Ati Technologies Ulc

Advanced Micro Devices Corporation

Kostantinos Danny Christidis

Jing Liu

Wade K Smith

HaiKun Dong

MinHua Wu

Gaojian Cong

Michael Mantor

Alexander J Branover

Benjamin Tsien

Anthony Asaro

James Raymond Magro

Mark Fowler

Lei Zhang

Angel E Socarras

Alexander Fuad Ashkar

Robert Scott Hartog

Sateesh Lagudu

Hui Zhou

Ling-Ling Wang

Dilip Kumar Jha

Rostyslav Kyrychynskyi

MingLiang Lin

Purushotham Niranjan Dasiga

Chi Yan Herburt Shek

Curtis Alan Gilgan

Wade Smith

Systems, apparatuses, and methods for implementing duplicated registers for access by initiators across multiple semiconductor dies are disclosed. A system includes multiple initiators on multiple semiconductor dies of a chiplet processor. One of the semiconductor dies is the master die, and this master die has copies of registers which can be accessed by the multiple initiators on the multiple semiconductor dies. When a given initiator on a given secondary die generates a register access, the register access is routed to the master die and a particular duplicate copy of the register maintained for the given secondary die. From the point of view of software, the multiple semiconductor dies appear as a single die, and the multiple initiators appear as a single initiator. Multiple types of registers can be maintained by the master die, with a flush register being one of the register types.

Duplicated registers in chiplet processing units

An apparatus and method for providing efficient power management for data transfer protocols between components. A source generates requests and a destination services the requests. The source and destination support a communication protocol that includes both a transfer channel and one or more transaction channels for each type of request. The source and destination rely on a valid signal and a ready signal of the transfer channels to autonomously manage power consumption. The source and destination remove any dependencies on an external power manager and make it unnecessary to add signal extensions to the communication protocol to support power management.

In-band communication interface power management fencing

A data processor includes a plurality of requestors, a plurality of responders, and a data fabric. The data fabric is for routing requests between the plurality of requestors and the plurality of responders and has a plurality of non-operational power states including a normal C-state and a light-weight C-state. The light-weight C-state has lower entry and exit latencies than the normal C-state. The data fabric monitors traffic through the data fabric and places the data fabric in the light-weight C-state in response to detecting an idle traffic state.

Data fabric C-state management

Apparatuses, systems and methods for routing requests and responses targeting a shared resource. A queue in a communication fabric is located in a path between the requesters and a shared resource. In some embodiments, the shared resource is a shared address translation cache stored in an endpoint. The physical channel between the queue and the shared resource supports multiple virtual channels. The queue assigns at least one entry to each virtual channel of a group of virtual channels where the group includes a virtual channel for each address translation request type from a single requester of the multiple requesters. When the at least one entry for a given requester is de-allocated, the queue allocates this entry only with requests from the assigned virtual channel even if the empty entry is the only available entry of the queue.

Transmission of address translation type packets

Systems, apparatuses, and methods for implementing a broadcast read response protocol are disclosed. A computing system includes a plurality of processing engines coupled to a memory subsystem. A first processing engine executes a read and broadcast response command, wherein the read and broadcast response command targets first data at a first address in the memory subsystem. One or more other processing engines execute a wait command to wait to receive the first data requested by the first processing engine. After receiving the first data from the memory subsystem, the plurality of processing engines process the first data as part of completing a first operation. In one implementation, the first operation is implementing a given layer of a machine learning model. In one implementation, the given layer is a convolutional layer of a neural network.

Broadcast command and response

Techniques are provided for using a translation lookaside buffer to provide low latency memory address translations for data streams. Clients of a memory system first prepare the address translation cache hierarchy by requesting that a translation pre-fetch stream is initialized. After the translation pre-fetch stream is initialized, the cache hierarchy returns an acknowledgment of completion to the client, which then begins to access memory. Pre-fetch streams are specified in terms of address ranges and are performed for large contiguous portions of the virtual memory address space.

Streaming translation lookaside buffer

Apparatuses, computer readable mediums, and methods of processor unit testing using cache resident testing are disclosed. The method may include loading a test program in a cache on a chip comprising one or more processor units. The method may include the one or more processor units executing the test program to generate one or more results. The method may include redirecting a first memory reference to the cache, wherein the first memory reference is generated during the execution of the test program. The method may include determining whether the one or more generated results match one or more test results. The method may include redirecting a memory request to a memory location resident in the cache if the memory request includes a memory location not resident in the cache. The method may include redirecting a memory request to the cache if the memory request is not directed to the cache.

System and method of testing processor units using cache resident testing

A system and method for efficient arbitration of memory access requests are described. One or more functional units generate memory access requests for a partitioned memory. An arbitration unit stores the generated requests and selects a given one of the stored requests. The arbitration unit identifies a given partition of the memory which stores a memory location targeted by the selected request. The arbitration unit determines whether one or more other stored requests access memory locations in the given partition. The arbitration unit sends each of the selected memory access request and the identified one or more other memory access requests to the memory to be serviced out of order.

Efficient arbitration for memory accesses

Techniques are provided for managing address translation request traffic where memory access requests can be made with differing quality-of-service levels, which specify latency and/or bandwidth requirements. The techniques involve translation lookaside buffers. Within the translation lookaside buffers, certain resources are reserved for specific quality-of-service levels. More specifically, translation lookaside buffer slots, which store the actual translations, as well as finite state machines in a work queue, are reserved for specific quality-of-service levels. The translation lookaside buffer receives multiple requests for address translation. The translation lookaside buffer selects requests having the highest quality-of-service level for which an available finite state machine is available. The fact that finite state machines are reserved to particular quality-of-service levels means that if all such finite state machines for a particular quality-of-service level are used by pending translation requests, then the translation lookaside buffer does not accept more translation requests for that quality-of-service level.

Sharing translation lookaside buffer resources for different traffic classes

A translation layer includes a plurality of first buffers and a controller to assert one or more ready signals corresponding to one or more of the plurality of first buffers in response to the one or more of the plurality of first buffers being less than full. The one or more of the plurality of first buffers receives data or control information from one or more corresponding components in response to the ready signal being asserted concurrently with one or more valid signals asserted by the one or more corresponding components.

Growing community of inventors

Toronto, Canada

Kostantinos Danny Christidis