Quote from: bigvito19 on Today at 07:20:30 PM

Quote from: NotATether on March 14, 2021, 04:32:48 AM

Quote from: WanderingPhilospher on March 13, 2021, 03:48:54 AM

This is just not true. The RAM consumed/used is not dependent on the bits being search...it comes down to your DP setting and how often you save the work file. I don't see your settings/flags when starting the program but you are using a very low -d setting or that plus using a long save to file time.

I can see now that searching a 256-bit interval will become feasible if we use 128+ bits of DP mask

Quote from: NotATether on March 23, 2021, 06:25:10 PM

Quote from: bigvito19 on March 15, 2021, 07:16:11 PM

How feasible are you talking? So you are working on searching 256-bit interval using 128 + bits of DP mask too?

I forgot about this thread - sorry about that.

Yeah I am, so far I have the actual program running on GPU, it's running at ~260MKeys/s with the expanded dpmask on my T4 though but it's making quite a large number of same herd collisions and dead kangaroos. I had actually expected the speed to be much faster, like around ~1500MKeys/s given that I saw someone's V100 do ~1100MKeys/s.

I doubt checking three more uint64's for equality within the main loop is what's causing this speed drop but it's a good opportunity to peek into the CUDA accelerated Int class and see what else can be sped up.

NotATether, Did you ever work on the 128 + bits DP mask?

That was a joke...it would take you 2^128 jumps to find a 128 DP mask. So let's say your CPU (1 thread) can do 4,000 jumps per second. Now divide 2^128 / 4000 jumps per second = 85,070,591,730,234,615,865,843,651,857,942,052,864 seconds to find one 128 DP.