An Abaqus 2016 GPU Computing Benchmark

A real life benchmark of Abaqus 2016 GPU Computing
While preparing an Abaqus pricing configuration for one of our customers, we wanted to have a clear view of how much benefit GPU Computing in Abaqus 2016 would bring. The goal of our study was to make sure we delivered the best bang for the buck in terms of solver execution speed to our customer.

Unlike a lot of the benchmark data published using enormous dynamics models, we wanted to create a more "down to earth" one that would be more representative of what our customers could use Abaqus for. These published benchmarks are a great illustration of the incredible scalability of Abaqus (one example can be viewed here), but do I gain anything at all if I don't have a 40M dof model with 2500 modes to calculate?

We put together a relatively small model (3M dof) and went on to test both linear statics and non-linear statics analysis with Abaqus 2016.

Abaqus NVIDIA GPU COmputing

The hardware used is a modest dual 6-core Xeon HP Z800 workstation with 96GB of RAM and an nvidia Quadro 4000. Nothing in this machine screams "state of the art": it is a (very good) 5 year old machine, with a (very good) 2 year old graphics card. Several configurations were tried, 4-core, 8-core, 12-core with and without GPU Computing. Below is a graph of the first results, by configuration:

ABAQUS Linear Statics GPU

First conclusion: Abaqus is pretty darn fast! But we knew that already. Second conclusion: Abaqus scales pretty well with the number of cores. Even with a 4 minute solve time, we managed to shorten it by 30% going from 4-core to 8-core solution! We also knew that already…

But here comes the SIMULIA magic, all the way from Providence, RI: adding our NVidia Quadro 4000 to the mix via Abaqus' GPU Computing and we're able to shorten the 4-core solution time by 33%!

Being able to shrink a 4 minutes solve by 33% is pretty remarkable, but is it really that much of a gain? Not exactly… Hence the next benchmark.

Using the same hardware, we this time ran the model with large displacement non-linear settings, and enforced time steps to make sure we'd get a fair amount of computation time to look at. The results are even more impressive in favor of GPU Computing:


We were able to shorten the solution time by 43% by using an 8-core solution instead of 4-core, that's already great. But, even better, by using GPU computing along with our 4-core solution, we're able to shorten the calculation time by 49%.

The conclusion to draw from all of this is that given the impressive performance of Abaqus 2016 GPU Computing, this is certainly a great way of getting the most bang for your buck for the Abaqus token-conscious as GPU Computing requires a fraction of the tokens a multi-core solution requires.

blog comments powered by Disqus