Distributed AI training topology based on flexible cable connection
A data processing system includes a central processing unit (CPU) and accelerator cards coupled to the CPU over a bus, each of the accelerator cards having a plurality of data processing (DP) accelerators to receive DP tasks from the CPU and to perform the received DP tasks. At least two of the accelerator cards are coupled to each other via an inter-card connection, and at least two of the DP accelerators are coupled to each other via an inter-chip connection. Each of the inter-card connection and the inter-chip connection is capable of being dynamically activated or deactivated, such that in response to a request received from the CPU, any one of the accelerator cards or any one of the DP accelerators within any one of the accelerator cards can be enabled or disabled to process any one of the DP tasks received from the CPU.
1. A data processing system, comprising:
a central processing unit (CPU); and
a plurality of accelerator cards coupled to the CPU over a bus, each of the accelerator cards having a plurality of data processing (DP) accelerators to receive DP tasks from the CPU and to perform the received DP tasks, wherein at least two of the accelerator cards are coupled to each other via an inter-card connection, wherein at least two of the DP accelerators are coupled to each other via an inter-chip connection, wherein each of the plurality of DP accelerators is in least two accelerator rings;
wherein each of the inter-card connection is capable of being dynamically activated or deactivated, such that in response to a request received from the CPU, any one of the accelerator cards can be enabled or disabled to process any one of the DP tasks received from the CPU.
2. The system of claim 1 , wherein each of the DP accelerators of each of the accelerator cards includes a plurality of inter-chip interfaces, which can be utilized to interconnect with another one of the DP accelerators of the accelerator card via a respective inter-chip connection.
3. The system of claim 1 , wherein each of the DP accelerators of each of the accelerator cards includes a plurality of inter-chip interfaces, which can be utilized to interconnect with another one DP accelerator of another accelerator card via a respective inter-card connection.
4. The system of claim 1 , wherein the DP accelerators in each of the accelerator cards are arranged in a plurality of rows and columns coupled to each other via one or more inter-chip connections.
5. The system of claim 4 , wherein each row of the DP accelerators of a first of the accelerator cards are coupled in series via respective horizontal inter-chip connections, forming a horizontal accelerator ring.
6. The system of claim 5 , wherein each column of the first accelerator card is coupled to a corresponding column of a second of the accelerator cards via one or more respective inter-card connections, forming a vertical accelerator ring.
7. The system of claim 4 , wherein each column of the DP accelerators of a first accelerator card are coupled in series via respective vertical inter-chip connections, forming a vertical accelerator ring.
8. The system of claim 7 , wherein each row of the first accelerator card is coupled to a corresponding row of a second of the accelerator cards via one or more respective inter-card connections, forming a horizontal accelerator ring.
9. The system of claim 1 , wherein each of the DP accelerators comprises an artificial intelligence (AI) accelerator chip.
10. The system of claim 1 , wherein the bus comprises a peripheral component interconnect express (PCIe) link or an Ethernet connection, and wherein an inter-chip connection or an inter-card connection comprises a cache coherent interconnect for accelerators (CCIX) link.
11. An accelerator card, comprising:
a host interface to be coupled to a central processing unit (CPU) over a bus; and
a plurality of data processing (DP) accelerators to receive DP tasks from the CPU and to perform the received DP tasks, wherein at least two of the DP accelerators are coupled to each other via an inter-chip connection, wherein each of the plurality of DP accelerators is in least two accelerator rings;
wherein each inter-chip connection is capable of being dynamically activated or deactivated, such that in response to a request received from the CPU via the host interface, any one of the DP accelerators can be enabled or disabled to process any one of the DP tasks received from the CPU.
12. The accelerator card of claim 11 , wherein each of the DP accelerators includes a plurality of inter-chip interfaces, which can be utilized to interconnect with another one of the DP accelerators via a respective inter-chip connection.
13. The accelerator card of claim 11 , wherein each of the DP accelerators includes a plurality of inter-chip interfaces, which can be utilized to interconnect with another one DP accelerator of another accelerator card via a respective inter-card connection.
14. The accelerator card of claim 11 , wherein the DP accelerators are arranged in a plurality of rows and columns coupled to each other via one or more inter-chip connections.
15. The accelerator card of claim 14 , wherein the DP accelerators of each row are coupled in series via respective horizontal inter-chip connections, forming a horizontal accelerator ring.
16. The accelerator card of claim 15 , wherein each column is coupled to a corresponding column of a second of the accelerator cards via one or more respective inter-card connections, forming a vertical accelerator ring.
17. The accelerator card of claim 14 , wherein the DP accelerators of each column are coupled in series via respective vertical inter-chip connections, forming a vertical accelerator ring.
18. The accelerator card of claim 17 , wherein each row is coupled to a corresponding row of a second of the accelerator cards via one or more respective inter-card connections, forming a horizontal accelerator ring.
19. The accelerator card of claim 11 , wherein each of the DP accelerators comprises an artificial intelligence (AI) accelerator chip.
20. The accelerator card of claim 11 , wherein the bus comprises a peripheral component interconnect express (PCIe) link or an Ethernet connection, and wherein an inter-chip connection or an inter-card connection comprises a cache coherent interconnect for accelerators (CCIX) link.