Abstract: Techniques for decoupled access-execute near-memory processing include examples of first or second circuitry of a near-memory processor receiving instructions that cause the first circuitry to implement system memory access operations to access one or more data chunks and the second circuitry to implement compute operations using the one or more data chunks.

Abstract: Techniques for decoupled access-execute near-memory processing include examples of first or second circuitry of a near-memory processor receiving instructions that cause the first circuitry to implement system memory access operations to access one or more data chunks and the second circuitry to implement compute operations using the one or more data chunks.

Abstract: Examples herein relate to checkpoint replication and copying of updated checkpoint data. For example, a memory controller coupled to a memory can receive a write request with an associated address to write or update checkpoint data and track updates to checkpoint data based on at least two levels of memory region sizes. A first level is associated with a larger memory region size than a memory region size associated with the second level. In some examples, the first level is a cache-line memory region size and the second level is a page memory region size. Updates to the checkpoint data can be tracked at the second level unless an update was previously tracked at the first level. Reduced amounts of updated checkpoint data can be transmitted during a checkpoint replication by using multiple region size trackers.

Abstract: An apparatus is described. The apparatus includes a memory controller to interface with a multi-level memory, where, an upper level of the multi-level memory is to act as a cache for a lower level of the multi-level memory. The memory controller has circuitry to determine: i) an original address of a slot in the upper level of memory from an address of a memory request in a direct mapped fashion; ii) a miss in the cache for the request because the slot is pinned with data from another address that competes with the address; iii) a partner slot of the slot in the cache in response to the miss; iv) whether there is a hit or miss in the partner slot in the cache for the request.

Abstract: A processor that includes compression instructions to compress multiple adjacent data blocks of uncompressed read-only data stored in memory into one compressed read-only data block and store the compressed read-only data block in multiple adjacent blocks in the memory is provided. During execution of an application to operate on the read-only data, one of the multiple adjacent blocks storing the compressed read-only block is read from memory, stored in a prefetch buffer and decompressed in the memory controller. In response to a subsequent request during execution of the application for an adjacent data block in the compressed read-only data block, the uncompressed adjacent block is read directly from the prefetch buffer.

Abstract: A method is described. The method includes receiving a read or write request for a cache line. The method includes directing the request to a set of logical super lines based on the cache line's system memory address. The method includes associating the request with a cache line of the set of logical super lines. The method includes, if the request is a write request: compressing the cache line to form a compressed cache line, breaking the cache line down into smaller data units and storing the smaller data units into a memory side cache. The method includes, if the request is a read request: reading smaller data units of the compressed cache line from the memory side cache and decompressing the cache line.

Abstract: A two-level main memory that includes a persistent memory and a cache is provided. Locations of dirty cache lines in the cache are tracked through the use a dirty cache line tracker. The dirty cache line tracker is stored in the cache and can be cached in a memory controller for the persistent memory. The dirty cache line tracker can be used to bypass cache lookup, perform efficient dirty cache line scrubbing and to decouple battery power and capacity of the cache in the two-level main memory.

Abstract: A two-level main memory in which both volatile memory and persistent memory are exposed to the operating system in a flat manner and data movement and management is performed in cache line granularity is provided. The operating system can allocate pages in the two-level main memory randomly across the first level main memory and the second level main memory in a memory-type agnostic manner, or, in a more intelligent manner by allocating predicted hot pages in first level main memory and predicted cold pages in second level main memory. The cache line granularity movement is performed in a “swap” manner, that is, a hot cache line in the second level main memory is swapped with a cold cache line in first level main memory because data is stored in either first level main memory or second level main memory not in both first level main memory and second level main memory.

Abstract: A method is described. The method includes receiving a read or write request for a cache line. The method includes directing the request to a set of logical super lines based on the cache line's system memory address. The method includes associating the request with a cache line of the set of logical super lines. The method includes, if the request is a write request: compressing the cache line to form a compressed cache line, breaking the cache line down into smaller data units and storing the smaller data units into a memory side cache. The method includes, if the request is a read request: reading smaller data units of the compressed cache line from the memory side cache and decompressing the cache line.

Abstract: One embodiment provides an apparatus. The apparatus includes last level cache circuitry and cache management circuitry. The last level cache circuitry stores cache blocks that at least partially include a subset of cache blocks stored by near memory circuitry. The near memory circuitry is configured in an n-way set associative format that references the cache blocks stored by the near memory circuitry using set identifiers and way identifiers. The cache management circuitry stores way identifiers for the cache blocks of the near memory circuitry within the cache blocks in the last level cache circuitry. Storing way identifiers in the cache blocks of the last level cache enables the cache management circuitry or memory controller circuitry to write back a cache block without reading tags in one or more ways of the near memory circuitry.

Abstract: Provided are an apparatus and system to cache data in a first cache and a second cache that cache data from a shared memory in a local processor node, wherein the shared memory is accessible to at least one remote processor node. A cache controller writes a block to the second cache in response to determining that the block is more likely to be accessed by the local processor node than a remote processor node. The first cache controller writes the block to the shared memory in response to determining that the block is more likely to be accessed by the one of the at least one remote processor node than the local processor node without writing to the second cache.

Abstract: An apparatus is described that includes a memory controller to couple to a multi-level memory characterized by a faster higher level and a slower lower level. The memory controller having early demotion logic circuitry to demote a page from the higher level to the lower level without system software having to instruct the memory controller to demote the page and before the system software promotes another page from the lower level to the higher level.

Abstract: A device is configured to be in communication with one or more host cores via a first communication path. A first set of processing-in-memory (PIM) cores and a second set of PIM cores are configured to be in communication with a memory included in the device over a second communication path, wherein the first set of PIM cores have greater processing power than the second set of PIM cores, and wherein the second communication path has a greater bandwidth for data transfer than the first communication path. Code offloaded by the one or more host cores are executed in the first set of PIM cores and the second set of PIM cores.

Abstract: To reduce overhead for cache coherence for shared cache in multi-processor systems, adaptive granularity allows tracking shared data at a coarse granularity and unshared data at fine granularity. Processes for adaptive granularity select how large of an entry is required to track the coherence of a block based on its state. Shared blocks are tracked in coarse-grained region entries that include a sharer tracking bit vector and a bit vector that indicates which blocks are likely to be present in the system, but do not identify the owner of the block. Modified/unshared data is tracked in fine-grained entries that permit ownership tracking and exact location and invalidation of cache. Large caches where the majority of blocks are shared and not modified create less overhead by being tracked in the less costly coarse-grained region entries.

Abstract: Examples herein relate to checkpoint replication and copying of updated checkpoint data. For example, a memory controller coupled to a memory can receive a write request with an associated address to write or update checkpoint data and track updates to checkpoint data based on at least two levels of memory region sizes. A first level is associated with a larger memory region size than a memory region size associated with the second level. In some examples, the first level is a cache-line memory region size and the second level is a page memory region size. Updates to the checkpoint data can be tracked at the second level unless an update was previously tracked at the first level. Reduced amounts of updated checkpoint data can be transmitted during a checkpoint replication by using multiple region size trackers.

Abstract: An apparatus is described. The apparatus includes a memory controller to interface with a multi-level memory, where, an upper level of the multi-level memory is to act as a cache for a lower level of the multi-level memory. The memory controller has circuitry to determine: i) an original address of a slot in the upper level of memory from an address of a memory request in a direct mapped fashion; ii) a miss in the cache for the request because the slot is pinned with data from another address that competes with the address; iii) a partner slot of the slot in the cache in response to the miss; iv) whether there is a hit or miss in the partner slot in the cache for the request.

Abstract: Techniques for decoupled access-execute near-memory processing include examples of first or second circuitry of a near-memory processor receiving instructions that cause the first circuitry to implement system memory access operations to access one or more data chunks and the second circuitry to implement compute operations using the one or more data chunks.

Abstract: Provided are an apparatus, system and method to determine whether to use a low or high read voltage. First level indications of write addresses, for locations in the non-volatile memory to which write requests have been directed, are included in a first level data structure. For a write address of the write addresses having a first level indication in the first level data structure, the first level indication of the write address is removed from the first level data structure and a second level indication for the write address is added to a second level data structure to free space in the first level data structure to indicate a further write address. A first voltage level is used to read data from read addresses mapping to one of the first and second level indications in the first and the second level data structures, respectively.

Abstract: Systems, apparatuses and methods may provide for technology to maintain a prediction table that tracks missed page addresses with respect to a first memory. If an access request does not correspond to any valid page addresses in the prediction table, the access request may be sent to the first memory. If the access request corresponds to a valid page address in the prediction table, the access request may be sent to the first memory and a second memory in parallel, wherein the first memory is associated with a shorter access time than the second memory.

Abstract: A two-level main memory that includes a persistent memory and a cache is provided. Locations of dirty cache lines in the cache are tracked through the use a dirty cache line tracker. The dirty cache line tracker is stored in the cache and can be cached in a memory controller for the persistent memory. The dirty cache line tracker can be used to bypass cache lookup, perform efficient dirty cache line scrubbing and to decouple battery power and capacity of the cache in the two-level main memory.