Centralized-distributed mixed organization of shared memory for neural network processing
AssigneeAlibaba Group Holding Limited
InventorsLiang HAN, Xiaowei JIANG, Jian CHEN
Life of this patent
Estimating the term…
Abstract
The present disclosure provides a processor providing a memory architecture having M-number of processing elements each having at least N-number of processing units and a local memory. The processor comprises a first processing element of the M-number of processing elements comprising a first set of N-number of processing units configured to perform a computing operation, and a first local memory configured to store data utilized by the N-number of processing units. The processor further comprises a data hub configured to receive data from the M-number of processing elements and to provide shared data to each processing element of the M-number of processing elements.
Claims22
A processor, comprising:
- a memory architecture having M-number of processing elements each having at least N-number of processing units and a local memory, wherein M, N comprising integers greater than 1 and the memory architecture comprising:
- a first processing element of the M-number of processing elements comprising:
- a first set of N-number of processing units configured to perform a computing operation, and
- a first local memory configured to store data capable of being utilized by the N-number of processing units;
- and
- a data hub capable of receiving data from the M-number of processing elements and capable of providing shared data from one of the M-number of processing elements to other processing elements of the M-number of processing elements.
A method performed by a centralized data hub capable of communicating with M-number of processing elements each having at least N-number of processing units and a local memory, the method comprising:
- acquiring data from a processing element of M-number of processing elements;
- determining a distribution pattern for the acquired data;
- and
- distributing the acquired data to at least one or more of the M-number of processing elements using the distribution pattern.
A method performed by a processing element of multiple processing elements coupled to a centralized data hub, the method comprising:
- acquiring shared data from the centralized data hub;
- acquiring private data local to the processing element;
- storing the private data in a local memory of the of the processing element;
- providing the shared data and the stored private data as operands to one or more processing units of the processing element;
- and
- generating one or more results from one or more operations performed by corresponding one or more processing units.
Dependencies are inferred from the claim text.