Memory-hard function
A key derivation function that needs large amounts of RAM, so specialised hardware cannot guess passwords much more cheaply than a normal computer.
A memory-hard function keeps a large block of data in memory throughout the computation and reads it in an unpredictable pattern. That makes GPU and ASIC guessing rigs expensive, since each parallel guess needs as much memory as a general-purpose processor would use.
Colin Percival presented scrypt in 2009, one of the first functions of this kind. Argon2 won the Password Hashing Competition in 2015. It is specified in RFC 9106, while scrypt is in RFC 7914. Both let you choose memory size and time cost, so you can decide how heavy the function should be.
It is not a magic fix. The memory limit is a trade-off against how much RAM a legitimate server can spare, and choosing too little memory leaves roughly the same attack surface as a plain hash. For passwords, where every guess matters, it is still a sensible way to raise the price for whoever is doing the guessing.