Techniques to manage training or trained models for deep learning applications

Abstract

Examples include techniques to manage training or trained models for deep learning applications. Examples include routing commands to configure a training model to be implemented by a training module or configure a trained model to be implemented by an inference module. The commands routed via out-of-band (OOB) link while training data for the training models or input data for the trained models are routed via inband links.

1 - 9 . (canceled) 
     
     
         10 . Computing system configurable for use in artificial intelligence related operations, the computing system comprising:
 training resource circuitry comprising graphics processing unit (GPU) circuitry, the training resource circuitry being configurable for use in one or more training-related operations;   inference resource circuitry comprising central processing unit (CPU) circuitry, the inference resource circuitry being configurable for use in one or more inference-related operations;   switch chip circuitry for use, when the computing system is in operation, in direct GPU to GPU communication, via respective communication links, associated with the GPU circuitry; and   network interface and switch circuitry to receive, when the computing system is in the operation, training-related data and inference request data, the training-related data being for use in configuring the training resource circuitry to implement, via the one or more training-related operations, one or more training models, the one or more inference-related operations are to be implemented based upon the inference request data;   wherein:
 the one more training models are to be configured, based upon input training data, via the one or more training operations, to generate one or more trained models; 
 the one or more inference-related operations are to implement the one or more trained models; 
 the inference request data is to request that the one or more trained models be implemented via the one or more inference-related operations; 
 the training resource circuitry and the inference resource circuitry comprise certain components; 
 different respective sets of data communication links are to be used in data transmissions involving the certain components; 
 the different respective sets of data communication links comprise one or more sets of data communication links and one or more other sets of data communication links; 
 the one or more sets of data communication links supports one or more communication bandwidths that are one or more relatively higher bandwidths compared to one or more other relatively lower bandwidths supported by the one or more other sets of data communication links; 
 the different respective sets of data communication links are to be used in data communication between or among respective subsets of the certain components; and 
 the computing system is configured for insertion in and coupling to another computing system via a backplane and/or rack. 
   
     
     
         11 . The computing system of  claim 10 , wherein:
 the switch chip circuitry is optimized for the direct GPU to GPU communication.   
     
     
         12 . The computing system of  claim 10 , wherein:
 the artificial intelligence-related operations are at least partially related to:
 facial recognition; 
 voice recognition; and/or 
 image recognition. 
   
     
     
         13 . The computing system of  claim 10 , wherein:
 the GPU circuitry comprises graphics processing units;   the CPU circuitry comprises central processing units;   the computing system comprises one or more power supplies; and   the computing system comprises PCIe communication resources.   
     
     
         14 . Computing system configurable for use in artificial intelligence-related operations, the computing system comprising:
 training resource circuitry configurable for use in one or more training-related operations, the training resource circuitry comprising artificial intelligence accelerator circuitry, the artificial intelligence accelerator circuitry comprising neural network processing accelerator circuitry and field programmable gate array circuitry;   inference resource circuitry configurable for use in one or more inference-related operations, the inference-related resource circuitry comprising central processing unit (CPU) circuitry;   direct accelerator-to-accelerator communication circuitry for use in direct accelerator to accelerator communication via respective accelerator-related links; and   network communication resource circuitry for use in receiving network-related data;   wherein:
 the training resource circuitry and the inference resource circuitry comprise certain components; 
 different respective sets of data communication links are to be used in data transmissions involving the certain components; 
 the different respective sets of data communication links comprise one or more sets of data communication links and one or more other sets of data communication links; 
 the one or more sets of data communication links supports one or more communication bandwidths that are one or more relatively higher bandwidths compared to one or more other relatively lower bandwidths supported by the one or more other sets of data communication links; 
 the different respective sets of data communication links are to be used in data communication between or among respective subsets of the certain components; and 
 the computing system is configured to be coupled via a backplane and/or rack. 
   
     
     
         15 . The computing system of  claim 14 , wherein:
 the direct accelerator-to-accelerator communication circuitry is optimized for the direct accelerator to accelerator communication.   
     
     
         16 . The computing system of  claim 14 , wherein:
 the artificial intelligence-related operations are at least partially related to:
 at least one security-related application; 
 facial recognition; 
 voice recognition; and/or 
 image recognition. 
   
     
     
         17 . The computing system of  claim 14 , wherein:
 the one or training-related operations and/or the one or more inference-related operations are to be implemented in accordance with one or more service agreements.   
     
     
         18 . The computing system of  claim 14 , wherein:
 the CPU circuitry comprises multiple central processing units;   the computing system includes one or more power supplies;   the computing system comprises PCIe communication resources; and/or   the one or more inference-related operations are to implement, at least in part, one or more trained models.