The tensor core accelerates mostly matrix operations and is the big block you can see has 4 per SM. Cuda core refers to the thread per SM, which you can see as

The tensor core accelerates mostly matrix operations and is the big block you can see has 4 per SM. Cuda core refers to the thread per SM, which you can see as FP32 or INT32 units, so there are (32*4) per SM on that diagram. Like you said, tensor core is similar to a special purpose ALU and is at a lower level of abstraction than something with an instruction pointer.