With this level of hardware power, if you had the knowledge to calculate the locations of the Prefixes.
1- Would you prefer the entire range or RANDOM scanning?
2- Would you chase the 1 in 1024 chance, even though you know the locations of close to 1024 with average probability?
I would be happy if you give a short answer. 1? 2?
If you choose 1, I can say a lot of words for you.
If you choose 2, you are thinking right. I wish you success in your studies.
It depends on what I'm looking for, assuming your strategy works.
If I'm looking for an exact total match, I'd scan the full range, first to last key. No exceptions. It's simply the fastest, most efficient way to do it. It's also the most parallel-friendly way to do it at massive (and I mean 68-bit massive) scale, since there is no need for complex logic, like stopping 65536 threads in the middle of their work just because some prefix matches somewhere, and then setting things up again, recomputing ranges, etc. Maybe you do that manually, which is a time waste, which can be used to simply scan more stuff, IDK.
---
If I'm after a nice prefix but I don't care about a perfect match, just find it really fast - I'd go for 2. But in all honesty I would not even need to go for 2 - because a VanitySearch would run many times faster, for this purpose. For example, an RTX 4090 that searches for some target prefix, hard-coded, and using endomorphism, can do 9.6 billion hashes per second, and find a key having the desired prefix.
And here is why I wouldn't use 2 to find a complete match (I said this already on some of your first posts): even if you take that "1 in 1024" route, there is a chance you don't find what you wanted, but something else. But the overhead of repeating the process to exhaustion (starting over with the second "1 in 1024 chances", then the third ... then the millionth, etc), overall, results in a longer process than simply scanning the range start to finish. And the probability is identical in both cases, because you can't cheat your way out of searching linearly for something that is random. You may disagree here - it's OK.
Meanwhile, Bram already started hunting down 68 for several days now, since the crowd funding was completed, so I guess you should hurry up.