Techniques to manage training or trained models for deep learning applications
AssigneeIntel Corporation
InventorsFrancesc GUIM BERNAT, Suraj PRABHAKARAN, Kshitij A. DOSHI, Da-Ming CHIANG
Life of this patent
Estimating the term…
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.
Claims27
Computer apparatus usable in association with artificial intelligence-related operations, the computer apparatus comprising:
- a computing platform that houses components of the computer apparatus, the components comprising:
- a plurality of circuit boards comprising circuitry-based resources for use, when computing platform is in operation, in training-related operations and inference-related operations, the circuitry-based resources comprising graphics processing unit (GPU) circuitry and central processing unit (CPU) circuitry;
- resource switch circuitry for use in direct GPU to GPU data communication among the circuitry-based resources via respective switch-related links;
- network communication and switching circuitry for use in receiving training-related data and at least one inference-related request;
- wherein:
- the training-related data is for use in association with configuring at least certain of the circuitry-based resources to implement, via the training-related operations, at least one training model;
- the training-related operations also comprise configuring, based upon training input data, the at least one training model so as to generate at least one trained model;
- the at least one trained model is to be implemented via the inference-related operations;
- the at least one training model is for use in association with at least one deep learning application;
- the inference-related operations are to be implemented based upon the at least one inference-related request to be received via the network communication and switching circuitry;
- and
- the computing platform is configured for insertion in another computing platform.
At least one machine-readable medium storing instructions executable by at least one machine comprised in a computer apparatus, the computer apparatus being usable in association with artificial intelligence-related operations, the instructions, when executed by the at least one machine, resulting in the computing apparatus being configured for performance of operations comprising:
- training-related operations and inference-related operations, the training-related operations and inference-related operations to be performed by circuitry-based resources of the computer apparatus, the circuitry-based resources being comprised in circuit boards housed by a computer platform of the computer apparatus, the circuitry-based resources comprising graphics processing unit (GPU) circuitry and central processing unit (CPU) circuitry;
- direct GPU to GPU data communication, via resource switch circuitry and respective switch-related links housed by the computer platform, among the circuitry-based resources;
- and
- receipt, via network communication and switching circuitry housed by the computer platform, of training-related data and at least one inference-related request;
- wherein:
- the training-related data is for use in association with configuring at least certain of the circuitry-based resources to implement, via the training-related operations, at least one training model;
- the training-related operations also comprise configuring, based upon training input data, the at least one training model so as to generate at least one trained model;
- the at least one trained model is to be implemented via the inference-related operations;
- the at least one training model is for use in association with at least one deep learning application;
- the inference-related operations are to be implemented based upon the at least one inference-related request received via the network communication and switching circuitry;
- and
- the computing platform is configured for insertion in another computing platform.
A method implemented using a computer apparatus, the computer apparatus being usable in association with artificial intelligence-related operations, the method comprising:
- executing training-related operations and inference-related operations, the training-related operations and inference-related operations to be performed by circuitry-based resources of the computer apparatus, the circuitry-based resources being comprised in circuit boards housed by a computer platform of the computer apparatus, the circuitry-based resources comprising graphics processing unit (GPU) circuitry and central processing unit (CPU) circuitry;
- carrying out direct GPU to GPU data communication, via resource switch circuitry and respective switch-related links housed by the computer platform, among the circuitry-based resources;
- and
- receiving, via network communication and switching circuitry housed by the computer platform, training-related data and at least one inference-related request;
- wherein:
- the training-related data is for use in association with configuring at least certain of the circuitry-based resources to implement, via the training-related operations, at least one training model;
- the training-related operations also comprise configuring, based upon training input data, the at least one training model so as to generate at least one trained model;
- the at least one trained model is to be implemented via the inference-related operations;
- the at least one training model is for use in association with at least one deep learning application;
- the inference-related operations are to be implemented based upon the at least one inference-related request received via the network communication and switching circuitry;
- and
- the computing platform is configured for insertion in another computing platform.
Dependencies are inferred from the claim text.