Google Inc.

Sun Microsystems, Inc.

Oracle America, Inc.

James Laudon

Shen Wang

Jeffrey Adgate Dean

Quoc V Le

Chian-Min Richard Ho

Azalia Mirhoseini

Anna Darling Goldie

Wenjie Jiang

Mustafa Nazim Yazgan

Roger David Carpenter

Young-Joon Lee

Ebrahim Songhori

Omkar Pathak

Kavya Srinivasa Setty

Luiz Andre Barroso

Sanjay Patel

Michael Roger Marty

Nicolai Kosche

Adam R Talcott

Ramaswamy Sivaramakrishnan

Sunil Kumar Vemula

Sudip Roy

Hanxiao Liu

Farnad Sajjadian

Curtis Roger McAllister

Qiumin Xu

Yanqi Zhou

Amirali Abdolrashidi

Daniel Lin-Kit Wong

Chao Ma

Phitchaya Mangpo Phothilimthana

Aspects of the disclosure are directed to automatically determining floor planning in chips, which factors in memory macro alignment. A deep reinforcement learning (RL) agent can be trained to determine optimal placements for the memory macros, where memory macro alignment can be included as a regularization cost to be added to the placement objective as a RL reward. Tradeoffs between the placement objective and alignment of macros can be controlled by a tunable alignment parameter.

Alignment cost for integrated circuit placement

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for compiler optimizations using a compiler optimization network. One of the methods includes receiving an input program, wherein the input program defines a graph of operation modules, wherein each node in the graph is a respective operation module, and each edge between nodes in the graph represents one operation module receiving the output generated by another operation module. The input program is processed by a compiler optimization network comprising a graph-embedding network that is configured to encode operation features and operation dependencies of the operation modules of the input program into a graph embedding representation and a policy network that is configured to generate an optimization action for each of one or more nodes encoded in the graph embedding representation. The compiler optimization network generates an output optimization plan comprising one or more optimization actions for the input program.

Learned graph optimizations for compilers

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating a computer chip placement. One of the methods includes obtaining netlist data for a computer chip; and generating a computer chip placement, comprising placing a respective macro node at each time step in a sequence comprising a plurality of time steps, the placing comprising, for each time step: generating an input representation for the time step; processing the input representation using a node placement neural network having a plurality of network parameters, wherein the node placement neural network is configured to process the input representation in accordance with current values of the network parameters to generate a score distribution over a plurality of positions on the surface of the computer chip; and assigning the macro node to be placed at the time step to a position from the plurality of positions using the score distribution.

Generating integrated circuit placements using neural networks

A method includes performing one or more operations as requested by a thread executing on a processor, the thread having a thread context; receiving a park request from the thread, the park request received following a request from the thread for a low latency resource, wherein the cache response time is less than or equal to a resource response threshold so as to allow the thread context to be stored and retrieved from the cache in less time than the portion of time it takes to complete the request for the low latency resource; storing the thread context in the cache; detecting that the resume condition has occurred; retrieving the thread context from the cache; and resuming execution of the thread.

Low latency thread context caching

Methods and systems are provided for managing thread execution in a processor. Multiple instructions are fetched from fetch queues. The instructions satisfy the condition that they involve fewer bits than the integer processing pathway that is used to execute them. The instructions are decoded, and divided into groups. The instructions are processed simultaneously through the pathway, such that part of the pathway is used to execute one group of instructions and another part of the pathway is used to execute another group of instructions. These parts are isolated from one another so the execution of the instructions can share the pathway and execute simultaneously and independently.

Doubling thread resources in a processor

A system includes a bus, a processor operably coupled to the bus, a memory operably coupled to the bus, a plurality of input/output (I/O) devices operably coupled to the bus, where each of the I/O devices has a set of control registers, and a first shared I/O unit operably coupled to the bus. The first shared I/O unit has a plurality of shared functions and is configured to perform the shared functions, where the shared I/O functions are not included as functions on the I/O devices and the I/O devices and the processor interact with the first shared I/O unit to use one or more of the shared functions performed by the first shared I/O unit.

Shared input/output (I/O) unit

A system and method for profiling runtime system events of a computer system may include associating a data source type with detected system events. The system events may be detected dependent on information included in a reply message received by a processor in response to a data request or other transaction request message. The reply message may include information characterizing a source type of a source of data included in the reply message. The source type information may indicate that the source is remote or local; that it is a shared or a private storage location; that the data is supplied via a cache-to-cache transfer; or that the data is sourced from a coherency domain other than that of the requesting process. Instructions, events, messages, and replies may be sampled, and extended address information corresponding to the samples may be stored in an event set database for performance analysis.

Apparatus and method for profiling system events in a fine grain multi-threaded multi-core processor

A memory controller including a control unit for limiting the number of memory requests that are executed within a predetermined time period to regulate power consumption. The control unit may determine a memory request limit indicating the maximum number of memory requests that are allowed to be executed during the predetermined time period based on at least a carry-over limit and a new request limit. The carry-over limit may indicate the maximum number of carry-over memory requests that are allowed during the predetermined time period. The new request limit may indicate the maximum number of new memory requests that are allowed during the predetermined time period. The control unit may further control the number of memory requests that are executed in each of a sequence of predetermined time periods.

System and method for metering requests to memory

Software-based cache coherence protocol. A processing unit may execute a memory request using a processor thread. In response to detecting a cache hit to shared or a cache miss associated with the memory request, a cache may provide both a trap signal and coherence information to the processor thread of the processing unit. After receiving the trap signal and the coherence information, the processor thread may perform a cache coherence operation for the memory request using at least the received coherence information. The processing unit may include a plurality of processor threads and a load balancer. The load balancer may receive coherence requests from one or more remote processing units and distribute the received coherence requests across the plurality of processor threads. The load balance may preferentially distribute the received coherence requests across the plurality of processor threads based on the operation state of the processor threads.

System and method for efficient software cache coherence

A system is provided for detecting when a temperature of a multiprocessor chip approaches an established threshold temperature indicating an imminent overheat condition. When the threshold temperature is reached, a number of active threads are idled in order to remove their contribution from the overall power consumption of the multiprocessor chip. Idling of the threads serves to prevent the multiprocessor chip from reaching the overheat condition. Once the temperature of the multiprocessor chip drops to an acceptable level, execution of the previously idled threads is resumed. Detection of the imminent overheat condition and corresponding idling of the threads to avoid reaching the overheat condition is conducted by hardware to ensure timely reduction of the multiprocessor chip temperature.

Method and apparatus for controlling power consumption in multiprocessor chip

Growing community of inventors

Madison, WI, United States of America

James Laudon