Project: Numerical non-linear simulations adapted to GPU technology

The project shall take advantage of the vast computational capacity of GPUs (Graphic Processing Units) in a novel software for non-linear finite element (FE) simulations. It will focus on developing platforms for running the software on a single GPU as well as multi GPUs. This capability is expected to give the software superior performance compared to using traditional High Performing Computing (HPC), which is based on using large data clusters based on CPU microprocessors. Further advantages are that the GPU technology will reduce both the hardware costs and power consumption, which is kinder to both the pocket and the environment._x000D__x000D_Today, industry relies heavily on numerical tools for efficient product development and documentation. In this regard accuracy, user-friendliness, industrial robustness and high computational speed are key elements which have to be considered for our product. _x000D__x000D_Our aim is to develop a new non-linear simulation tool for mechanical and multi-physics problems with unsurpassed accuracy, user-friendliness and industrial robustness, adapted near-100% to the latest GPU hardware technology. The new simulation tool will be based on 1) novel algorithms used for solving the fundamental mechanical equations and 2) adaptation of the code to GPU hardware technology, which is a new technology for high computational speed._x000D__x000D_No commercial finite element codes exist today dedicated to GPU technology. This is partly because GPU technology requires the use of a special programming language and a careful methodology in order to maximise performance. For this reason it is difficult to take existing codes consisting of millions of lines and adapt them to GPU technology. In our approach we start from scratch and construct a new code especially adapted to GPU technology._x000D__x000D_The current project is motivated by the excellent results gained by IMPETUS Afea AS during the last three years. In this period, IMPETUS Afea AS has been developing a novel non-linear finite element code (IMPETUS Afea Solver) which partly runs on a single GPU and the complementary element on traditional CPUs (Central Processing Units). The code itself has unique features for the solving of mechanical equations, such as the successful integration of higher order elements with the explicit integration schemes, an approach which gives computational cells (finite elements) an unsurpassed degree of accuracy for problems involving large deformations. Moreover, a unique discrete particle approach is used to model gas dynamics (air and high explosives) as well as sand materials. During 2010, IMPETUS Afea AS cooperated with SARL IMPETUS Afea, a French SME, and integrated an SPH algorithm into the GPU framework. SPH is short for "Smoothed Particle Hydrodynamics", and is used to represent compressible fluids, such as water._x000D__x000D_The overall status so far is a significant speeding up of multi-physics computations when using GPU technology. Currently, important algorithms, such as those solving the material constitutive equations as well as contact treatment between deformable parts, are not adapted to the GPU technology within the code. For the current project, we seek a near-100% adaptation of the approach to GPU technology, and thus unsurpassed computational speed in combination with low-cost hardware solutions. _x000D__x000D_The use of GPUs mounted in a standard desktop computer is an efficient and low-cost approach for SMEs to gain access to HPC resources. Owing to the investment level required, HPC has long been restricted to large companies. However in the future, these market leaders will also rely on computational scalability (use of several CPUs or GPUs for the computations). Therefore, the project will also develop technology which makes it possible to utilize multi-GPUs in numerical simulations as well as utilizing GPU technology for massive-parallel processing (MPP) computational approaches controlled by MPI (Message Passing Interface), hence utilizing several CPU cores and GPUs in parallel._x000D__x000D_The following Ps make up the consortium:_x000D__x000D_IMPETUS Afea AS (Norway - SME)_x000D_IMPETUS Afea AB (Sweden - SME)_x000D_CRF - Centro Ricerche Fiat (Italy)_x000D_NTNU - Norwegian University of Science and Technology (Norway)_x000D_NVIDIA (USA)_x000D__x000D_NVIDIA is the world's leading manufacturer of GPU hardware. The novel software solutions are developed by IMPETUS Afea AS and its daughter company IMPETUS Afea AB. This cooperation will establish the optimal technological solution between software, operating systems and hardware, and reinforce the existing professional collaboration. The 2nd aspect to consider is the accuracy of the software itself, i.e., how well the fundamental mechanical equations are solved as well as the validation of how complex material behaviour is represented in the code. This aspect is covered by NTNU as the university P. Finally, it is the role of CRF, as a typical demanding end user, to prove the industrial compatibility and robustness of the technology.

Acronym GEPEU (Reference Number: 6951)
Duration 01/02/2012 - 31/01/2015
Project Topic Our aim is to develop a new non-linear simulation tool for mechanical and multi-physics problems with unsurpassed accuracy, user-friendliness and industrial robustness adapted near-100% to computing on graphics processing units (GPU), which is a new hardware technology for high computational speed.
Project Results
(after finalisation)
1. In the project it is achieved a good performance in estimation of real phenomena. The industrial use case applied in a honeycomb structure crash simulation and in a full aluminum frontal crossbar with the crash-box, had a good correlation with experimental data coming from internal tests (TASK 1.1). _x000D_2. The FE analysis executed on the GPU server, that was purchased in the first period of the project, actually demonstrated a relevant time reduction in the modeling process. In fact, the effect was relevant and evaluated in a -75% in solution time of the FE model with 1 GPU (TASK 3.2,3.3). The difference in single GPU (MS2) and multi GPU (MS3) is less relevant and not interesting on industrial point of view. The strategy emerged from the work done therefore would be to run in parallel a set of FE analysis on 1 GPU up to the maximum number of GPUs installed in the system._x000D_3. The difference in output between the CPU and the GPU implementation of the software IMPETUS are marginal, so by this point of view, the global quality of the GPU transposition is adapt for the industrial use of the solver._x000D_4. The solver has been successfully implemented in a Mutiobjective Design Optimization cycle run, with the resulting output of an optimized design of the model tested. This approach demonstrates the possibility to operate on the software in an open way, taking to an integration with other software in a real multiphisical optimization, the actual frontier of the industrial design._x000D_
Network Eurostars
Call Eurostars Cut-Off 7

Project partner

Number Name Role Country
5 NVIDIA Corporation Partner United Kingdom
5 IMPETUS Afea AB Partner Sweden
5 IMPETUS Advanced Finite Element Analyses AS Coordinator Norway
5 Centro Ricerche Fiat S.C.p.A. Partner Italy
5 Norwegian University of Science and Technology Partner Norway