Massively-parallel computing on the the desktop
In the minds of engineers, simulation is an afterthought. We concern ourselves primarily with solving designs geometrically, and many of our designs don't need to be tested against failure, anyhow. For those that do, we let someone else worry about it after it fails.
And so companies that sell simulation software (and those that want to expand into the realm) are frustrated by our insufficient use of their software. The idea expressed by companies like ANSYS (which sells simulation primarily) and Autodesk (which sells it on the side) is this: if we can make simulation easier to use, then more people will use it (and buy more). This leads to phrases like "anyone can use it" and "software democracy!"
There is, however, the counter-argument from experienced engineers: simulation is sufficiently critical that if designers don't understand what the are doing, then the designs can screw up badly -- like a spreadsheet with a bad formula. The most critical aspect is load assumptions. Get them wrong, and the structure fails, no matter what other loads it resisted successfully. And so we engineers prefer that simulation specialists do the work.
ANSYS is nevertheless hopeful that one day the number of designers using simulation will eventually reach 1 in 1. They quote the following trend line from one of their customers:
- 2005: 1 in 22
- 2015: 1 in 6
ANSYS Discovery Live
To reach the 1-in-1 goal, ANSYS created software that doesn't need geometry that's been "fixed" for simulation, does not need users to apply meshing, and offers users no waiting for solving. Discovery Live performs simulation tasks in real-time, reducing the time hours to seconds. See figure 1.
Figure 1: ANSYS Discovery Live doing real time fluid flow analysis
Discovery Live integrates its real-time simulation with a history-free modeler, SpaceClaim, which ANSYS owns. It operates on the desktop, not in the cloud; it gets its speed by running on a single CUDA-compatible GPU from nVidia.
To use Discovery Live, the designer needs to specify only the inlets and outlets (for fluid flows) or forces and attachment points (for stresses):
- The "Discovery" part of the name means that its extreme speed lets designers modify the design and discover in real-time how design changes affect the simulation.
- The "Live" part of the name means the software is interactive and that the analysis is performed live, letting designers adjust pressures and stresses, and watching the system change. (The Live name has nothing to do with Microsoft's line of Live software.)
The software makes assumptions for nearly everything, like velocities and temperatures, and then shows the results in a cross-section view, automatically so that designers don't need to take the time to set them up. You can adjust the assumed values, of course.
How It Works, and Where It Doesn't
ANSYS wrote Discovery Live to run on GPUs, a process that took several years. GPU is short for graphics processing unit, the chip that powers graphics boards, such as from AMD and NVIDIA. The fascinating part is that GPUs contain an excess capacity of computing power, largely untapped in the CAD world.
It is untapped, because GPUs process data differently from the CPU that runs CAD, Windows, and MacOS on our computers: GPUs process data in parallel, running as many as 3,500 operations at the same time -- a.k.a. massively parallel. (CPUs do between only one and about a dozen operations at a time.)
This makes GPUs suitable for processing graphics used for real time 3D rotations and renderings, but is unsuitable for most other software. This is because most software can only do one thing at a time. In order to do 3,500 things at once, the program needs to split the task into 3,500 threads, run them in parallel, and then put together the result afterwards.
The catch is that programs need to set up this task-splitting ahead of time, which means the program needs to know ahead of time what the user wants to do. In CAD, only a very few operations lend themselves to task-splitting (a.k.a. multiple threads), such as loading a drawing file from the hard drive or generating a photo-realistic rendering.
Another task in CAD that benefits greatly from multi-threaded operations is FEA, finite element analysis, the technology behind simulations. FEA breaks up models into tiny chunks, and then operates on each chunk in each thread. This make FEA an ideal application for GPUs.
AMD and NVIDIA have long dreamed to mainstream software running on their powerful GPUs, because that would boost sales of their boards; most users have no need for them, because the graphics that Intel throws in for free is good enough for most users.
To use Discovery Live, your computer needs a graphics board that runs the CUDA [Compute Unified Device Architecture] engine -- nVidia only, not AMD. For example, a GeForce GTX 1080 Ti graphics board (see figure 2) has 3,584 cores, meaning it can run 3,584 operations at the same time, each at a speed of 1.5GHz. This is like running one operation at 5.4THz (5,400GHz) -- that that, Intel!
The graphics board carries 11GB of its own RAM running at 11gigabits per second, moving 352 bits simultaneously -- 5.5x the 64 bits in Intel CPUs. The price is a reasonable $699, but it needs a desktop computer with a 600W power supply. https://www.nvidia.com/en-us/geforce/products/10series/geforce-gtx-1080-ti/
Figure 2: GeForce GTX 1080 Ti graphics board from nVidia runs certain software about 1,500x faster than an Intel CPU
And so Discovery Live gets its extreme speed by running in parallel on a GPU:
- The more on-board RAM, the larger a model that Discovery Live can handle
- The more cores, the faster Discovery Live runs simulations
There is a cloud option for those who do not have the right graphics board, "but the best experience is running it on the desktop," says ANSYS senior product marketing manager John Graham.
The other part of Discovery Live is SpaceClaim, the direct modeler that shook up the MCAD industry when it was introduced in 2005. (I say shook up, because in the aftermath PTC released Creo, Autodesk released Fusion, and Siemens released Synchronous Technology.) The SpaceClaim part lets you edit models interactively during the simulation, as well as build new models from scratch. You can watch the simulation interact as you add and subtract elements.
The software assumes a default material, but you can assign a specific one. Meshing is done automatically through a proprietary process developed by ANSYS.
A simple line graph cleverly tracks the change in efficiency as you make changes. Another part of Discovery Live is, however, manual: you have to repeatedly make changes in order to approach some optimum.
"Simulation is not cheap, but in Discovery Live model-complexity is free," says Mr Graham. Fillets, sharp edges, and other details are a problem for meshing, but not for Discovery Live. "Start learning physics instead of simulation, " he adds.
What Ralph Grabowski Thinks
My first thought following the demo went into the future: "This does away with the need for cloud computing." And so I wrote an editorial on the implication of using a GPU for edge computing, and how the cloud suffers from a slow data bus speed. See http://www.worldcadaccess.com on WorldCAD Access
Another editor put it this way: "This is the most impressive software demo I've seen in two decades."
Now, this software isn't for everyone, and even ANSYS says so. It is meant primarily for use at the start of the design process, and somewhat in the middle. Discovery Live is accurate in a certain direction, but should not be used for final validation. It is fine to ensure a product doesn't break, but is not suited for validation where human lives are involved. For studies that require deeper physics, ANSYS will sell you AIM. (See http://www.ansys.com/products/3d-design/ansys-aim .)
The hidden advantage of running software on GPUs is that as more cores are added to graphics boards in the future, the software just gets faster. It scales upwards automatically.
Discovery Live will be free during the beta phase, and then is due to be released and sold in the first quarter of 2018. ANSYS plans to add more simulation analyses to the software in the future.
Good news indeed about ANSYS & Discovery Live - it's about time! GPU computing has been a hot topic in certain circles for well over 5 years: http://fastra.ua.ac.be/en/index.html
Two points to add:
1. A decent motherboard can host multiple video cards, which can be linked together to double the power and speed of GPU computation. AMD calls this linking "Crossfire", while nVidia calls it "SLI", and it works very well.
2. AMD used to have their own proprietary code for GPU computation, but switched over to an open source package called OpenCL : https://www.khronos.org/opencl/
ANSYS may want to port to that software as well.
Thanks as always fro your excellent reporting.
Posted by: Peter in Maryland | Sep 11, 2017 at 04:22 PM