Megh Computing, Inc.

Intel Corporation

Duncan Moss

Martin Langhammer

Gregg William Baeckler

Dipankar Das

Sasikanth Avancha

Sudarshan Srinivasan

Suchit Subhaschandra

Jonathan Beare

A computing system, including a processor configured to, at development time, receive a machine learning model topology including a plurality of layers. The processor may be further configured to generate an internal representation graph of the machine learning model topology. The internal representation graph may include a plurality of internal representation layers. By performing one or more modifications to the internal representation graph, the processor may be further configured to generate an inferencer graph including a plurality of inferencer layer blocks. Each inferencer layer block may indicate an input buffer size, a logic function, and an output buffer size. At deployment time, the processor may be further configured to transmit, to a plurality of processing devices, instructions to implement the machine learning model topology with the respective input buffer sizes, logic functions, and output buffer sizes selected for the plurality of inferencer layer blocks of the inferencer graph.

Inferencer graph for implementing machine learning model topology

A computing system is provided, including a processor configured to receive a directed acyclic graph (DAG) template specifying a data pipeline of a plurality of processing stages. For each processing stage, the processor may be further configured to select a respective processing device of a plurality of communicatively linked processing devices. The processor may be further configured to determine a routing sequence between the plurality of processing devices according to the DAG template. The processor may be further configured to transmit a plurality of input packets encoding the plurality of processing stages to the respective processing devices selected for the processing stages as specified by the routing sequence. In response to transmitting the plurality of input packets, the processor may be further configured to receive, from a processing device of the plurality of processing devices, one or more output packets encoding a processing result of the data pipeline.

Directed acyclic graph template for data pipeline

The present disclosure relates generally to techniques for improving the implementation of certain operations on an integrated circuit. In particular, deep learning techniques, which may use a deep neural network (DNN) topology, may be implemented more efficiently using low-precision weights and activation values by efficiently performing down conversion of data to a lower precision and by preventing data overflow during suitable computations. Further, by more efficiently mapping multipliers to programmable logic on the integrated circuit device, the resources used by the DNN topology to perform, for example, inference tasks may be reduced, resulting in improved integrated circuit operating speeds.

Circuitry for low-precision deep learning

Growing community of inventors

Hillsboro, OR, United States of America

Duncan Moss