In cpython:

DUP_TOP:

            PyObject *top = TOP();
            Py_INCREF(top);
            PUSH(top);
            FAST_DISPATCH();

TOP:

            (stack_pointer[-1])

LOAD_FAST:

            PyObject *value = GETLOCAL(oparg);
            if (value == NULL) { /* throw */ }
            Py_INCREF(value);
            PUSH(value);
            FAST_DISPATCH();

GETLOCAL:

            (fastlocals[i])

So looks like either way, you have an array access of something definitely in cache, a Py_INCREF, a PUSH, and a FAST_DISPATCH. The walrus operator saves you a null-check, but that check is probably skipped right over by the branch predictor, as it always throws. I'd bet the performance is indistinguishable, but I'd be interested to see for real.