Google Inc.

Liqun Cheng

Parthasarathy Ranganathan

Rama Krishna Govindaraju

David Lo

Norman Paul Jouppi

Sheng Li

Nishant P Patil

Richard Yoo

Benjamin Charles Serebrin

Quoc V Le

Michael Roger Marty

Andrew Gallatin

Yun Ni

Brian Zhang

Haishan Zhu

Sundar Jayakumar Dev

Martin Guy Dixon

Sridhar Lakshmanamurthy

Ruoming Pang

Mingxing Tan

Daniel Stodolsky

Andrew Li

Erik Karl Norden

Aspects of the disclosure are directed to a heterogeneous machine learning accelerator system with compute and memory nodes connected by high speed chip-to-chip interconnects. While existing remote/disaggregated memory may require memory expansion via remote processing units, aspects of the disclosure add memory nodes into machine learning accelerator clusters via the chip-to-chip interconnects without needing assistance from remote processing units to achieve higher performance, simpler software stack, and/or lower cost. The memory nodes may support prefetch and intelligent compression to enable the use of low cost memory without performance degradation.

Heterogeneous ML accelerator cluster with flexible system resource balance

The subject matter described herein provides systems and techniques to address the challenges of growing hardware and workload heterogeneity using a Warehouse-Scale Computer (WSC) design that improves the efficiency and utilization of WSCs. The WSC design may include an abstraction layer and an efficiency layer in the software stack of the WSC. The abstraction layer and the efficiency layer may be designed to improve job scheduling, simplify resource management, and drive hardware-software co-optimization using machine learning techniques and automation in order to customize the WSC for applications at scale. The abstraction layer may embrace platform/hardware and workload diversity through greater coordination between hardware and higher layers of the WSC software stack in the WSC design. The efficiency layer may employ machine learning techniques at scale to realize hardware/software co-optimizations as a part of the autonomous WSC design.

Autonomous warehouse-scale computers

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for determining an architecture for a task neural network that is configured to perform a particular machine learning task on a target set of hardware resources. When deployed on a target set of hardware, such as a collection of datacenter accelerators, the task neural network may be capable of performing the particular machine learning task with enhanced accuracy and speed.

Hardware-optimized neural architecture search

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for scheduling operations represented as a computational graph on a distributed computing network. A method includes: receiving data representing operations to be executed in order to perform a job on a plurality of hardware accelerators of a plurality of different accelerator types; generating, for the job and from at least the data representing the operations, features that represent a predicted performance for the job on hardware accelerators of the plurality of different accelerator types; generating, from the features, a respective predicted performance metric for the job for each of the plurality of different accelerator types according to a performance objective function; and providing, to a scheduling system, one or more recommendations for scheduling the job on one or more recommended types of hardware accelerators.

Generating recommendations based on performance metrics for scheduling jobs on distributed computing devices

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, are described for performing asymmetric data communication at a host-device interface of a system. The methods include identifying devices coupled to a host of the system and generating a system topology that identifies a connectivity of the devices and identifies bus lanes that enable data transfers at the system. The host determines that a first connection between the host and a first device of the multiple devices has an asymmetric bandwidth requirement. The host configures a set of bus lanes of a data bus connecting the first device and the host to allocate a different number of the bus lanes to data egress from the host than to data ingress to the host. The bus lanes are configured to allocate the differing number of bus lanes based on the asymmetric bandwidth requirement of the first connection.

Asymmetric data communication for host-device interface

Methods, systems, and computer storage media storing instructions for managing processing system efficiency. One of the methods includes obtaining data splitting a plurality of general-purpose processing units in a processing system into a high-priority domain and a low-priority domain, wherein the general-purpose processing units in the high-priority domain are assigned to perform one or more tasks comprising one or more high-priority tasks, and the general-purpose processing units in the low-priority domain are assigned to perform one or more low-priority tasks; and during runtime of the processing system, obtaining memory usage measurements that characterize usage of system memory by the high-priority domain and the low-priority domain; and adjusting, based on the memory usage measurements, a configuration of (i) the high-priority domain, (ii) the low-priority domain, or (iii) both to adjust utilization of the system memory by the general-purpose processing units.

Managing processing system efficiency

Recommendations for scheduling jobs on distributed computing devices

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for caching data not frequently accessed. One of the methods includes receiving a request for data from a component of a device, determining that the data satisfies an infrequency condition, in response to determining that the data satisfies the infrequency condition: determining a target cache level which defines a cache level within a cache level hierarchy of a particular cache at which to store infrequently accessed data, the target cache level being lower than a highest cache level in the cache level hierarchy, requesting and receiving the data from a memory that is not a cache of the device, and storing the data in a level of the particular cache that is at or below the target cache level in the cache level hierarchy, and providing the data to the component.

Data caching

An example method includes during execution of a software application by a processor, receiving, by a copy processor separate from the processor, a request for an asynchronous data copy operation to copy data within a memory accessible by the copy processor, wherein the request is received from a copy manager accessible by the software application in a user space of an operating system managing execution of the software application; in response to the request, initiating, by the copy processor, the asynchronous data copy operation; continuing execution of the software application by the processor; determining, by the copy processor, that the asynchronous data copy operation has completed; and in response to determining that the asynchronous copy operation has completed, selectively notifying, by the copy processor, the software application that the asynchronous copy operation has completed.

Asynchronous copying of data within memory

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for pre-fetching data. The methods, systems, and apparatus include actions of providing a request for data to an input-output device and receiving a set of memory addresses for the requested data. Additional actions include determining a subset of the memory addresses, providing a request for a processor to pre-fetch or inject data corresponding to the subset of the memory addresses, and receiving the requested data and the set of memory addresses. Additional actions include determining that the received data includes data for the subset of memory addresses that has been requested to be pre-fetched or injected, storing the data for the subset of memory addresses in a cache of the processor, and storing remaining data of the received data for the memory addresses in a main memory.

Controlled cache injection of incoming data

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for machine level resource distribution are disclosed. In one aspect, a method is implemented in a data processing apparatus, which includes, for each server computer in a set of two or more server computers within a data center, wherein each server computer includes a plurality of processing cores, receiving wear data describing, for each processing core of the server computer, a wear level for the processing core that is indicative of accumulated wear of the processing core, and moderating accumulation of wear in the processor cores based on the wear level of the processing cores from at least two different server computers.

Growing community of inventors

San Jose, CA, United States of America

Liqun Cheng