Message boards : Puzzles and science : KERAUNOS
Message board moderation
| Author | Message |
|---|---|
|
25 years old Send message Joined: 27 Sep 26 Posts: 349 Credit: 5,209,993 RAC: 137,017 |
β‘ Keraunos -- what it is and how it works β‘ Keraunos (Greek: "thunderbolt" -- sudden, striking, leaps across the sky) is BitBOINC's kangaroo-based discrete-log search π¦. It's the app that actually has a realistic shot at finding a key in practical time -- but it can only be pointed at a puzzle whose public key is already visible on the blockchain. π Why it's so much faster than brute force Brute-forcing a private key in a range of width W takes, on average, W/2 tries. Keraunos instead runs Pollard's kangaroo method (the van Oorschot/Wiener parallel version): expected effort is only about 2*sqrt(2)*sqrt(W) group operations -- the SQUARE ROOT of the range width, not the range width itself. For a 70-bit-wide range that's the difference between "some multiple of the age of the universe" β³ and "genuinely reachable with enough combined compute" πͺ. This efficiency is exactly why it NEEDS the public key exposed: the algorithm's math only works starting from a known point on the curve, not just an address hash. π¦ How the search actually works πΎ Every host releases "kangaroos" (tame and wild herds) that take pseudorandom jumps across the range. π Whenever a kangaroo lands on a "distinguished point" (a point meeting a rarity condition, e.g. specific low bits all zero), that point plus how far it's travelled gets written into a table shared by EVERY host working that puzzle, across every platform. π₯ When a tame kangaroo and a wild kangaroo land on the exact SAME distinguished point, comparing their recorded distances hands you the private key directly. π Because the jumps are randomized, hosts never need to coordinate or split up the range explicitly -- everyone just jumps and reports distinguished points, and more total jumps (more hosts, more hardware) means a collision gets found faster, same idea as more people buying lottery tickets together ποΈ. Right now Keraunos is working puzzle #140. π§ π§ Recent fixes worth knowing about β A real checkpoint bug was fixed: resuming a task used to accidentally reset all progress back to zero because of how the checkpoint was fingerprinted. Checkpoints now resume correctly. β Linux CPU builds are now statically linked, which fixed real tasks failing on volunteer machines with an older glibc than our build toolchain assumed. β Keraunos now gets scheduling priority over Potamos (roughly 3:1) on hosts that could run either -- since it's mathematically much more likely to actually find a key once a puzzle's pubkey is exposed, that's where we'd rather your hardware spend its time first. π₯οΈ Platforms supported CPU: πͺ Windows -- x86, x64, ARM64 π§ Linux -- x86_64, i686, aarch64, armhf, ppc64le, riscv64 π macOS -- Apple Silicon π FreeBSD -- x86_64 π€ Android -- aarch64 GPU-accelerated π (Windows/Linux/macOS): π© NVIDIA -- CUDA, Kepler/Volta-class and Turing-or-newer (separately tuned) π₯ AMD -- OpenCL π¦ Intel -- OpenCL β¬ Apple Silicon -- Metal (cross-checked step-for-step against the CUDA/OpenCL implementations on real hardware before release β ) π οΈ Coming soon: Moore Threads MUSA GPU support -- the port is in progress, not yet built/released. ARM+CUDA (Jetson-class devices) is on the roadmap but not started yet. π Found a bug? Reply below -- especially anything around checkpoint/resume, distinguished-point counts that look wrong, or crashes on a specific GPU vendor. Include your OS/GPU and, if you can, the task's stderr output. π
|
|
Send message Joined: 11 Oct 26 Posts: 3 Credit: 7,015,268 RAC: 132,892 |
Re: NVIDIA -- CUDA, Kepler/Volta-class and Turing-or-newer (separately tuned) If I'm not mistaken, the Pascal architecture is more recent than Kepler, but precedes Volta. Can you confirm if any of the apps support Pascal-class NVIDIA GPUs? Thanks! |
|
25 years old Send message Joined: 27 Sep 26 Posts: 349 Credit: 5,209,993 RAC: 137,017 |
OpenCL works on all of them.
|
|
Send message Joined: 11 Oct 26 Posts: 3 Credit: 7,015,268 RAC: 132,892 |
Thanks for confirming. I haven't had any luck so far getting any opencl_nvidia WUs assigned for my Pascal-class GPU: NVIDIA GeForce GTX 1050 Host: https://bitboinc.athena.org.tr/show_host_detail.php?hostid=623 I do see there are opencl_nvidia apps listed on Windows x64 for Nostos, Phygas, and Planodia Deep. And Nostos is showing unsent tasks on the Server Status page. Perhaps my project or computing preferences are preventing download, as this host already has one CPU and two intel GPU tasks in progress. I will try adjusting my preferences for this host. Thanks! |
Tamagoch47 years old Send message Joined: 29 Sep 26 Posts: 68 Credit: 14,373,389 RAC: 423,227 |
In reply to hugo75's message of 12 Oct 2026: I haven't had any luck so far getting any opencl_nvidia WUs assigned for my Pascal-class GPU: NVIDIA GeForce GTX 1050 I am running Nostos on 1070 GPU. I think 100% CPU on that host could do better than 1 CPU + 1 GPU in this configuration. For your particular host it also may be better to just utilize 100% CPU. And it surely can run faster than Ryzen if you switch to Linux ) I will publish more performance results later. |