A theoretically perfect TPM is way better. The strength of your KDF is limited by the user experience. If you want to limit an attacker to trying one password per minute, then you have to make the user wait one minute to log in, if you assume the attacker only has the same amount of computing power you do. A TPM can just say “you have ten tries and that’s it.”
Of course, a real TPM won’t be perfect. But there’s no reason you can’t do both. Have the TPM use a good KDF and then an attacker has to break both.