Pretty much all the standard performance advice from other languages like C/C++/Rust applies. The concepts driving "data oriented design" will get you a long way. The biggest difference is learning how to play nice with the GC- avoiding allocations where necessary, keeping the heap simple so that collections are cheap, or just setting things up so the GC never has to run.

I wrote a bit of stuff a few years ago: https://www.bepuentertainment.com/blog/2018/8/30/modern-spee...

Some of the implementation details (like the operator codegen) are surprisingly outdated now given the speed at which the runtime has moved (and the library is now ~4x faster or something silly like that), but the fundamentals are still there.

Since I wrote that, codegen has improved a lot, the hardware intrinsic vector APIs have offered a way to optimize critical codepaths further, additional cross platform vectorization helpers have made it even easier, NativeAOT is getting close, and all sorts of other stuff. The C# discord #allow-unsafe-blocks channel contains a variety of denizens who might be able to offer more resources.