□ ABCI System Outline
□ Block Diagram of the Computing Node
|FUJITSU PRIMERGY Server( 2 servers in 2U)|
|CPU||Intel Xeon Gold 6148(27.5M Cache, 2.40 GHz, 20 Core)×2|
|GPU||NVIDIA Tesla V100(SXM2)×4|
|Local Storage||1.6TB NVMe SSD(Intel SSD DC P4600 u.2)×1|
□ Computing Node
Each computing node accommodates: two Intel Xeon Gold 6148, four Nvidia Tesla V100 connected by SXM2, 1.6TB SSD connected with NVMe, and 384GiB main memory.
The theoretical performance of the computing node is 506 AI-TFLOPS for half precision (required for AI machine learning) and 34.2 TFLOPS for double precision (required for Engineering science and technology calculations).
□ High-Speed Interconnect
InfiniBand EDR is used for interconnecting among the computing nodes, the multi-platform nodes, the interactive nodes, and the large-capacity storage systems. 34 computing nodes are mounted on a rack. There are 32 racks for a total of 1,088 nodes. The interconnection within a rack is full bisection, while the interconnection between the 32 racks are over subscripted (1/3 bandwidth).
□ Large-Capacity Storage
22PB Large-capacity storage system for handling AI and Big Data consists of three sets of DDN SFA14KX. GPFS is supported by DDN GRIDScaler.
□ Interconnection Network
Since ABCI is connected to SINET-5(100Gbps), ABCI users may access ABCI through the internet. The connection is secured by firewalls (FortiGate 1500D) and two-stage authentication is adopted.