What is offloaded to the CPU specifically? Afaik GPUs are great at parallel processes and most linear algebra can be done in such a way. So why not have everything on the GPU?
I'm not sure which part of the post you're referring to, so I might be answering the wrong part.
If it's about rendering, even nowdays, with GPU particles, tessellation, etc, the process of setting up for executing the render is still CPU. Less than previously (5-10 years ago) but it's still there.
Think of it like the CPU being the secretary and the GPU being the CEO/Associate/Partner/whatever.
Agreed, but I’m still convince proper physics will help to make AI more realistic.
No. There's still so much room left to fine-tune the AI as it exists, without 'creating' any new information. Think of it like this: the AI will hit the curb at the apex of T2 at Red Bull Ring, causing it to hope on the left 2 wheels a bit, then oversteer on power-down. It'll react in (say) 1/50th of a second to countersteer; a micro-correction.
What do humans do? We predict that oversteer is a likely outcome for such an action (curb-hitting), and will...queue up (???) some counter-steer in readiness of that.
I said 'create' like that cause it could be able to have that information from the previous lap. Whether it's learning things like that, I doubt. But the point is, it has more than enough information and variables to drastically improve upon what it is now. And not just improve, but start to show behaviours like that, which we associate with human driving and driver development.
I'm not saying our B-Spec Bob has a brain that big, but it would be very interesting to see Big Think Bob, who was trained on SS replays.
I've been hoping that some indie dev would try this approach, to mimic mid-90's in-car footage - the level of 'realistic' effects you could get at those low resolutions would be incredible.
Sort of like a Cuphead approach, maybe?
(also cheers for the expert contribution)
It's unlikely that an indie dev would try, simply because it's not just a problem of running the application, but building it. Paying the rendering performance cost is one thing, the R&D cost for integrating the latest and greatest teach into a realtime engine, then creating content for it, would still be the same as the R&D + content creation cost a large VFX studio would incur.
Just wanted to thanks you for the post. Imo just "like" the post wasn't enough. Everyone should read it
No problem.