This is not JeanLucPons version of Kangaroo
this have a lot of unnecessary checks that slow down the process.

The plain cuda version of secp256k1 and several optimize changes have this speed.

RTX 3070 with power limit 170w 2.4 GKey/s  GPU Ram Usage 3.5GB
RTX 3090 with power limit 350w 4.8 GKey/s  GPU Ram Usage 8.0GB

And one dedicated server for check
have 32GB RAM ( 8gb+ ram need for bloom filters )
~140Gb of nvme used for kvrocks storage for store 2^31.10 of DP32
or 35Gb if use DP34

ps4all. Possibly I'll find time for compare if you have dp's (or work file fromJeanLucPons Kangaroo ) for still determine 120 and check wild/tame and later
(in compared dp's included buyed from zielar)

But i rarely come here.