Remote artificial intelligence (AI) acceleration system

1. A remote artificial intelligence (AI) acceleration system, comprising:

a plurality of application servers, wherein each of the plurality of application server is configured to execute AI applications over an AI software framework;

at least one artificial intelligence accelerator (AIA) appliance server configured to execute AI processing tasks in response to requests from the AI applications; and

at least one switch configured to allow connectivity between the plurality of application servers and the at least one AIA appliance server, wherein each of the at least one AIA appliance server includes:

a plurality of network-attached artificial intelligence accelerator (NA-AIA) engines; and

at least one least one AIA switch directly coupled to the plurality of NA-AIA engines and a network, wherein the at least one AIA switch allows direct communication between the plurality of application servers and the plurality of NA-AIA engines through the network,

wherein each of the plurality of NA-AIA engines is configured to execute at least one processing AI task upon a request from an AI application executed by an application server, wherein data required for execution of an AI task is directly copied, over the network, from the application server to a NA-AIA engine executing the AI task,

and wherein each of the plurality of NA-AIA engines includes an embedded network interface controller (eNIC), the eNIC is configured to connect the NA-AIA engine to the AIA switch and to implement a remote direct memory access to allow the application server to copy data directly to the NA-AIA engine over the network.

2. The system of claim 1 , wherein the at least one switch is an Ethernet switch.

3. The system of claim 1 , wherein the AI software platform is any one of: TensorFlow, Caffe2, Pytorch, and Microsoft Cognitive Toolkit (CNTK).

4. The system of claim 1 , wherein each of the plurality of application servers comprises:

a network interface card (NIC);

a processor; and

a memory coupled to the processor.

5. The system of claim 4 , wherein the at least one AIA appliance server further comprises:

a plurality of memory units, each of the plurality of memory units is coupled to a respective NA-AIA engine of the plurality of NA-AIA engines.

6. The system of claim 5 , wherein application data of an AI application is directly copied from a memory of an application server executing the AI Application to a memory associated with a respective NA-AIA engine in the AIA appliance server.

7. The system of claim 6 , wherein the application data is copied to the NA-AIA engine in response to a requested AI processing task by the AIA appliance server.

8. The system of claim 7 , wherein data of processing results is directly copied from the NA-AIA engine to the memory of the AIA appliance server executing the AI Application.

9. The system of claim 1 , wherein an AI processing task of the AI processing tasks includes an AI sequence.

10. The system of claim 5 , wherein each of the plurality of NA-AIA engines are connected to the at least one AIA switch on-demand.

11. The system of claim 1 , wherein each of the plurality of the NA-AIA engines is realized as a system on chip (SoC).

12. The system of claim 11 , wherein each of the plurality of the NA-AIA engines comprises:

a processing circuitry;

an AI sequencer; and

a memory.

13. The system of claim 12 , wherein each of the plurality of the NA-AIA engines further comprises:

an AI engine configured to control and manage a pipeline of processing AI task;

a pre-processing engine configured to preform pre-processing of functions that are not neural network related; and

a post-processing engine configured to preform post-processing of functions that are not neural network related.

14. The system of claim 1 , wherein the communication between the at least one AIA appliance server and an application server is over a network connection, wherein the network connection includes any one of: an Ethernet connection, a Transmission Control Protocol (TCP) connection, a remote direct memory access (RDMA) connection, and a RDMA over Converged Ethernet (RoCE) connection.

15. The system of claim 14 , wherein a protocol utilized to copy data over the network connection is an AI over Fabric (AIoF) protocol, wherein the AIoF protocol is a communication protocol designed to support installation and execution of AI models and AI operations.

16. A method for accelerating the operations of artificial intelligence (AI) processing tasks, comprising:

receiving, from an application server, a request to execute an AI task, the request is received at an artificial intelligence accelerator (AIA) appliance server, wherein the request is copied directly from the application server to the AIA appliance server over a network connection by means of at least one AIA switch connected in the AIA appliance server, wherein the at least one AIA switch allows direct communication between the application server and the AIA appliance server;

processing the received request by the AIA appliance server; and

copying the processing results directly to the application server over the network connection.

17. The method of claim 16 , wherein the AI task includes an AI sequence.

18. The method of claim 16 , wherein the network connection includes any one of: an Ethernet connection, a Transmission Control Protocol (TCP) connection, a remote direct memory access (RDMA) connection, and a RDMA over Converged Ethernet (RoCE) connection.

19. The method of claim 16 , wherein a protocol utilized to copy data over the network connection is an AI over Fabric (AIoF) protocol, wherein the AoF protocol is a communication protocol designed to support installation and execution of AI models and AI operations.