of a programmable atomic operator by the memory controller. The PAU can then perform a verification on a programmable atomic operator partition for the programmable atomic operator. Here, the programmable atomic operator partition is located in a memory of the PAU. The PAU can then signal a trap in response to the verification indicating that the programmable atomic operator partition is not prepared.

Abstract: A system comprises an interposer including interconnect and multiple chiplets arranged on the interposer. Each chiplet includes multiple chiplet input-output (I/O) channels interconnected to I/O channels of other chiplets by the interposer; a chiplet I/O interface for the chiplet I/O channels that includes multiple interface layers; and initialization logic circuitry configured to advance initialization of the chiplet interface sequentially through the interface layers starting with a lowest interface layer.

Abstract: A chiplet system can include a Serial Peripheral Interface (SPI) bus for communication. A controller or primary device coupled to the SPI bus can generate a message with read or write instructions for one or more secondary devices. In an example, the primary device can be configured to use information on a data input port or data input bus to determine a communication status of one or multiple secondary devices on the bus.