Advanced Micro Devices Corporation

Jagadish B Kotra

John Kalamatianos

Sergey Blagodurov

Marko Scrbak

Paul J Moyer

William Louie Walker

Kishore Punniyamurthy

Mahzabeen Islam

Nuwan S Jayasena

Gabriel H Loh

Yasuko Eckert

Shaizeen Dilawarhusen Aga

Marius Evers

Michael T Clark

Jay E Fleischman

Joseph Lee Greathouse

Ganesh Suryanarayan Dasika

Steven E Raasch

Karthik Rao

Matthew R Poremba

Vignesh Adhinarayanan

Johnathan R Alsop

Seyedmohammad SeyedzadehDelcheh

Michael Wayne LeBeane

Onur Kayiran

Patrick J Shyvers

Ashwin Aji

Sriram Srinivasan

Travis Henry Boraten

Sriseshan Srikanth

David Andrew Werner

Weon Taek Na

Dong Kai Wang

Gagandeep Panwar

Nagadastagiri Reddy Challapalle

Nicholas Dean Lance

Asmita Pal

Ioannis Papadopoulos

Yonghae Kim

The disclosed device uses a control circuit configured to (i) receive branch misprediction information corresponding to a mispredicted branch window of instructions and (ii) send a misprediction status of a memory access from the mispredicted branch window of instructions, and a cache prefetcher of a cache configured to train using a set of memory accesses that are updated in response to receiving the misprediction status from the control circuit.

Systems and methods for branch misprediction aware cache prefetcher training

An apparatus and method for creating less computationally intensive nodes for a neural network. An integrated circuit includes a host processor and multiple memory channels, each with multiple memory array banks. Each of the memory array banks includes components of a processing-in-memory (PIM) accelerator and a scatter and gather circuit used to dynamically perform quantization operations and dequantization operations that offload these operations from the host processor. The host processor executes a data model that represents a neural network. The memory array banks store a single copy of a particular data value in a single precision. Therefore, the memory array banks avoid storing replications of the same data value with different precisions to be used by a neural network node. The memory array banks dynamically perform quantization operations and dequantization operations on one or more of the weight values, input data values, and activation output values of the neural network.

Method and apparatus for quantization and dequantization of neural network input and output data using processing-in-memory

Systems, apparatuses, and methods for compacting multiple groups of micro-operations into individual cache lines of a micro-operation cache are disclosed. A processor includes at least a decode unit and a micro-operation cache. When a new group of micro-operations is decoded and ready to be written to the micro-operation cache, the micro-operation cache determines which set is targeted by the new group of micro-operations. If there is a way in this set that can store the new group without evicting any existing group already stored in the way, then the new group is stored into the way with the existing group(s) of micro-operations. Metadata is then updated to indicate that the new group of micro-operations has been written to the way. Additionally, the micro-operation cache manages eviction and replacement policy at the granularity of micro-operation groups rather than at the granularity of cache lines.

Implementing a micro-operation cache with compaction

A memory request issue counter (MRIC) is maintained that is incremented for every memory access a central processing unit core makes. A region reuse distance table is also maintained that includes multiple entries each of which stores the region reuse distance for a corresponding region. When a memory access request for a physical address is received, a reuse distance for the physical address is calculated. This reuse distance is the difference between the current MRIC value and a previous MRIC value for the physical address. The previous MRIC value for the physical address is the MRIC value the MRIC had when a memory access request for the physical address was last received. A region reuse distance for a region that includes the physical address is generated based on the reuse distance for the physical address and used to manage the cache.

Managing a cache using per memory region reuse distance estimation

An entry of a last level cache shadow tag array to track pending last level cache misses to private data in a previous level cache (e.g., an L2 cache), that also are misses to an exclusive last level cache (e.g., an L3 cache) and to the last level cache shadow tag array. Accordingly, last level cache miss status holding registers need not be expended to track cache misses to private data that are already being tracked by a previous level cache miss status holding register. Additionally or alternatively, up to a threshold number of last level cache pending misses to the same shared data from different processor cores are tracked in the last level cache shadow tag array, and any additional last level cache pending misses are tracked in a last level cache miss status holding register.

Method and apparatus for increasing memory level parallelism by reducing miss status holding register allocation in caches

Address translation is performed to translate a virtual address targeted by a memory request (e.g., a load or memory request for data or an instruction) to a physical address. This translation is performed using an address translation buffer, e.g., a translation lookaside buffer (TLB). One or more actions are taken to reduce data access latencies for memory requests in the event of a TLB miss where the virtual address to physical address translation is not in the TLB. Examples of actions that are performed in various implementations in response to a TLB miss include bypassing level 1 (L1) and level 2 (L2) caches in the memory system, and speculatively sending the memory request to the L2 cache while checking whether the memory request is satisfied by the L1 cache.

Accessing a cache based on an address translation buffer result

Selectively bypassing cache directory lookups for processing-in-memory instructions is described. In one example, a system maintains information describing a status—clean or dirty—of a memory address, where a dirty status indicates that the memory address is modified in a cache and thus different than the memory address as represented in system memory. A processing-in-memory request involving the memory address is assigned a cache directory bypass bit based on the status of the memory address. The cache directory bypass bit for a processing-in-memory request controls whether a cache directory lookup is performed after the processing-in-memory request is issued by a processor core and before the processing-in-memory request is executed by a processing-in-memory component.

Bypassing cache directory lookups for processing-in-memory instructions

An apparatus configured for offloading system service tasks to a processing-in-memory ('PIM') device includes an agent configured to: receive, from a host processor, a request to offload a memory task associated with a system service to the PIM device; determine at least one PIM command and at least one memory page associated with the host processor based upon the request; and issue the at least one PIM command to the PIM device for execution by the PIM device to perform the memory task upon the at least one memory page.

Method and apparatus to expedite system services using processing-in-memory (PIM)

Methods, devices, and systems for retrieving information based on cache miss prediction. It is predicted, based on a history of cache misses at a private cache, that a cache lookup for the information will miss a shared victim cache. A speculative memory request is enabled based on the prediction that the cache lookup for the information will miss the shared victim cache. The information is fetched based on the enabled speculative memory request.

Speculative dram request enabling and disabling

Processor-guided execution of offloaded instructions using fixed function operations is disclosed. Instructions designated for remote execution by a target device are received by a processor. Each instruction includes, as an operand, a target register in the target device. The target register may be an architected virtual register. For each of the plurality of instructions, the processor transmits an offload request in the order that the instructions are received. The offload request includes the instruction designated for remote execution. The target device may be, for example, a processing-in-memory device or an accelerator coupled to a memory.

Processor-guided execution of offloaded instructions using fixed function operations

In accordance with the described techniques for leveraging processing in memory registers as victim buffers, a computing device includes a memory, a processing in memory component having registers for data storage, and a memory controller having a victim address table that includes at least one address of a row of the memory that is stored in the registers. The memory controller receives a request to access the row of the memory and accesses data of the row from the registers based on the address of the row being included in the victim address table.

Leveraging processing in memory registers as victim buffers

A method for operating a memory having a plurality of banks accessible in parallel, each bank including a plurality of grains accessible in parallel is provided. The method includes: based on a memory access request that specifies a memory address, identifying a set that stores data for the memory access request, wherein the set is spread across multiple grains of the plurality of grains; and performing operations to satisfy the memory access request, using entries of the set stored across the multiple grains of the plurality of grains.

Tag and data configuration for fine-grained cache memory

Offloading computations from a processor to remote execution logic is disclosed. Offload instructions for remote execution on a remote device are dispatched in the form of processor instructions like conventional instructions. In the processor, an offload instruction is inserted in an offload queue. The offload instruction may be inserted at the dispatch stage or the retire stage of the processor pipeline. Metadata for the offload instruction is added to the offload instruction in the offload queue. After retirement of the offload instruction, the processor transmits an offload request generated from the offload instruction.

Offloading computations from a processor to remote execution logic

In accordance with the described techniques for data compression and decompression for processing in memory, a page address is received by a processing in memory component that maps to a first location in memory where data of a page is maintained. The data of the page is compressed by the processing in memory component. Further, compressed data of the page is written by the processing in memory component to a compressed block device responsive to the compressed data satisfying one or more compressibility criteria. The compressed block device is a portion of the memory dedicated to storing data in a compressed form.

Data compression and decompression for processing in memory

Arbitrating atomic memory operations, including: receiving, by a media controller, a plurality of atomic memory operations; determining, by an atomics controller associated with the media controller, based on one or more arbitration rules, an ordering for issuing the plurality of atomic memory operations; and issuing the plurality of atomic memory operations to a memory module according to the ordering.

Growing community of inventors

Austin, TX, United States of America

Jagadish B Kotra