Intel Corporation

Rahul Pal

Manoj K Arora

Dheemanth Nagaraj

Henk G Neefs

Sreenivas Subramoney

Bahaa Fahim

Sailesh Kottapalli

Ashok Jagannathan

Ragavendra Natarajan

David M Puffer

Yen-Cheng Liu

Saurabh Gupta

Mahesh K Kumashikar

Ajaya V Durg

Niranjan K Soundararajan

Philip Abraham

Ravindra P Saraf

Joydeep Ray

Altug Koker

Rajesh M Sankaran

Hong Wang

Robert G Blankenship

Andrew J Herdrich

Dheeraj R Subbareddy

Aravindh V Anantaraman

Md Altaf Hossain

Ankireddy Nalamalpu

Joseph Nuzman

Sridhar Muthrasanallur

Lakshminarayana Pappu

Ronak Singhal

Vedaraman Geetha

Swadesh Choudhary

Pratik M Marolia

Jared Warner Stark, Iv

Atul Maheshwari

Anshuman Thakur

Binata Bhattacharyya

Baskaran Ganesan

Sayantan Sur

Dongsheng Bi

Suresh Sugumar

Ling Cen

Niranjan Soundararajan

Umberto Santoni

Guadalupe J Garcia

Andrew Y Sun

Nayan Amrutlal Suthar

Nithiyanandan Bashyam

Mahesh Mamidipaka

Mahak Gupta

Binoy Balan

Sanilkumar Mm

Daniel Deng

C Santhosh

Rahul Joshi

Shrikul Atulkumar Joshi

Rajesh Vivekanandham

In a further embodiment, a system on a chip integrated circuit (SoC) is provided that includes an active base die including a first cache memory, a first die mounted on and coupled with the active base die, and a second die mounted on the active base die and coupled with the active base die and the first die. The first die includes an interconnect fabric, an input/output interface, and an atomic operation handler. The second die includes an array of graphics processing elements and an interface to the first cache memory of the active base die. At least one of the graphics processing elements are configured to perform, via the atomic operation handler, an atomic operation to a memory device.

Atomic handling for disaggregated 3D structured SoCs

The present disclosure is directed to 3-D stacked architecture for Programmable Fabrics and Central Processing Units (CPUs). The 3-D stacked orientation enables reconfigurability of the fabric, and allows the fabric to function using coarse-grained and fine-grained acceleration for offloading CPU processing. Additionally, the programmable fabric may be able to function to interface with multiple other compute chiplet components in the 3-D stacked orientation. This enables multiple compute components to communicate without the need for offloading the data communications between the compute chiplets.

Three-dimensional stacked programmable logic fabric and processor design architecture

Methods and apparatus for an accelerator controller hub (ACH). The ACH may be a stand-alone component or integrated on-die or on package in an accelerator such as a GPU. The ACH may include a host device link (HDL) interface, one or more Peripheral Component Interconnect Express (PCIe) interfaces, one or more high performance accelerator link (HPAL) interfaces, and a router, operatively coupled to each of the HDL interface, the one or more PCIe interfaces, and the one or more HPAL interfaces. The HDL interface is configured to be coupled to a host CPU via an HDL link and the one or more HPAL interfaces are configured to be coupled to one or more HPALs that are used to access high performance accelerator fabrics (HPAFs) such as NVlink fabrics and CCIX (Cache Coherent Interconnect for Accelerators) fabrics. Platforms including ACHs or accelerators with integrated ACHs support RDMA transfers using RDMA semantics to enable transfers between accelerator memory on initiators and targets without CPU involvement.

Accelerator controller hub

A processor unit comprising a first controller to couple to a host processing unit over a first link; a second controller to couple to a second processor unit over a second link, wherein the second processor unit is to couple to the host central processing unit via a third link; and circuitry to determine whether to send a cache coherent request to the host central processing unit over the first link or over the second link via the second processing unit.

Peer-to-peer link sharing for upstream communications from XPUS to a host processor

A processor, including a core; and a cache-coherent memory fabric coupled to the core and having a primary cache agent (PCA) configured to provide a primary access path; and a secondary cache agent (SCA) configured to provide a secondary access path that is redundant to the primary access path, wherein the PCA has a coherency controller configured to maintain data in the secondary access path coherent with data in the main access path.

Redundant cache-coherent memory fabric

A method to check for redundancy in two or more data lines comprises receiving data on a first data line, computing a first cyclic redundancy check (CRC) value on the data of the first data line, performing an exclusive OR (XOR) function on the first CRC value with a stored memory value, and updating the stored memory value with a result of the XOR function, and repeating on additional data lines until a last line is processed such that an error is indicated if a final stored memory value is not zero. An apparatus to check that two cores are operating in lockstep comprises a first core comprising a first data checker, a second core comprising a second data checker, and a lockstep checker to compare an output of the first data checker with an output of the second data checker.

Hardware lockstep checking within a fault detection interval in a system on chip

A hetero-mesh architecture is provided to enable varying densities of tile in a multi-core processor. The hetero-mesh architecture includes areas with different tile sizes and wire densities operating and different bandwidths. A split merge switch is utilized between the different parts of the hetero-mesh to enable the sending of packets from tiles in one area of the hetero-mesh to another area of the hetero-mesh while employing a single end to end communication protocol.

Method and apparatus to build a monolithic mesh interconnect with structurally heterogenous tiles

Embodiments of apparatuses, methods, and systems for misprediction-triggered local history-based branch prediction are described. In one embodiments, an apparatus includes a current pattern table and a local pattern table. The current pattern table has a plurality of entries, each entry in which to store a plurality of pattern lengths of a current pattern of one of a plurality of branch instructions. The local pattern table is to provide a first branch prediction based on the current pattern.

Misprediction-triggered local history-based branch prediction

Operational faults, including transient faults, are detected within computing hardware for mission-critical applications. Operational requests received from a requestor node are to be processed by shared agents to produce corresponding responses. A first request is duplicated to be redundantly processed independently and asynchronously by distinct shared agents to produce redundant counterpart responses including a first redundant response and a second redundant response. The first redundant response is compared against the second redundant response. In response to a match, the redundant responses are merged to produce a single final response to the first request to be read by the requestor node. In response to a non-match, an exception response is performed.

Mission-critical computing architecture

One embodiment provides an apparatus. The apparatus includes a store direct dependent (SDD) branch prediction circuitry and an SDD management circuitry. The store direct dependent (SDD) branch prediction circuitry is to store an SDD branch table. The SDD branch table is to store at least one record. Each record includes a branch instruction pointer (IP) field, a load IP field, a store IP field, a comparison info field and at least one of a store value field and/or a predicted outcome field. The SDD management circuitry is to populate the SDD branch table at runtime and to override a baseline branch prediction associated with an incoming branch IP with an SDD branch prediction, if the SDD branch table contains a first record populated with the incoming branch IP and at least one of a store value and/or an SDD predicted outcome.

Dynamic detection and prediction for store-dependent branches

A processor may include a baseline branch predictor and an empirical branch bias override circuit. The baseline branch predictor may receive a branch instruction associated with a given address identifier, and generate, based on a global branch history, an initial prediction of a branch direction for the instruction. The empirical branch bias override circuit may determine, dependent on a direction of an observed branch direction bias in executed branch instruction instances associated with the address identifier, whether the initial prediction should be overridden, may determine, in response to determining that the initial prediction should be overridden, a final prediction that matches the observed branch direction bias, or may determine, in response determining that the initial prediction should not be overridden, a final prediction that matches the initial prediction. The predictor may update an entry in the global branch history reflecting the resolved branch direction for the instruction following its execution.

Branch predictor with empirical branch bias override

A chip multiprocessor may include a first cluster and a second cluster, each having multiple cores of a processor, multiple co-located cache slices, and a memory controller. The processor stores directory information in a memory to indicate cluster cache ownership of a first address space to the first cluster. In response to a request to change the cluster cache ownership of the first address space to a second address space of the second cluster, the processor provides a quiesce period during which to block new read or write requests to the first cluster and the second cluster; drain read or write requests issued on the first cluster and the second cluster; and remove the block on new read or write requests. The processor may also update the directory information to change the cluster cache ownership of the first address space to the second address space of the second cluster.

Changing cache ownership in clustered multiprocessor

An apparatus and method are described for distributed snoop filtering. For example, one embodiment of a processor comprises: a plurality of cores to execute instructions and process data; first snoop logic to track a first plurality of cache lines stored in a mid-level cache ('MLC') accessible by one or more of the cores, the first snoop logic to allocate entries for cache lines stored in the MLC and to deallocate entries for cache lines evicted from the MLC, wherein at least some of the cache lines evicted from the MLC are retained in a level 1 (L1) cache; and second snoop logic to track a second plurality of cache lines stored in a non-inclusive last level cache (NI LLC), the second snoop logic to allocate entries in the NI LLC for cache lines evicted from the MLC and to deallocate entries for cache lines stored in the MLC, wherein the second snoop logic is to store and maintain a first set of core valid bits to identify cores containing copies of the cache lines stored in the NI LLC.

Method and apparatus for distributed snoop filtering

Resolving coherency issues inherent in sharing distributed cache is described. A chip multiprocessor may include at least first and second processing clusters, each having multiple cores of a processor, multiple cache slices co-located with the multiple cores, and a memory controller (MC). The processor stores directory information in a memory coupled to the processor to indicate cluster cache ownership of a first address space to the first cluster. In response to a request to change the cluster cache ownership of the first address space, the processor may remap first lines of first cache slices, corresponding to the first address space, to second lines in second cache slices of the second cluster, and update the directory information (e.g., a state of the first cache lines) to change the cluster cache ownership of the first address space to the second cluster. One of the MCs may manage such updating of the directory.

A request associated with a particular cache record is generated to be sent to a system component associated with a cache bank over an interconnect. A wake-warn signal is sent over a dedicated wake-warn channel to indicate to the system component that the request is to arrive. Wake-warn signals cause a disabled clock to be ungated to an enabled state. The request is then sent to the system component.

Early wake-warn for clock gating control

Growing community of inventors

Bengaluru, India

Rahul Pal