Dr. Zsolnai shows a new particle data structure that speeds up GPU simulations across multiple graphics cards. A 100-million-particle concrete crush runs at 4 minutes per frame; a 134-million-particle bomb ensemble runs under 1 minute per frame. Smaller time steps multiply the compute cost dramatically.
Questions this video answers
Nearly 23 million particles; about 4 seconds per frame on a system with four graphics cards.
48 million particles; about 15 seconds per frame.
55 million particles; about 30 seconds per frame.
Nearly 100 million particles; nearly four minutes (≈4 minutes) per frame.
The scene uses around 6,000 bombs which together total 134 million particles; less than one minute per frame.
Chapters
Dear fellow scholars, this is two minute papers with Dr. Zsolnai. If we study the laws of fluid motion and implement them in a computer program, we can create and enjoy these beautiful fluid simulations. And not only that, but today, with the amazing progress in computer graphics research, we can even enrich our physics simulations with anisotropic damage and elasticity. So what does that mean exactly? This means that we can simulate more extreme topological changes in these virtual objects. This leads to better material separation when the damage happens. So it appears that today we can do a great deal, but these techniques are pretty complex and take quite a while to compute, and they typically run on your processor. That's a pity, because many powerful consumer computers also have a graphics card. And if we could restate some of these algorithms to be able to run them on those, they would run significantly faster. So do we have any hope for that? Well, today's paper promises a new particle data structure that is better suited for number crunching on the graphics card and is hence much faster than its predecessors. As a result. This runs the material point method, the algorithm that is capable of simulating these wondrous things you are seeing here not only on your graphics card, but the work on one problem can also be distributed between many, many graphics cards. This means that we get crushing concrete, falling soil, candy balls, sand armadillos, oh my, you name it, and all this much faster than before. Now, since these are some devilishly detailed simulations, please do not expect several frames per second kind of performance. We are still in the seconds per frame region, but we are not that far away. For instance, hold on to your papers, because this candy ball example contains nearly 23 million particles. And despite that, it runs in about four seconds per frame on a system equipped with four graphics cards. Four seconds per frame. My goodness. If somebody told this to me today without showing me this paper, I would have not believed a word of it. But there is more. You know what? Let's double the number of particles and pull up this dambreak scene. What you see here is 48 million particles. That runs in 15 seconds per frame. Let's do even more. These sand armadillos contain a total of 55 million particles and take about 30 seconds per frame. And in return, look at that beautiful mixture of the two sand materials. And with half a minute per frame, that's a great deal. I'll take this any day of the week. And if we wish to simulate crushing this piece of concrete with a hydraulic press, that will take nearly a hundred million particles. Just look at that footage this is an obscene amount of detail. And the price to be paid for this is nearly four minutes of simulation time for every frame that you see on the screen here. Four minutes, you say? Hmm, that's a little more than expected. Why is that? We had several seconds per frame for the others, not minutes per frame. Well, it is because the particle count matters a great deal. But that's not the only consideration for such a simulation. For instance, here you see with Delta T, something that we call time step size. The smaller this number is, the, the tinier the time steps with which we can advance the simulation when computing every interaction, and hence the more steps there are to compute. In simpler words, generally, time step size is also an important factor in the computation time. And the smaller this is, the slower the simulation will be. As you see, we have to simulate five times more steps to make sure that we don't miss any particle interactions, and hence this takes much longer. Now this one appears to be perhaps the simplest simulation of the bunch, isn't it? No, no, no, quite the opposite if you have been holding onto your paper so far. Now squeeze that paper and watch carefully. They wiggle. They wiggle. Why do they wiggle? These are just around 6,000 bombs, which is not a lot. However, wait a minute. Each bomb is a collection of particles, giving us a total of not a measly 6,000, but a whopping 134 million particle simulation. And hence we may think that it's nearly impossible to perform in a reasonable amount of time. The time steps are not that far apart for this one, so we can do it in less than one minute per frame. This was nearly impossible when I started my PhD and today, less than a minute for one frame. It truly feels like we are living in a science fiction world. What a time to be alive. I also couldn't resist creating a slow motion version of some of these videos. So if this is something that you wish to see, make sure to visit our Instagram page in the video description for more.
What you see here is an instrumentation.
For a previous paper that we covered.
In this series, which was made by weights and biases. I think organizing these experiments really showcases the usability of their system. Weights and Biases provides tools to track your experiments in your deep learning projects. Their system is designed to save you a ton of time and money and it is actively used in projects at prestigious labs such as OpenAI, Toyota Research, GitHub and more. And the best part is that if you have an open source, academic or personal project, you can use their tools for free. It really is as good as it gets. Make sure to Visit them through wnb.com papers or click the link in the video description to start tracking your experiments in five minutes. Our thanks to WeightsAndBiases for their long standing support and for helping us make.
