The desktop RTX 3070 has 46 SMs, which are the most comparable thing to Apple's cores.

NVIDIA defines any SIMD lane to be a core. They recently have gotten more creative with definition, they were able to double FP32 executions per unit (versus previous gen) and hence in marketing materials, doubled the number of "CUDA cores".

Funny, all these years I've been wondering how they possibly packed so many "cores" into those things.

The apples to apples comparison would be CUDA cores to execution units. Basically how many units which can perform a math operation. Apple's architecture has 128 units per core, so a 32 core M1 Max has the same theoretical compute power as 4096 CUDA cores. This of course doesn't take into consideration clock speed or architectural differences.

CPU core == Cuda SM