Sun Microsystems, Inc.

Nvidia Corporation

Advanced Micro Devices Corporation

Digital Equipment Corporation

Oracle America, Inc.

Emulex Design & Manufacturing Corporation

Glenn A Dearth

Mark Douglas Hummel

Paul M Whittemore

George R Plouffe, Jr

Anwar Kashem

Thomas D Bissett

Nan Jiang

Bennet H Ih

Gregory Michael Thorson

Samuel Hammond Duncan

Denis J Foley

John R Feehrer

Larry Robert Dennison

William F Bruckert

Brian Kelleher

Edoardo Prete

Vyas Venkataraman

Kenneth A Ward

Ram Gummadi

Gerry Talbot

Daniel Joseph Lustig

David A Medeiros

John Wortman

Rich Reeves

Brian W Amick

Gerald Robert Talbot

Eitan Zahavi

Gil Bloch

Sudhir Srinivasan

Warren Anderson

Miriam Menes

Christopher J Jackson

Thomas P Webber

David E Kovalcin

Carl John Lindeborg

Dane Mrazek

Shwetal Arvind Patel

Richard W Reeves

James Duda

Joseph James Ervin

Stephen A Jay

Ahmad Atamli

Dennis Mazur

Mark R Johnson

Jonathan Owen

Robin Lee Brown

David M Kaffine

John Munzer

Karan Gupta

Mike Osborn

Domenic Dimeo

Gyorgy Rubin

Janet Y Zheng

John P Pabisz

Tushar A Parikh

Scott R Meeth

Sean Cummins

Peter R Mott, Jr

Ann Katan

Alex Ishii

Carlos Alonso

Paul Paternoster

Meni Orenbach

Eric Anderson

Richard Reeves

Warren R Anderson

A method is provided for operating a network switch comprising a plurality of input ports and a plurality of output ports. The method comprises receiving a first data packet received via a first input port and a second data packet received via a second input port to be delivered to an egress endpoint connected to a first output port, configuring a plurality of crossbar switch units arranged in a tiled architecture to pass the first data packet to the first output port via a primary path and pass the second data packet to the first output port via a secondary path, and transmitting the first data packet and the second data packet to the egress endpoint. The first data packet and the second data packet pass through the plurality of crossbar switch units simultaneously.

Crossbar multipathing for multicast performance in tiled switches

In various examples, a single notification (e.g., a request for a memory access operation) that a processing element (PE) has reached a synchronization barrier may be propagated to multiple physical addresses (PAs) and/or devices associated with multiple processing elements. Thus, the notification may allow an indication that the processing element has reached the synchronization barrier to be recoded at multiple targets. Each notification may access the PAs of each PE and/or device of a barrier group to update a corresponding counter. The PEs and/or devices may poll or otherwise use the counter to determine when each PE of the group has reached the synchronization barrier. When a corresponding counter indicates synchronization at the synchronization barrier, a PE may proceed with performing a compute task asynchronously with one or more other PEs until a subsequent synchronization barrier may be reached.

Efficient multi-device synchronization barriers using multicasting

In various examples, a memory model may support multicasting where a single request for a memory access operation may be propagated to multiple physical addresses associated with multiple processing elements (e.g., corresponding to respective local memory). Thus, the request may cause data to be read from and/or written to memory for each of the processing elements. In some examples, a memory model exposes multicasting to processes. This may include providing for separate multicast and unicast instructions or shared instructions with one or more parameters (e.g., indicating a virtual address) being used to indicate multicasting or unicasting. Additionally or alternatively, whether a request(s) is processed using multicasting or unicasting may be opaque to a process and/or application or may otherwise be determined by the system. One or more constraints may be imposed on processing requests using multicasting to maintain a coherent memory interface.

Multicast and reflective memory behavior for memory model consistency

Systems and techniques for performing multicast-reduction operations. In at least one embodiment, a network device receives first network data associated with a multicast operation to be collectively performed by at least a plurality of endpoints. The network device reserves resources to process second network data to be received from the endpoints, and sends the first network data to a plurality of additional network devices. The network device receives the second network data, and processes the second network data using the reserved resources.

Multicast-reduction assisted by network devices

Apparatuses, systems, and techniques are presented to configure computing resources to perform various tasks. In at least one embodiment, an approach presented herein can be used to verify whether a network of computing nodes is properly configured based, at least in part, on one or more expected data strings generated by the network of computing nodes.

Path attestation for computing resources

Fabric Attached Memory (FAM) provides a pool of memory that can be accessed by one or more processors, such as a graphics processing unit(s) (GPU)(s), over a network fabric. In one instance, a technique is disclosed for using imperfect processors as memory controllers to allow memory, which is local to the imperfect processors, to be accessed by other processors as fabric attached memory. In another instance, memory address compaction is used within the fabric elements to fully utilize the available memory space.

Techniques for an efficient fabric attached memory

Introduced herein is a routing technique that, for example, routes a transaction to a destination port over a network that supports link aggregation and multi-port connection. In one embodiment, two tables that can be searched based on the target and supplemental routing IDs of the transaction are utilized to route the transaction to the proper port of the destination endpoint. In an embodiment, the first table provides a list of available ports at each hop/route point that can route the transaction to the destination endpoint, and the second table provides a supplemental routing ID that can select a specific group of ports from the first table that can correctly route the transaction to the proper port.

Network routing using aggregated links

Multiple processors are often used in computing systems to solve very large, complex problems, such as those encountered in artificial intelligence. Such processors typically exchange data among each other via an interconnect fabric (such as, e.g., a group of network connections and switches) in solving such complex problems. The amount of data injected into the interconnect fabric by the processors can at times overwhelm the interconnect fabric preventing some of the processors from communicating with each other. To address this problem, techniques are disclosed to enable, for example, processors that are connected to an interconnect fabric to coordinate and control the amount of data injected so that the interconnect fabric does not get overwhelmed.

Techniques for reducing congestion in a computer network

An endpoint in a network may make posted or non-posted write requests to another endpoint in the network. For a non-posted write request, the target endpoint provides a response to the requesting endpoint indicating that the write request has been serviced. For a posted write request, the target endpoint does not provide such an acknowledgment. Hence, posted write requests have lower overhead, but they suffer from potential synchronization and resiliency issues. While non-posted write requests do not have those issues, they cause increased load on the network because such requests require the target endpoint to acknowledge each write request. Introduced herein is a network operation technique that uses non-posted transactions while maintaining a load overhead of the network as a manageable level. The introduced technique reduces the load overhead of the non-posted write requests by collapsing and reducing a number of the responses.

Techniques for reducing the overhead of providing responses in a computing network

Systems and techniques for synchronizing transactions between processing devices on an interconnection network are provided. Upon receiving a stream of posted transactions followed by a flush transaction from a source processing device connected to the interconnection network, the flush transaction is trapped before it enters the interconnecting network. Subsequently, based on monitoring for responses received from a destination processing device for transactions corresponding to the posted transactions, a flush response is generated and returned to the source processing device. The described techniques enable efficient synchronizing posted writes, posted atomics and the like over complex interconnection fabrics such that a first GPU can write data to a second GPU so that a third GPU can safely consume the data written to the second GPU.

Techniques for efficiently synchronizing data transmissions on a network

A plurality of registers implemented in association with a memory physical layer interface (PHY) can be used to store one or more instruction words that indicate one or more commands and one or more delays. A training engine implemented in the memory PHY can generate at-speed programmable sequences of commands for delivery to an external memory and to delay the commands based on the one or more delays. The at-speed programmable sequences of commands can be generated based on the one or more instruction words.

Memory physical layer interface logic for generating dynamic random access memory (DRAM) commands with programmable delays

In one form, an apparatus comprises a delay circuit and a controller. The delay circuit delays a plurality of command and address signals according to a first delay signal and provides a delayed command and address signal to memory interface. The controller performs command and address training in which the controller provides an activation signal and a predetermined address signal with first timing according to the first delay signal, and the plurality of command and address signals besides the predetermined address signal with second timing according to the first delay signal, wherein the second timing is relaxed with respect to the first timing. The controller determines an eye of timing for the predetermined address signal by repetitively providing a predetermined command on the command and address signals, varying the first delay signal, and measuring a data signal received from the memory interface.

Address and control signal training

A controller integrated in a memory physical layer interface (PHY) can be used to control training used to configure the memory PHY for communication with an associated external memory such as a dynamic random access memory (DRAM), thereby removing the need to provide training sequences over a data pipeline between a BIOS and the memory PHY. For example, a controller integrated in the memory PHY can control read training and write training of the memory PHY for communication with the external memory based on a training algorithm. The training algorithm may be a seedless training algorithm that converges on a solution for a timing delay and a voltage offset between the memory PHY and the external memory without receiving, from a basic input/output system (BIOS), seed information that characterizes a signal path traversed by training sequences or commands generated by the training algorithm.

Integrated controller for training memory physical layer interface

Various method and apparatus embodiments for training a delay for enabling a data strobe signal in a memory subsystem are disclosed. In one embodiment, a system includes a memory controller configured to receive a data strobe signal. The memory controller includes a training circuit. The training circuit includes a first storage circuit coupled to receive the data strobe signal on a data input and an enable signal on a clock input, and a training unit configured to, based on an output signal received from the first flip-flop, adjust a phase of the enable signal until an assertion of the enable signal coincides with a preamble indication in the data strobe signal.

Method and apparatus for memory access delay training

Growing community of inventors

Groton, MA, United States of America

Glenn A Dearth