14–18 Jul 2015
Kobe International Conference Center
Asia/Tokyo timezone

Performance of Maxwell GPUs and Optimization of Non-Perturbative Renormalization codes.

15 Jul 2015, 18:30
2h 30m
4th floor (Kobe International Conference Center)

4th floor

Kobe International Conference Center

6-9-1 Minatojima-nakamachi, Chuo-ku, Kobe, Hyogo 650-0046, Japan
Board: 31
Poster Algorithms and Machines Poster Session

Speaker

Mr Jeonghwan Pak (Seoul National University)

Description

Matching factors for the four-fermion operators relevant to B_K are obtained using the Non-Perturbative Renormalization(NPR) method in the RI-MOM scheme. Our NPR codes are calculated in Double Precision(DP) and optimized on Fermi GPU. NVIDIA released The GTX Titan X based on the Maxwell architecture. The Single Precision(SP) calculation performance of the GTX Titan X is improved, but the ratio of DP performance to SP performance is low. Hence, We optimize NPR codes on GTX Titan X and measure its scalability on Multi GPUs. And, we measure the performance of GTX Titan X using our conjugate gradient(CG) solver.

Primary authors

Mr Jeonghwan Pak (Seoul National University) Prof. Weonjong Lee (Seoul National University)

Co-authors

Mr Hwancheol Jeong (Seoul National University) Mr Jangho Kim (Seoul National University) Mr Sangbaek Lee (Seoul National university)

Presentation materials