An older CAE machine gets new life with NVIDIA GPU

An older workgroup CAE server gets a second life, thanks to NVIDIA Tesla K40

Apart from being Subject Matter Experts in Structural Dynamics and Thermal Modeling, Analysis & Simulation for the Aerospace & Defense sector, TEN TECH LLC enjoys a good reputation as a reliable independent hardware partner for High Performance Computing and GPU Computing. Some of our duties include marketing material creation and benchmarking activities for companies such as NVIDIA or AMD and OEM such as Next Computing:

AMD NVIDIA Next Computing

One of our primary area of interest is GPU Computing, more specifically GPU Computing applied to CAE Software with Abaqus and MSC Nastran, two advanced FEA solvers we use in our Simulation Services division and/or resell, support and offer training for through our Software Solutions division. Dassault Systèmes and MSC Software both have invested a lot of development effort to offer solver acceleration with GPU Computing, and the results are staggering.

We have posted several benchmark results involving different types of hardware combinations representative of a Small Business environment. Our last set of results highlighted how an NVIDIA Quadro GP100 can double the performance of Abaqus 2017 on the fastest workstation available at the time:

NVIDIA Quadro GP100 Abaqus 2017

In this article, we present the results of our experiment at revitalizing one of our older workgroup server with the help of a GPU Compute Accelerator: the NVIDIA Tesla K40. With 1.43 Tflops of Double Precision power, 12GB of memory and 2,880 parallel processing CUDA cores, the K40 should have a significant impact on the solution time.

NVIDIA Tesal K40 Specs

The machine in question is an HP Z800, a high-quality and powerful machine for its time, but grossly inadequate in terms of computing power for today's demanding applications. It had been decommissioned for quite some time. As we were in need of extra computing resources for a particular project, we decided to test the hybrid solution 12-core with Tesla K40 as a viable solution instead of investing in additional hardware.

Released in 2010, our HP Z800 boasts dual Hexacore Xeon X5650 2.67GHz, 96GB of RAM and 1TB of RAID0 SSD scratch running Windows 10. The Z800 is tuned to allow for maximum computing performance: Hyperthreading off, NUMA off and memory interleaving on. It can run Abaqus and Nastran, but not very quickly.

The solver used for our evaluation is Abaqus 2017. Our test model consists of a quasi-static acceleration and thermal expansion analysis of 26M dof FEM of a jet engine:

Abaqus FEA of a jet engine

We performed a total of 6 runs, corresponding to different core usage and GPU usage:

  • 4-core, with and without GPU acceleration
  • 8-core, with and without GPU acceleration
  • 12-core, with and without GPU acceleration

The graph below summarizes the results of the 6 runs, the longer bars being naturally the longer runs. As a point of reference, the 4-core solve took about 2h30min.

NVIDIA Tesla K40 Abaqus Benchmark

As we can see, we achieve a 3x solution speedup, essentially turning our 2h30 minutes solve on a 4 core solution into a 51min affair by adding the Tesla K40. This means we can get 8 runs in a normal day by using the Tesla K40 vs. 2 without it. We typically have to run 12-16 load cases for any given model, the benefits of the K40 are clear, and the ROI for a GPU accelerator is pretty evident and quick.

Similarly, the K40 provides a 2.5x solution speedup with 8-core, and still a 2x speedup with a 12 core solution. The final and fastest run being a little bit over 30min for a 12-core solution with GPU.

In conclusion, the NVIDIA Tesla K40, allowed us to obtain reasonable solution time from a clearly outclassed and outdated computer, tripling its computing power. The Z800 has now been recommissioned is a valuable resource as a secondary workgroup server for Abaqus and Nastran.

blog comments powered by Disqus