Technical Report - GPU acceleration 01. What is GPU operation? 02. GPU operation in midas Gen/Civil 03. Recommended Sys
Views 266 Downloads 9 File size 457KB
Technical Report - GPU acceleration
01. What is GPU operation? 02. GPU operation in midas Gen/Civil 03. Recommended System Environment
MIDAS Information Technology Co. Ltd, Copyrights © since 2000 All rights reserved
01.
What is GPU operation?
GPU is a device related to the graphical process in computer. Due to the difference of usage purpose, CPU and GPU have been developed in a different direction. CPU has been developed to improve each core capacity whereas GPU has been developed to operate simple calculation in parallel with multiple cores. As per the recent development of computer technology, one GPU can handle up to 2,000 cores. Each GPU chip has capacity equal to one cluster.
02.
GPU operation in midas Gen/Civil In case of GPU, single precision operation is very fast for graphical process. For example, in case of GeForce,
operation speed of single precision is 12~24 times faster than the operation speed of double precision. As shown in the figure below, in case of Tesla graphic card, operation speed of single precision is 2~3 times faster than that of double precision. In case of midas Gen/Civil, double precision operation is frequently used. Therefore Tesla graphic card which has high performance in double precision is recommended. In order to use GPU solver, high-performance GPU is required. GPU has been applied in Multi Frontal Solver (MFS) of midas Gen/Civil in order to accelerate the calculation of simultaneous equations which is highest proportion in FEM analysis. As GPU maximize their capacity with large number of cores, remarkable different analysis time can be obtained with larger stiffness matrix.
[GPU performance graph – Theoretical performance]
http://en.MidasUser.com
1
[GPU structure: smallest box represents a core]
03.
Recommended System Environment Table 1. Recommended Computer Capacity Type
Capacity
GPU Model
Tesla C20xx, Tesla K20
CPU
Quad Core or more
RAM
Larger memory than GPU memory(Recommendation: 6GB or more)
Driver
Latest Nvidia graphics driver
Misc.
1 GPU for display output + 1 GPU for calculation
64 bit
GPU operation is valid in midas Gen/Civil n 64-bit version. Nvidia GPU is supported. Using Nvidia control panel, CUDA-GPU can be controlled by graphic card. As shown in the figure below, other GPU should be checked off in CUDA setting window in order to prevent operation error in GPU used in display. In order to increase capability, quad core or larger CPU with larger memory than GPU are recommended. In case of Tesla, additional option (ECC) increasing result accuracy and stabilizing the operation is available. If this option is applied, the computer will use additional memory (about 700Mbyte) to find data error. Telsa and Quadro GPU has high-performance, however their price is quite expensive. Even though there are various GPU is GeForce, GPU which supports double precision is quite limited. Therefore, in order to obtain high-performance, Tesla GPU is recommended.
http://en.MidasUser.com
2
http://en.MidasUser.com
3