Google Inc.

Xiaoyu Ma

Jason Rupert Redgrave

Benjamin K Dodge

Manu Gulati

Albert Meixner

Keith Robert Pflederer

Allan Douglas Knies

Vinod Chamarty

Hongil Yoon

Rahul Jagdish Thakur

Weiping Liao

David Patterson

Aspects of the disclosure are directed to a tree-based network architecture for serving and/or training machine learning models. The architecture includes one or more multi-chip packages having a plurality of compute-memory stacks connected via an input/output (I/O) die. The I/O die includes an aggregator to aggregate computations from the compute-memory stacks. The architecture can further include a plurality of the multi-chip packages connected on a server via a server level aggregator and a plurality of the servers connected on a rack via a rack level aggregator for further aggregation of the computations from the compute-memory stacks. The tree-based network architecture allows for fewer hops, resulting in lower latency and savings in bandwidth when serving and/or training machine learning models.

Tree-based network architecture for accelerating machine learning collective operations

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating, by an image sensor of a computing device, frame data comprising sub-frames of image pixel data. A first resource of the system-on-chip provides the frame data to a second resource of the system-on-chip. The frame data is provided to the second resource using a first data path included in the system-on-chip. The first resource provides a token to the second resource using a second data path included in the system-on-chip. A processor of the system-on-chip, uses the token to synchronize production of sub-frames of image pixel data provided by the first resource to the second resource and to synchronize consumption of the sub-frames of image pixel data received by the second resource from the elastic memory buffer.

Synchronized data chaining using on-chip cache

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for a system-level cache to allocate cache resources by a way-partitioning process. One of the methods includes maintaining a mapping between partitions and priority levels and allocating primary ways to respective enabled partitions in an order corresponding to the respective priority levels assigned to the enabled partitions.

Way partitioning for a system-level cache

Growing community of inventors

Mountain View, CA, United States of America

Xiaoyu Ma