US 12,271,326 B2
Data flow-based neural network multi-engine synchronous calculation system
Li Jiao, Guangdong (CN); Yuanchao Li, Guangdong (CN); Kuen Hung Tsoi, Guangdong (CN); and Xinyu Niu, Guangdong (CN)
Assigned to Shenzhen Corerain Technologies Co., Ltd., Shenzhen (CN)
Appl. No. 18/011,528
Filed by Shenzhen Corerain Technologies Co., Ltd., Guangdong (CN)
PCT Filed Jun. 4, 2021, PCT No. PCT/CN2021/098289
§ 371(c)(1), (2) Date Dec. 20, 2022,
PCT Pub. No. WO2021/259040, PCT Pub. Date Dec. 30, 2021.
Claims priority of application No. 202010574436.7 (CN), filed on Jun. 22, 2020.
Prior Publication US 2023/0305976 A1, Sep. 28, 2023
Int. Cl. G06F 13/28 (2006.01); G06F 12/0897 (2016.01)
CPC G06F 13/28 (2013.01) [G06F 12/0897 (2013.01)] 7 Claims
OG exemplary drawing
 
1. A data flow-based neural network multi-engine synchronous calculation system, comprising:
a plurality of calculation engines, wherein each calculation engine comprises a plurality of calculation modules and at least one cache module located at different layers, and wherein each of the plurality of calculation modules is configured to calculate an input calculation graph provided by the at least one cache module or by the calculation module of a preceding layer of the same calculation engine, to obtain an output calculation graph; and
at least one synchronization module, wherein each synchronization module is configured to monitor a data amount of the input calculation graph stored by each cache module of the same layer across the plurality of calculation engines, and control, when the data amount of the input calculation graph stored in each cache module reaches a preset value corresponding to that cache module, each cache module on the same layer to output the stored input calculation graph to the calculation module in a next layer of the same calculation engine, wherein the at least one synchronization module is configured to monitor the data amounts of the input calculation graphs stored by a first-layer cache module of the at least one cache module and a second-layer cache module of the at least one cache module, and when the difference between the data amounts of the input calculation graphs stored by the first-layer cache module and the second-layer cache module is within a respective preset range, data flows are synchronous.