Cuckoo Cycle https://github.com/tromp/cuckoo rates as the most memory intensive and slowest PoW ever. With the default graph size of 2^30 (denoted cuckoo30) nodes (roughly a billion), it takes the fastest CPU 2.5 seconds for look for a 42-cycle. For the past few months I’ve been working on porting that code to CUDA. It’s running well enough now to convincingly outperform CPUs. Today I finally had a chance to run it on a 1080ti, where it runs each attempt in a cool 1.3 seconds.
Considering my amateur level GPU coding skills, this bodes well for a future sub-second miner.
The race is on!
Where will we find experts in GPU coding?
They hang out in https://bitcointalk.org/index.php?board=160.0 and https://forum.z.cash/c/mining
People like Claymore, EWBF_, optiminer, dstm, and realbminer. All anonymous, but can be left private messages. There’s also Marc Bevand who wrote a solid OpenCL Equihash miner.
Indeed, today was the first time I was able to light up the yeastmonster building a proper cuckoo-30chain on all 3 cards at once at a blazing 1.7 solutions per second! Even though the race is still on for sub-second, we’re at a point where cuckoo 30 is just about ready for deployment on testnet2. Well done!
Ha! Are you actually paying bounties out in Bitcoin Cash?
I hope to… still waiting for the bounty claims…
Anybody has final numbers for GTX 1070? I guess around 1.2s, but I lost my thumbdrive with Linux configured for CUDA cuckoo compile…
And the race to break the 1-second barrier is over!
Using 128 thread blocks, and about 5.4GB of memory, the solver runs in under 0.98 seconds on an NVIDIA 1080 Ti.
I started playing round to try get the cuckoo miner running on android (simpleminer). Not got any performance stats yet or worked through if this could even be used with grin. but managed to find some solutions.
How did you go about building the grin miner for Android? Have you had time to try it with the latest grin cpu miner? Any problem solving notes you can share about how to build grin for android?