NVIDIA Tesla V100 と SDPARA

NVIDIA Tesla V100 (1枚)の性能を SDPARA 7.6.1 を用いて測定してみました。

nb = 3584 ぐらいの値が良さそうです。

[gpdpotrf] ### END n=107206, nb=1024, 1x1 procs, ver 50: 204.023sec --> 2013.060GFlops ###
[gpdpotrf] ### END n=107206, nb=1536, 1x1 procs, ver 50: 146.845sec --> 2796.903GFlops ###
[gpdpotrf] ### END n=107206, nb=2048, 1x1 procs, ver 50: 119.595sec --> 3434.169GFlops ###
[gpdpotrf] ### END n=107206, nb=2560, 1x1 procs, ver 50: 104.077sec --> 3946.207GFlops ###
[gpdpotrf] ### END n=107206, nb=3072, 1x1 procs, ver 50: 94.764sec --> 4334.026GFlops ###
[gpdpotrf] ### END n=107206, nb=3584, 1x1 procs, ver 50: 94.713sec --> 4336.360GFlops ###
[gpdpotrf] ### END n=107206, nb=4096, 1x1 procs, ver 50: 96.872sec --> 4239.726GFlops ###
[gpdpotrf] ### END n=107206, nb=4608, 1x1 procs, ver 50: 99.286sec --> 4136.625GFlops ###
[gpdpotrf] ### END n=107206, nb=5120, 1x1 procs, ver 50: 102.152sec --> 4020.567GFlops ###


◯ソフトウェア SDPARA 7.6.1

◯計算サーバ
CPU : Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz x 2
メモリ：256GB
GPU : NVIDIA Tesla V100 x1 & GeForce GTC 1080 Ti x 2
OS : CentOS 7.4