Benchmark on GPU HD7970 / LAMMPS

AMD FX-8150@4.086GHz / GPU XFX HD7970 GHz edition / MPICH 3.0.1 / Open64 4.5.2 (AMD) / APP SDK 2.8 / LAMMPS 20130121 (compiled with -O3)

Results for Rhodopsin
On the FX-8150
Single core 45.9495 seconds
4 cores 11.94 seconds
8 cores 7.06722 seconds

On the GPU HD7970 GHz edition
1 proc 4.8 seconds
2 procs 3.67 seconds

Results with FirePro3D V8800 and Tesla C2050
https://sites.google.com/site/akohlmey/news-and-announcements/gpuacceleratedlammpsonamdgpus

Results for EAM

1 procs
mpirun -n 1 ./lmp_mpich  -sf gpu -c off -v g 1 -v x 80 -v y 80 -v z 80 -v t 100 < in.eam.gpu
--------------------------------------------------------------------------
GPU 0: Tahiti, 256 cores, 2.9 GB, 1.1 GHZ (Double Precision)
--------------------------------------------------------------------------

Initializing GPU and compiling on process 0...Done.
Initializing GPU 0 on core 0...Done.

Setting up run ...
Memory usage per processor = 426.905 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -7249920            0   -6826360.6    18704.149
      50    780.81547   -7031660.9            0   -6824959.8    52291.364
     100    798.21786   -7036295.3            0   -6824987.4    51479.467
Loop time of 27.893 on 1 procs for 100 steps with 2048000 atoms

2 procs
mpirun -n 2 ./lmp_mpich  -sf gpu -c off -v g 1 -v x 80 -v y 80 -v z 80 -v t 100 < in.eam.gpu
--------------------------------------------------------------------------
GPU 0: Tahiti, 256 cores, 2.9 GB, 1.1 GHZ (Double Precision)
--------------------------------------------------------------------------

Initializing GPU and compiling on process 0...Done.
Initializing GPU 0 on core 0...Done.
Initializing GPU 0 on core 1...Done.

Setting up run ...
Memory usage per processor = 225.757 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -7249920            0   -6826360.6    18704.149
      50    780.81547   -7031660.9            0   -6824959.8    52291.364
     100    798.21786   -7036295.3            0   -6824987.4    51479.467
Loop time of 22.4219 on 2 procs for 100 steps with 2048000 atoms


mpirun -n 1 ./lmp_cuda  -sf cuda -v g 2 -v x 80 -v y 80 -v z 80 -v t 100 < in.eam.cuda

LAMMPS (21 Jan 2013)
# Using LAMMPS_CUDA
USER-CUDA mode is enabled (lammps.cpp:393)
# CUDA: Activate GPU
# Using device 0: GeForce GTX 580
Lattice spacing in x,y,z = 3.615 3.615 3.615
Created orthogonal box = (0 0 0) to (289.2 289.2 289.2)
  1 by 1 by 1 MPI processor grid
Created 2048000 atoms
# CUDA: VerletCuda::setup: Allocate memory on device for maximum of 2050000 atoms...
# CUDA: Using precision: Global: 8 X: 8 V: 8 F: 8 PPPM: 8
Setting up run ...
# CUDA: VerletCuda::setup: Upload data...
# CUDA: Total Device Memory useage post setup: 1314.117188 MB
Memory usage per processor = 416.017 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -7249920            0   -6826360.6    18704.149
      50    780.81547   -7031660.9            0   -6824959.8    52291.364
     100    798.21786   -7036295.3            0   -6824987.4    51479.467
Loop time of 20.0248 on 1 procs for 100 steps with 2048000 atoms
 

LAMMPS (CUDA/GTX580)

Uma simulação com mais de 2 milhões de atomos usando o potencial EAM em uma placa de vídeo GTX 580 no LAMMPS com CUDA.

--------------------------------------------------------------------------
RESULTADOS
--------------------------------------------------------------------------
--------------------------------------------------------------------------
Na GPU GTX 580 1.544 GHz CUDA
--------------------------------------------------------------------------
# CUDA: Activate GPU
# Using device 0: GeForce GTX 580
Lattice spacing in x,y,z = 3.615 3.615 3.615
Created orthogonal box = (0 0 0) to (289.2 289.2 285.585)
  1 by 1 by 1 MPI processor grid
Created 2022400 atoms
# CUDA: VerletCuda::setup: Allocate memory on device for maximum of 2030000 atoms...
# CUDA: Using precision: Global: 8 X: 8 V: 8 F: 8 PPPM: 8
Setting up run ...
# CUDA: VerletCuda::setup: Upload data...
# CUDA: Total Device Memory useage post setup: 1316.125000 MB
Memory usage per processor = 412.306 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -7159296            0   -6741031.1    18704.149
      50    780.71972   -6943740.2            0   -6739647.9    52292.948
     100    798.21531   -6948340.9            0     -6739675    51478.838
Loop time of 19.7667 on 1 procs for 100 steps with 2022400 atoms
--------------------------------------------------------------------------
Na CPU AMD Phenom(tm) II X6 1100T
--------------------------------------------------------------------------
  2 by 3 by 1 MPI processor grid
Created 2022400 atoms
Setting up run ...
Memory usage per processor = 130.318 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -7159296            0   -6741031.1    18704.149
      50    780.71972   -6943740.2            0   -6739647.9    52292.948
     100    798.21531   -6948340.9            0     -6739675    51478.838
Loop time of 108.489 on 6 procs for 100 steps with 2022400 atoms

Uma GTX580 usando OpenCL  é mais lento que o HD7970 para esse tipo de cálculo. O estranho é que só aparecem 256 cores no HD7970, e tinha visto em algum lugar que são 2048 para precisão simples e 512 para precisão dupla. Já o GTX580 aparecem os 512 cores.

--------------------------------------------------------------------------
GPU 0: GeForce GTX 580, 512 cores, 1.4/1.5 GB, 1.5 GHZ (Double Precision)
--------------------------------------------------------------------------

Initializing GPU and compiling on process 0...Done.
Initializing GPUs 0-1 on core 0...Done.

Setting up run ...
Memory usage per processor = 215.701 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -3537961            0   -3331264.1    18704.139
      50    780.24619   -3431376.3            0   -3330579.7    52317.163
     100    798.09276   -3433696.3            0   -3330594.2    51480.829
Loop time of 17.601 on 1 procs for 100 steps with 999424 atoms

--------------------------------------------------------------------------
GPU 0: Tahiti, 256 cores, 2.8 GB, 1.1 GHZ (Double Precision)
GPU 1: AMD FX(tm)-8150 Eight-Core Processor, 8 cores, 3.6 GHZ (Double Precision)
--------------------------------------------------------------------------

Initializing GPU and compiling on process 0...Done.
Initializing GPUs 0-1 on core 0...Done.

Setting up run ...
Memory usage per processor = 215.69 Mbytes
Step Temp E_pair E_mol TotEng Press
       0         1600     -3537961            0   -3331264.1    18704.139
      50    780.24619   -3431376.3            0   -3330579.7    52317.163
     100    798.09276   -3433696.3            0   -3330594.2    51480.829
Loop time of 14.9655 on 1 procs for 100 steps with 999424 atoms

Olhando o desmpenho da GTX580(CUDA) vs HD7970(OpenCL post anterior) parece que o ganho está no clock da GPU, só com a aparencia não dá para concluir nada, mas está faltando cores no HD7970, está, desde os primordios sempre uma ATI teve mais cores do que as GPUs da NVIDIA, e essa que tinha um clock mais elevado na briga pelo desempenho.