BROWSE
Detail
Conference Paper
고집적 GPU 서버의 데이터 병렬 딥러닝 성능 및 확장성
Cited - time in 
Share 
Share
- Issue Date
- 2022-07
- Citation
- 대한전자공학회 학술 대회 (하계) 2022, pp.2072-2075
- Publisher
- 대한전자공학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- In this paper, we investigate and execute the performance and scalability analysis for deep learning application especially on a highly integrated many-GPU server environment. In general, a conventional GPU server can equip four to ten GPUs in a server chassis. There are many restrictions to hider equipping many GPUs in a server for example, physical chassis dimension, number of PCIe connections, and etc. In this paper, we integrate the GPU scalability in a server over ten GPUs per server by utilizing proprietary external PCIe expansion hardware, and investigate GPU scale-up performance by applying a deep learning application. The implementation and experimentation result show that a GPU server can equip up to twenty-six GPUs - the total number of GPUs in a server is limited by BIOS capability, and its performance scaled up linearly in a server.
- KSP Keywords
- Deep learning application, Performance and scalability, Scalability Analysis, Scale-up, deep learning(DL)