ANSYS AND NVIDIA: CLOSING THE GAP BETWEEN CAE AND GAMES?
By September 7th, ANSYS released its new Product: Discovery Live. And I suppose many - just like me - got impressed. The GPU Parallelization allows Real Time Simulations, as if the CAE System were an XBox or PS4.
But one question didn't get out of my head: what was that Technology used for Domain Discretization - or Meshing - and by the Solver?
After talking with some Engineering colleagues and Professors, I have a guess.
NVIDIA FLEX LIBRARY
If we go back in time and check the Stanford's work of Ian Buck (NVIDIA's Director of GPU Computing) we'll understand that, basically, CUDA is a GPU Platform for C Programming Language. Here are some Videos:
In order to fully unleashes the power of Real Time Parallel Computing, NVIDIA developed lots of GPU Accelerated Libraries - PhysX, Flow and Blast, for example. One of the most interesting Libraries NVIDIA offers is called FleX. For more: FLEX 1.1.0 MANUAL.
As you can read in the Manual, FleX uses Position Based Dynamics (PBD) and Position Based Fluids (PBF). Solids are modeled by Voxels - a very primitive way of Discretization. And Fluids, by particles. When both make contact, the Fluid Particle is treated as a Solid.
PBD, PBF and their Real Time unification were developed by Matthias Müller (NVIDIA's Research Leader):
M. MACKLIN, M. MüLLER; Position Based Fluids; ACM Transactions on Graphics (SIGGRAPH, 2013).
MY GUESS
My guess is that ANSYS is using NVIDIA FleX library to fully profits from the CUDA Platform. Gene Poole (ANSYS's Software Developer) doesn't confirm that, but gives hints in the same direction:
Another possibility is: ANSYS could have developed a completely new Solver, optimized for the CUDA architecture. Serban Georgescu, for example, is a Researcher that investigates the use of NVIDIA and AMD GPUs to accelerate the Finite Element Method.
CONCLUSION
If everything considered above is true, it's great news for the Engineering Community. It means that, soon, the barrier between Gaming Physics Engines and CAE will be removed. In the near future, most likely, we'll have brand new and Real Time Methods - like PBD and PBF - as part of our arsenal.
However, we still have to keep something important in mind: like any Method, PBD and PBF have their weak points. Our Mission, as Engineers and Leaders, is to study and understand well the Technology we're using. For example: Voxels may reduce computational time, but they sacrifice the accuracy of the solution.
Let's wait and see how the concurrence will reacts. For sure, ALTAIR, MSC SOFTWARE, DASSAULT and SIEMENS will show their new Solver Technologies soon.
KEEP AN EYE
Lately, the Engineering Community has noticed a considerable effort in eliminating the task of Meshing. ANSYS did an excellent job adding Real Time to the idea. In terms of Domain Discretization, there are already two interesting Technologies growing: External Finite Element Approximation Method (EFEAM) and Lattice Boltzmann Method (LBM).
EFEAM is less known and, for the moment, SIMSOLID is the only software I've heard of that uses it. SIMSOLID works directly over the CAD Model. It can be used together with ONESHAPE, FUSION or SOLIDWORKS.
An interesting feature of SIMSOLID is its capability of detecting bolts and other joints, from the imported Assembly. Moreover, analyses are super fast - taking only seconds for simple Assemblies. For the moment, the Solver does not support Fluids.
For more: EXTERNAL APPROXIMATIONS BY FINITE ELEMENTS, SIMSOLID TECHNOLOGY OVERVIEW and SIMSOLID VALIDATION MANUAL.
As for LBM, I've seen some works from NASA; and a couple of companies are already selling robust solutions - such as XFLOW, POWERFLOW and ULTRAFLUIDX. Even though there are Researchers trying to create a Code capable of dealing with Solids, at the moment, LBM based Solvers deal only with Fluids. Due to its Lagrangian approach, LBM is particularly powerful for Turbulence problems.
Currently, some Aircraft Manufacturers - such as EMBRAER, AIRBUS and BOEING - are testing both XFLOW and POWERFLOW. Some Results can be checked at AIAA CFD High Lift Prediction Workshops.
Simulation Engineer | Entrepreneur | Ansys Partner
7 年I have been having this conversation with Joel Koster for about 10 years now as we have watched these tools converge in their ability to simulate physics... It's a fun time to be in this business! Thanks for the great investigative work, Leonardo!
Product Manager | Product Owner | Scrum Master
7 年For more about Real Time Simulations: < https://bit.ly/2z0Zj2o >. The Post has a video by Prof. Chinesta, speaking about Model Order Reduction. There are many Methods for doing that. The one he develops is called Proper Generalized Decomposition (PGD). He is using that for Real Time simulations of Surgeries and Manufacturing Processes. One of the examples I liked the most was the Real Time prediction of bending, due to Residual Stresses, during Milling Operations.
Dystrybucja ANSYSa w Polsce od ponad 25 lat
7 年Great article, thank for that! I can only say that after talking to some guys in Ansys I learned that is it lattice Boltzmann method that is being used in the Ansys Discovery. Still, these flow simulations look as if they were run on large viscosities. Yet, the LBM theory has been very well developed recently, hence progress will show up very soon.
Chargé d’affaire Engineering
7 年Thanks for the analysis Leo! I was aware of GPGU capability for simulation for a while. Ansys did a great job by leveraging this technology . The whole PhysX package use true physics approximation for the particles cinematic. But the output are not directly usable. That's why I think Ansys use both FleX(PhysX) library and proprietary code to leverage the gpgpu power. Abaques and Nastran are already capable of running on gpu with very interesting results for lower investment.
Stress & FEA Engineer (Aircraft Primary/Secondary & Gas Turbine Structures: Wing, Fuselage, Engine Hot Section, Nacelle & Actuation Systems; Automotive Durability & Crash FEA)
7 年Really valuable article. Thanks for sharing your comments and all reference videos regarding the transition of simulation today. I have a feeling that we will get rid of finite elements in 5 years or so. It is just a matter of time until a software team figures it out completely. If they can't figure it out until then, we might still have the chance to have more finite element fun in Quantum computers. Both possibilities look exciting though..