Solving ECDLP with Kangaroos: Part 1 + 2 + RCKangaroo

[Ykra]

I updated the first part:
SOTA v2: This method has same complexity but uses only two groups of kangaroos.
So we can get same K=1.15 with only two herds: one wild and one tame.

Interesting, tested this method in my GPU solution and can confirm it gets the same K as my 3 group implementation based on your original SOTA.
Thanks for the continued contribution.

I've been (for fun) trying to push my 5090's GKeys/s, currently sat at a peak record (for me) of 11.6 GKeys/s.
I'm aware you've focused mostly on 4090's but have you tried to push a 5090 yourself and if so did you achieve a lot higher? (understand it's less bang for buck so was maybe classed as a waste of your time ^^)

[bebejj]

What are the reported numbers for a single GPU? I'd like to know the estimated maximum performance per GPU and see if there are any bottlenecks in the upgrades I'm making, since the highest I've been able to count is a single RTX 5080 GPU at 6000 MK/s.

[RetiredCoder]

Interesting, tested this method in my GPU solution and can confirm it gets the same K as my 3 group implementation based on your original SOTA.
Thanks for the continued contribution.
I've been (for fun) trying to push my 5090's GKeys/s, currently sat at a peak record (for me) of 11.6 GKeys/s.
I'm aware you've focused mostly on 4090's but have you tried to push a 5090 yourself and if so did you achieve a lot higher? (understand it's less bang for buck so was maybe classed as a waste of your time ^^)

Currently I have code that has about 18G/s on 5090 for SOTA. Sources are difficult for understanding a bit, also there are some additional requirements to build them, but I will prepare everything for publishing as open source with explanations because I want to make my ideas public.

[mjojo]

Interesting, tested this method in my GPU solution and can confirm it gets the same K as my 3 group implementation based on your original SOTA.
Thanks for the continued contribution.
I've been (for fun) trying to push my 5090's GKeys/s, currently sat at a peak record (for me) of 11.6 GKeys/s.
I'm aware you've focused mostly on 4090's but have you tried to push a 5090 yourself and if so did you achieve a lot higher? (understand it's less bang for buck so was maybe classed as a waste of your time ^^)

Currently I have code that has about 18G/s on 5090 for SOTA. Sources are difficult for understanding a bit, also there are some additional requirements to build them, but I will prepare everything for publishing as open source with explanations because I want to make my ideas public.

I feel this is a sign if P135 bits already solved by Mr. RC.. Congrats Sir

[Cricktor]

I feel this is a sign if P135 bits already solved by Mr. RC.. Congrats Sir

Please use your brain.

If RetiredCoder had already solved puzzle 135, why would he wait to claim the coins? There's no need to use hidden mempool service like slipstream.mara.com in that case, because the public key is of course already known. Any solver of puzzle 135 could broadcast the withdrawal transaction publicly. It would be stupid to wait to withdraw the coins. And I don't locate RetiredCoder in the stupid camp.

For puzzles like #71 and moderately up, where the public keys are unknown, a solver should definitely not withdraw by public transaction, only via above mentioned Slipstream service and non-public mempool!


Even with impressive performance on a 5090 as RetiredCoder claims for his code, you need a heck of a lot cards, time and energy and thus a hefty investment to solve #135.


Post edit:

..., but I will prepare everything for publishing as open source with explanations because I want to make my ideas public.

Totally forgot to comment: looking forward and really appreciate you're willing to share your tricks and approaches for better performance!

[Ykra]

Currently I have code that has about 18G/s on 5090 for SOTA. Sources are difficult for understanding a bit, also there are some additional requirements to build them, but I will prepare everything for publishing as open source with explanations because I want to make my ideas public.

Very impressive, I was quite happy with my baby steps hitting 11.6 G/s (while keeping SOTA) haha.

Really curious to see where the largest gains stem from and particularly interested by "also there are some additional requirements to build them", this sparks some ideas that I will toy around with.

[Belkacar]

I updated the first part:
SOTA v2: This method has same complexity but uses only two groups of kangaroos.
So we can get same K=1.15 with only two herds: one wild and one tame.

Can the current publicly available version of the program find 135 puzzles? Or are there limitations?

[Cricktor]

...

Why don't you verify it yourself? You would blindly believe any answer that RetiredCoder gives you if he answers your question at all?

I would verify any tool that I use for a key recovery because I want to be sure my toolchain actually works.

Create your own sample #135 puzzle private key being masked appropriately to fall in that range, derive the public key and test the tool with a small enough start and end range so that you'll get a result in minutes or hours of computation.

If it finds the sample private key you already know, you verified it's capable to find a solution in the key space range of the real puzzle #135. (OK, there could still be weird bugs preventing it to find a solution in larger slices of the key range.)

But that's what I would do first with any solving tool.


You are aware that a single GPU or even only a few handful of them won't cut the cake for real puzzle #135?

[puzzlesolver878]

hi,

i want to know how i can save work, as i see if machine restarts, all work is lost and need to start by 0.

i also noticed that if i use thame fileles, required time is 3x less, but it requires host to be gigh ram.

so if i build a 1TB ram machine, it will help me solving the puzzle 135?

thanks

[kTimesG]

so if i build a 1TB ram machine, it will help me solving the puzzle 135?

Maybe. You definitely need to store DPs in RAM, it helps a lot. Just make sure the machine and the app never restarts or crashes for any reason, for the next ~5000 years or so, give it or take a few millennia.

[bebejj]

hi,

i want to know how i can save work, as i see if machine restarts, all work is lost and need to start by 0.

i also noticed that if i use thame fileles, required time is 3x less, but it requires host to be gigh ram.

so if i build a 1TB ram machine, it will help me solving the puzzle 135?

thanks

1TB of RAM won't help you at all, don't waste your money on it. You could build a system with up to 8TB, but it would still be useless. The only way to improve is to speed up the processing, but even that's very slow for Puzzle 135 unless you have a 2000 Mbps GPU. Currently, it's not viable with all known programs unless you have a 2000 or 3000 Mbps GPU.

[puzzlesolver878]

if program cant be restarted in the whole execution as all progress is lost, how did retired coder solved puzzle in 1 month?

all GPU programs loves crashing once a day. its mandatory to perform checkpoints at least once an hour of work

i am currently solving puzzle 71 saving work each 20 minutes as miners crash every time

thanks

[ficklebellpepper]

...

You are aware that a single GPU or even only a few handful of them won't cut the cake for real puzzle #135?

I think about that a lot, but part of my brain goes to the scene from dumb and dumber - "So you're saying there's a chance" and I wonder while I am heating my room when it's cold out that my 4090 might just find that magical cross over and get a hit. 

Trying some novel stuff out to see what happens, and seeing what kind of clever bits everyone else is thinking about.  Looking forward to retiredcoders updates

[Belkacar]

...

You are aware that a single GPU or even only a few handful of them won't cut the cake for real puzzle #135?

I think about that a lot, but part of my brain goes to the scene from dumb and dumber - "So you're saying there's a chance" and I wonder while I am heating my room when it's cold out that my 4090 might just find that magical cross over and get a hit. 

Trying some novel stuff out to see what happens, and seeing what kind of clever bits everyone else is thinking about.  Looking forward to retiredcoders updates

The boss wrote that the current version of the program is not designed to find the 135th puzzle.

[RetiredCoder]

The boss wrote that the current version of the program is not designed to find the 135th puzzle.

You can use RCKang as a start point, but you must apply some efforts to make it good for #135: add loading/saving points, add network part, create a server part because you can't crack it with a single system.
It's not so difficult comparing to efficient cuda secp256k1 programming.
If you want free open source ready-to-use solution to crack #135 quickly - I wish you luck 

[Ykra]

You can use RCKang as a start point, but you must apply some efforts to make it good for #135: add loading/saving points, add network part, create a server part because you can't crack it with a single system.
It's not so difficult comparing to efficient cuda secp256k1 programming.
If you want free open source ready-to-use solution to crack #135 quickly - I wish you luck  

This is exactly how I started and after getting comfortable with understanding most of the source (or at least being able to follow it), testing tuneable options and reading too much CUDA documentation it's led me down the path of building my own solution although heavily referenced against RCs project and borrowing a large amount still.
This has helped me (as of yesterday) reach a solution that hit 12.2 GKeys/s on a RTX 5090, vastly behind RCs private build (and likely far behind kTimesG's if he has tested on a 5090) but a good 20%+ ahead of the public RCKang solution.

There are a small amount of knowledgeable guys here that help, such as RC and kTimesG that have provided code and/or hints to improving a solution or what to look into so if you want to develop something better yourself, review the source and read what the guys in the know are saying without just throwing it into a LLM and copy pasting the output.

[ficklebellpepper]

...

You are aware that a single GPU or even only a few handful of them won't cut the cake for real puzzle #135?

I think about that a lot, but part of my brain goes to the scene from dumb and dumber - "So you're saying there's a chance" and I wonder while I am heating my room when it's cold out that my 4090 might just find that magical cross over and get a hit. 

Trying some novel stuff out to see what happens, and seeing what kind of clever bits everyone else is thinking about.  Looking forward to retiredcoders updates

The boss wrote that the current version of the program is not designed to find the 135th puzzle.

I did implement saving/loading, but not the network bit, but im not sure if RC is also saying you need a server farm because you need the computer power | you're looking for a single molecule in an ocean the size of galaxy | or a single rig just can't load/save store that big of data.  I got it working for a single GPU, and using a local QRNG, just for fun  - I plan to test on known puzzles to see if there is any advantage to use a QRNG, but lots of projects and so little time.  If I find any interesting results ill post them here