how do i make a computer powerful enough to simulate anything i want in realtime? sound modelling of strings, percussion membrames, flesh being hit, glass shattering, the stress moduli of rubber, concrete, metal, plastics, and flesh being ripped, punctured, sliced, and twisted, the lighting conditions and various optical effects of the human eye or camera capturing it, volumetric particle interactions with clouds, fog, fumes, and fluid dynamics for water, sludge, mists, and honey.
There's also stuff like leather, cotton and polyester for clothing. i think most of this is possible in cgi, but im looking for a realtime approach possibly with unreal, but with higher fidelity for realism that i could get in something like maya or houdini. what do?
The one you have with you. Do we even have software for this?
Nolan Johnson
unreal is a game engine so its tailored for realtime performance but the physics simulations and audio synthesis are a no go. i have a program called modalys which runs under Max that lets me model strings, membranes, but not flesh, and i have another program for vocal tract simulation to make voices. nothing for more organic materials though. i need to learn more about simulation programming.
Kevin Perez
Based pervert
Daniel Barnes
uh no it's for video game realism
Nathan Cook
how can you make a computer that can compute the universe?
Austin Scott
i figured a crapload of titan Vs would be the first step, then i'd figure out how to utilize only half precision computing to do it since it has 100TFLOPs each.
Aaron Wood
>simulate That word is not as clearly defined as you might think. All simulations use some sort of simplified model at some point. Sometimes due to constraints in the current understanding of physics, sometimes due to mathematical/numerical intractability, sometimes due to memory/CPU limitations. In any case, there is no "all the way down". You have to draw a line somewhere and fake it.
After all, the use of sampled sounds is a simulation of sorts. So is the use of all the simplifications you see in games nowadays. You need a better definition of your goals first, I think.
Owen Rogers
but the machine is in the universe so you'll have to simulate the entire machine you are using to simulate it
Gavin Green
Is it possible to do that book where someone had a quantum computer and used it to simulate what would happen in the past and found the secret of the universe?
Nathaniel Stewart
basically i want to make realistic but interactive game environments that take advantage of photorealism via not only raytracing but also with advanced shaders, vellum constraint soft body simulation, flip fluids, and volumetric fumes.
basically, i want the parity available now to viewport previewing of scenes in houdini, maya, and 3ds max for photorealistic production scenes to be what the engine displays at an interactive rate. it's called shader parity, and this is side by side with a full production render. so basically i want that realtime viewport to be what the core engine uses.
For real time stuff (say one frame latency at 60fps, i.e. 16ms), you cannot make computers too big. Synchronization overhead and memory bandwidth might end up being the biggest problems.
Austin Perry
but isn't HBM supposed to mitigate that?
Owen Rodriguez
synchronization loss is exponential so if you have 2000 cores it might only be 1% faster than 1000 cores depending on the sharing, this is even if your memory had zero latency
Easton Clark
so it's not the memory bus but the actual proximity between cores causing a delay and in turn limitation on concurrency? how do you deal with synchronization if you need a larger amount of operational performance?
Ethan Peterson
For realtime rendering you're probably stuck with one node. Get a dual socket board and a couple of either epyc rome or the double die cascade lake monsters. Might be best to pick whatever has the most pcie lanes. Cram as many quadro rtx 8000 cards in as you can fit because that's the bulk of your real compute power. No matter what it's going to be a huge challenge programming efficiently for that sort of beast. I'm sure there are some fun possibilities though.
Juan Martinez
Based manchild
Sebastian Campbell
that's what i figured, but i wanna try and get titan Vs instead of rtx because they're only about $2000.00USD versus the 5000+ quadro rtx cards ask for, but they have the volta architecture on them which allows for 100 teraflops of half precision performance apparently. so right now im just trying to figure out what i can viable do with half precision and what can't be done with it, and if i can make an 8 gpu build using two PCs, each with a threadripper and its 64 pci lanes and in turn 4 gpus, then i could beowulf link them and get some decent power for not so much cash, like 900+ teraflops. i should be able to run realtime but realistic raytraced lighting, physics for fluids/soft body, and advanced shader models with that right?
Asher Morris
The 100 teraflops refers to tensor performance not general half precision performance. Turing and Volta both have double rate half precision. Turing has somewhat fewer tensor cores than Volta but that's only useful if your primary performance consideration is neural networks. Turing has better performance on every other workload plus you get ray tracing acceleration. If you care at all about bang for your buck then a single socket Epyc or Threadripper system filled with 2080tis is definitely the way to go. You can do the price/perf math on Epyc vs Threadripper yourself. Epyc has twice the PCI-E lanes so you can get twice the GPUs in one box.
Jeremiah Allen
no it's memory sharing and virtual sharing
William Ramirez
yeah but what motherboard will allow me to connect 8 GPUs to an epyc chip without using extension modules?
Jackson Watson
Yeah I'm not sure what the best config is in precise detail. The strongest advice I have to give is that you should really go with RTX cards. 1x 2080ti = 4x Titan V in ray tracing and they are about the same in general purpose compute. BTW the reason I suggested the high end Quadro cards is they come with up to 48GB of memory which could easily be relevant in an extreme workload. It's a huge premium over the Geforce cards for no gain in raw compute though.
Matthew Turner
yeah that is true. i might try to buy a lot of them via alibaba for bulk discounting, but i think that the 2080ti should suffice. i might also just wait for lavina from chaos group simply because it might actually allow for advanced materials, but i don't know what to do for realtime advanced physics simulations.
If your intent is to go beyond what exists now you need to get ready to program it yourself. Try to find a good extensible platform because none of them will do everything out of the box.
Benjamin Gonzalez
do you have any reccomendations by any chance? i might be willing to work on my own stuff via C++ since there seems to be a crapload of APIs related to this, but doing it from scratch would be painful since i'd wanna make a UI for all this as well.
Kevin Brown
It's a tough question. Unreal has a good feature set for graphics and anything interactive but it has nothing like any of the high end physics you're interested in. Either you have to make a realtime renderer interoperate with your physics software or muscle an offline renderer into running in realtime.
Jace Green
Implying you even have the computational mathematics and physics knowledge to implement a paper from two decades ago.
John Sanders
Kinda pointless negativity isn't it? If he's fantasizing it's harmless fun and if he's serious it might turn into something interesting. We're all just browsing an internet forum anyway.
Jackson Russell
try and stop me.
Sebastian Cox
It's a bit harsh, but it should diagnose what he doesn't know. If he is really passionate about this, he should know numerical analysis, methods, and solutions to differential equations. Then he has to be able to code this in C++ with proper data structures and vectorizing algorithms whenever possible.
Josiah Hill
And numerical linear algebra top.
Levi Kelly
i'm not that knowledgable about numerical analysis, although i have a few pdfs on the matter. i just wasn't aware of how much of the computational mathematics i needed to pull out to retrofit existing libraries into one system, in comparison to making it from scratch. i mostly want fast efficient sims with high framerates. would nvidia's physics libraries help with that or is their accuracy too low for realism?
Luke Cox
YOLO
Jaxson Rodriguez
anyway OP post a thread on /3/ if you make something cool and don't forget that someone out here in the icy universe believes in you
Benjamin Martinez
will do. thanks. on a side note before i leave, i also had a plan for something else that isn't related. it's a program to control a robotic arm with a person strapped to it so that they can perform stunts without a supervisor. basically their stras have haptic feedback and a camera looks at them to see how they're moving to ensure they won't collide with anything, while letting their natural movements be enhances or lets them run on walls or jump higher than normal, you know extravagant matrix-esque stunts, but without a whole wire team. just a robotic arm attached to a vehicle that follows them.
Blake Stewart
Look at the book called Numerical Recipes. It will give you the high performance algorithms. I'm unsure if those libraries are the state of the art oh high performance algorithms. That's where knowing numerical analysis and numerical linear algebra comes in so you can come to your own conclusions.
this one is a collection of tools rather, for modelling the vocal tract. you could combine it with modalys which lets you create exciters and make a realistic vocal chord.
