...
you will never get 250 MK/s on any GPU, by computing H160 out of some list of private keys.
...
You just don't have enough experience in GPU coding or/and don't know about some efficient methods for EC scalar multiplication.
4090 can do >0.5GK/s at rndprivkey-to-h160, I can publish the code, just need to find some time to make my sources more readable to make them public.