Software version: Abaqus 2021,No fix pack. OS version: Windows 10 PRO x64 21H2, Solver: Abaqus Standard. Machine №1: Intel(R) Xeon(R) Gold 6242R CPU @ 3.10GHz (x2, 40 cores, 80 threads), 1Tb RAM, NVIDIA RTX A5000 24Gb VRAM, Machine №2: Intel(R) Xeon(R) Gold 5218 CPU @ 2.30GHz (x2, 32 cores, 64 threads), 128Gb RAM, NVIDIA QUADRO P4000 8Gb VRAM (x2).
There are two different models to calculate. The model №1 consists of 2 686 000 nodes and 1 603 000 elements. The model №2 consists of 943 000 of nodes and 520 000 of elements. The Machine №1 is used to perform analysis of model №1 and the Machine №2 is used to perform analysis of model №2. The calculations are done in double precision.
In the table below the results of calculation of model №1 is presented:
CPUs | GPUs | Time spent (hours, minutes) | Minimum memory required (MB) | Memory to minimize i/o (MB) | |
1 | 20 | 0 | 5:41 | 10131 | 203636 |
2 | 19 | 0 | 5:39 | 10131 | 203645 |
3 | 18 | 0 | 5:40 | 10131 | 203641 |
4 | 20 | 1 | 8:49 | 10131 | 203625 |
5 | 20 | 2 | 8:45 | 10131 | 203641 |
6 | 20 | 3 | 8:46 | 10131 | 203639 |
7 | 20 | 10 | 9:41 | 10131 | 207617 |
8 | 40 | 0 | 6:38 | 10131 | 207596 |
9 | 38 | 0 | 6:34 | 10131 | 207598 |
10 | 10 | 0 | 8:40 | 10131 | 207660 |
In the table below the results of calculation of model №2 is presented:
CPUs | GPUs | Time spent (hours, minutes) | Minimum memory required (MB) | Memory to minimize i/o (MB) | |
1 | 16 | 0 | 2:44 | 5244 | 86945 |
2 | 16 | 1 | 5:02 | 5244 | 86935 |
3 | 16 | 2 | 3:33 | 5244 | 86939 |
4 | 32 | 2 | 3:34 | 5244 | 86952 |
According to the results above the usage of GPU acceleration slows down the simulation. Also the number of GPUs refers to number of graphics cards instead of GPU cores as we expected. And GPU acceleration depends on VRAM used: 1 GPU uses 8Gb of VRAM, 2 GPUs use 16Gb of VRAM (see table of model №2 results, time spent).
The format of command used for launch the calculations (for example, with 16 CPUs and 2 GPUs):
“abq2021 job=test output_precision=full input=test cpus=16 gpus=2”
We followed recommendations of Abaqus White pages for GPU acceleration. However, have not succeed so far. Could you please advise what should we do to see GPU acceleration?
