xref: our C++ counterpart (cudax) NVIDIA/cccl#2084

Piotr suggested we base the exposure on cudaDeviceAttribtues rather than cudaDeviceProperties to reduce the minor version compatibility considerations. WDYT @leofang

Inside you DeviceProperties class we could consider using that. But I don't know how does this help with

to reduce the minor version compatibility considerations

I guess if cuda 13.x.y came out and .y had some great new property we wanted to expose, we would have to wait for 13.z if we used the property struct in the backend, whereas we could expose it right away if we used the attributes. I will look into it and wee what the trade-offs are

Yeah followed-up internally, the device prop struct is frozen by major.0, so any new attributes added in major.minor will not be exposed to the prop struct until the next major release, whereas the attribute query API can immediately gain support by the minor release that adds the new attribute. Let's move away from the device prop struct, and query & cache the attributes lazily (on the as-needed basis).