Interesting, I hadn't seen this before. I find it hard to believe, as Argon2 does psuedorandom access over a large array. As soon as this array gets larger than the local GPU cache (much smaller than CPU cache), we should get pretty good protection. What have I missed?

In particular, in terms of ASIC attacks, bcrypt and other non-memory-hard KDFs have extremely efficient implementations. Silicon is cheap, computation is cheap, memory access is extremely expensive- both in terms of time and power usage.