Bitcoin puzzle transaction ~32 BTC prize to who solves it

[Jorge54PT]

This is nothing new, and I've asked this question several times, and no satisfactory answer has been given in the forum. And it doesn't require coding, all calculations can be done with a calculator.
5HpHagT65TZzG1PH3CSu63k8DbpvM6sdcMk3rQ8hVnTJAphn1wQ
5Km2kuu7vtFDPpxywn4u3NLpbr5jTnTiwHZAqh7Go9SJu8y7XMR
1FAiLLQsUoJwVYeYPUuoceyAyNVENv3b5w

L5oLkpV3aqBjhki6LmvChTCV6odtRDex6c4zRv2gigEhzkSSVKEP
KwDiBf89QgGbjEhKnhXJuH7LrciWTivUb4Z1b36yk3d4nHheQ1AU
1PijtU6wcyyZYiPcjvTaRBbC6xMMkd5Cj

Why did you post these? You have WIFS/Private Keys that aren't in the curve. Maybe I am missing something.

´´Heklo friend WanderingPhilospher

Its possible you tell me wath diference i use or not the parameter in VBCr.exe  -r rekey? is better with rekey or dont use that paraneter? Any advantage if i use?  thank you so much  (my 3060 make 1070 Mkeys/s)

[Virtuose]

wtf i try the nvidia h200 and is just doing 2.8–3.0 billion keys/sec i was expecting more how can i gpu of  $31,000 to $32,000 do such low or im doing something wrong??

The H200 is optimized for AI and tensor workloads, not integer-heavy operations like secp256k1.
Most GPU bruteforce tools aren't tuned for Hopper architecture, so you're likely not using its full potential.

For this kind of task, RTX 4090 or 5090 performs better per dollar unless you write custom CUDA code optimized for H200.

[brainless]

wtf i try the nvidia h200 and is just doing 2.8–3.0 billion keys/sec i was expecting more how can i gpu of  $31,000 to $32,000 do such low or im doing something wrong??

The H200 is optimized for AI and tensor workloads, not integer-heavy operations like secp256k1.
Most GPU bruteforce tools aren't tuned for Hopper architecture, so you're likely not using its full potential.

For this kind of task, RTX 4090 or 5090 performs better per dollar unless you write custom CUDA code optimized for H200.

In my view h200 gpu is more powerful then 5090 gpu
Due to 140gb ram, not always hash calc, load of ram also calc and result would be 5 time higher then 24gb ram
He need adjust proper -b -p -t
If he reply if he need info about these setting for use h200 optimal speed...

[Virtuose]

wtf i try the nvidia h200 and is just doing 2.8–3.0 billion keys/sec i was expecting more how can i gpu of  $31,000 to $32,000 do such low or im doing something wrong??

The H200 is optimized for AI and tensor workloads, not integer-heavy operations like secp256k1.
Most GPU bruteforce tools aren't tuned for Hopper architecture, so you're likely not using its full potential.

For this kind of task, RTX 4090 or 5090 performs better per dollar unless you write custom CUDA code optimized for H200.

In my view h200 gpu is more powerful then 5090 gpu
Due to 140gb ram, not always hash calc, load of ram also calc and result would be 5 time higher then 24gb ram
He need adjust proper -b -p -t
If he reply if he need info about these setting for use h200 optimal speed...

That's true technically, but even with custom CUDA kernels, the H200 is not cost-effective for secp256k1 bruteforce.
You're paying $30k+ for AI-focused tensor performance, not INT32 throughput.

A few RTX 5090s will outperform it per watt, per dollar, and even in raw key/s with optimized tools. H200 is impressive, just not for this job.

[kTimesG]

In my view h200 gpu is more powerful then 5090 gpu
Due to 140gb ram, not always hash calc, load of ram also calc and result would be 5 time higher then 24gb ram

Maybe after you convince a book to read itself and present you the audio summary. Pretty much the same as what you said there.

Fastest code doesn't even use RAM, latency is too high, and guess what, it's not even needed, and if it unfortunately becomes needed, the speed drops. Like, big time. So, best practice is to limit to maximum any access to RAM. Having exabytes or zettabytes of it does not help with anything here (I hear the "bsgs" whispers, it is indeed a fantastic idea to do lookups in some memory-based huge table, after each EC point gets computed! It definitely won't go all over the place and add latencies).

[Tony8989]

true i few GeForce RTX 5090  will do the job for much less, the h200 is for Ai task using 4 RTX 5090  i can achieve  1_382_400_000_000_000  x day

for now i try few gpu

NVIDIA RTX A4000   818.15 MKey/s

GeForce RTX 5060 Ti  973.68 MKey/s

GeForce RTX 3080  1574.46

GeForce RTX 3060  656.09 MKey/s

GeForce RTX 3090 1324.28 MKey/s

GeForce RTX 4090   3061.96 MKey/s

GeForce RTX 4070 1286.45 MKey/s

GeForce RTX 5080  1988.08 MKey/s

GeForce RTX 3070  1014.12 MKey/s

NVIDIA RTX A5000  1338.56 MKey/s

GeForce RTX 5090  3956.87 MKey/s  top

GeForce RTX 4080 SUPER 1725.58 MKey/s

GeForce GTX 1070 119.88 MKey/s

NVIDIA H200 NVL  2871.53 MKey/s


any way to cut the key or skip consecutive number and letter smart way?


bro i try  -b -p -t  but
PWo3JeB9jrGwfHDNpdGK54CRas7fsVzXU -b 512 -t 512 -p 1090
[2025-08-03.19:30:45] [Info] Compression: compressed
[2025-08-03.19:30:45] [Info] Starting at: 00000000000000000000000000000000000000000000004EC2832F6F6B400000
[2025-08-03.19:30:45] [Info] Ending at:   00000000000000000000000000000000000000000000004EC2867589EDFFFFFF
[2025-08-03.19:30:45] [Info] Counting by: 0000000000000000000000000000000000000000000000000000000000000001
[2025-08-03.19:30:45] [Info] Initializing NVIDIA H200 NVL
[2025-08-03.19:30:45] [Info] Generating 285,736,960 starting points (10900.0MB)
Segmentation fault


b 256 -t 256 -p 1090
[2025-08-03.19:28:31] [Info] Compression: compressed
[2025-08-03.19:28:31] [Info] Starting at: 00000000000000000000000000000000000000000000004EC2832F6F6B400000
[2025-08-03.19:28:31] [Info] Ending at:   00000000000000000000000000000000000000000000004EC2867589EDFFFFFF
[2025-08-03.19:28:31] [Info] Counting by: 0000000000000000000000000000000000000000000000000000000000000001
[2025-08-03.19:28:31] [Info] Initializing NVIDIA H200 NVL
[2025-08-03.19:28:31] [Info] Generating 71,434,240 starting points (2725.0MB)
[2025-08-03.19:28:44] [Info] 10.0%
[2025-08-03.19:28:44] [Info] 20.0%
[2025-08-03.19:28:44] [Info] 30.0%
[2025-08-03.19:28:44] [Info] 40.0%
[2025-08-03.19:28:45] [Info] 50.0%
[2025-08-03.19:28:45] [Info] 60.0%
[2025-08-03.19:28:45] [Info] 70.0%
[2025-08-03.19:28:45] [Info] 80.0%
[2025-08-03.19:28:45] [Info] 90.0%
[2025-08-03.19:28:45] [Info] 100.0%
[2025-08-03.19:28:45] [Info] Done
NVIDIA H200 NVL  7070 / 143186MB | 1 target 2871.53 MKey/s

[2025-08-03.19:24:33] [Info] 80.0%
[2025-08-03.19:24:33] [Info] 90.0%
[2025-08-03.19:24:33] [Info] 100.0%
[2025-08-03.19:24:33] [Info] Done
[2025-08-03.19:24:33] [Info] Error: too many resources requested for launch

[brainless]

true i few GeForce RTX 5090  will do the job for much less, the h200 is for Ai task using 4 RTX 5090  i can achieve  1_382_400_000_000_000  x day

for now i try few gpu

NVIDIA RTX A4000   818.15 MKey/s

GeForce RTX 5060 Ti  973.68 MKey/s

GeForce RTX 3080  1574.46

GeForce RTX 3060  656.09 MKey/s

GeForce RTX 3090 1324.28 MKey/s

GeForce RTX 4090   3061.96 MKey/s

GeForce RTX 4070 1286.45 MKey/s

GeForce RTX 5080  1988.08 MKey/s

GeForce RTX 3070  1014.12 MKey/s

NVIDIA RTX A5000  1338.56 MKey/s

GeForce RTX 5090  3956.87 MKey/s  top

GeForce RTX 4080 SUPER 1725.58 MKey/s

GeForce GTX 1070 119.88 MKey/s

NVIDIA H200 NVL  2871.53 MKey/s


any way to cut the key or skip consecutive number and letter smart way?

What command you apply ?
And one same command for all gpu u tested ?
Or different command for above listed gpu
I am talking only about bitcrack

[Tony8989]

I used BitCrack with the same base command but manually tuned the -t, -b, and -p values for each GPU. I gradually increased these parameters until the GPU either crashed or was about to crash, then backed off slightly to get the maximum stable performance.

Someone in the group chat managed to push their RTX GPU to about 5356.87 MKey/s, which is higher than what I got. They were likely using a different GitHub project like KeyHunt-Cuda, which can sometimes achieve better raw speed. However, in my tests, I noticed that KeyHunt-Cuda occasionally skips some keys, especially with test/fake keys, which could lead to missing targets.

BitCrack, on the other hand, is very thorough and doesn't skip keys, so I prefer it for reliability, even if the speed is slightly lower

[brainless]

-b 132. -t 1024 - p 2048
-b 132. -t 1024 - p 1536
-b 132. -t 1024 - p 1024
-b 132. -t 1024 - p 512

-b 132. -t 512 - p 2048
-b 132. -t 512 - p 1536
-b 132. -t 512 - p 1024
-b 132. -t 512 - p 512

Start from top, one by one and tell me which one start working with ur h200

[Tony8989]

Bro Tried all these configs — all crash with too many resources requested for launch 

/bin/cuBitCrack --keyspace 5206e53083080a29e7:5206eef59f4d2c09e8 1PWo3JeB9jrGwfHDNpdGK54CRas7fsVzXU -b 132 -t 256 -p 6048
[2025-08-04.22:05:51] [Info] Compression: compressed
[2025-08-04.22:05:51] [Info] Starting at: 00000000000000000000000000000000000000000000005206E53083080A29E7
[2025-08-04.22:05:51] [Info] Ending at:   00000000000000000000000000000000000000000000005206EEF59F4D2C09E8
[2025-08-04.22:05:51] [Info] Counting by: 0000000000000000000000000000000000000000000000000000000000000001
[2025-08-04.22:05:51] [Info] Initializing NVIDIA H200 NVL
[2025-08-04.22:05:51] [Info] Generating 204,374,016 starting points (7796.2MB)
[2025-08-04.22:06:30] [Info] 10.0%
[2025-08-04.22:06:31] [Info] 20.0%
[2025-08-04.22:06:31] [Info] 30.0%
[2025-08-04.22:06:32] [Info] 40.0%
[2025-08-04.22:06:32] [Info] 50.0%
[2025-08-04.22:06:32] [Info] 60.0%
[2025-08-04.22:06:32] [Info] 70.0%
[2025-08-04.22:06:32] [Info] 80.0%
[2025-08-04.22:06:33] [Info] 90.0%
[2025-08-04.22:06:33] [Info] 100.0%
[2025-08-04.22:06:33] [Info] Done
NVIDIA H200 NVL  19244 / 143156MB | 1 target 3079.91 MKey/s (11,444,944,896 total) [00:00:01]

[neoman602]

I've updated the code

and made each hex char to be generated on its own(so its random from 0 to 15 for each hex char)

removed the left-shift transformations as it is slowing down and not effective

now there is only:

1) vertical hex rotation 16 times
2) reversing hex
3) hex inversion(in bits)

https://github.com/puzzleman22/bitcoin_puzzle_transformation - give it a taste, maybe you get lucky

[brainless]

Bro Tried all these configs — all crash with too many resources requested for launch 

/bin/cuBitCrack --keyspace 5206e53083080a29e7:5206eef59f4d2c09e8 1PWo3JeB9jrGwfHDNpdGK54CRas7fsVzXU -b 132 -t 256 -p 6048
[2025-08-04.22:05:51] [Info] Compression: compressed
[2025-08-04.22:05:51] [Info] Starting at: 00000000000000000000000000000000000000000000005206E53083080A29E7
[2025-08-04.22:05:51] [Info] Ending at:   00000000000000000000000000000000000000000000005206EEF59F4D2C09E8
[2025-08-04.22:05:51] [Info] Counting by: 0000000000000000000000000000000000000000000000000000000000000001
[2025-08-04.22:05:51] [Info] Initializing NVIDIA H200 NVL
[2025-08-04.22:05:51] [Info] Generating 204,374,016 starting points (7796.2MB)
[2025-08-04.22:06:30] [Info] 10.0%
[2025-08-04.22:06:31] [Info] 20.0%
[2025-08-04.22:06:31] [Info] 30.0%
[2025-08-04.22:06:32] [Info] 40.0%
[2025-08-04.22:06:32] [Info] 50.0%
[2025-08-04.22:06:32] [Info] 60.0%
[2025-08-04.22:06:32] [Info] 70.0%
[2025-08-04.22:06:32] [Info] 80.0%
[2025-08-04.22:06:33] [Info] 90.0%
[2025-08-04.22:06:33] [Info] 100.0%
[2025-08-04.22:06:33] [Info] Done
NVIDIA H200 NVL  19244 / 143156MB | 1 target 3079.91 MKey/s (11,444,944,896 total) [00:00:01]

It's mean your device max support -t 256
Now expand -p 6048 to on word
-p 6048 + 64. Or 32 every time test with add into -p 6112
When last stop that time just minus 32 for get final -p value workable
By these setting u will get highest speed print ur last final value of -p

[Tony8989]

Thanks for the tip! Quick question — wouldn’t it be easier to just use RTX 5090 or 4090 GPUs instead? They’re more stable and probably easier to tune, right? Or does the H200 NVL have big advantages despite the tricky tuning?

[brainless]

Thanks for the tip! Quick question — wouldn’t it be easier to just use RTX 5090 or 4090 GPUs instead? They’re more stable and probably easier to tune, right? Or does the H200 NVL have big advantages despite the tricky tuning?

In my view h200 better
If u r using 5090 -b is 170
Rest t and p adjust as per test
But h200 have 140gb ram, which can be used to make ur finding spot big definitely in puzzle 135

[brainless]

Ever gpu have their own structure, every need to set their b t p
It's like load design as per country state cities
Wrong set b t p, u will waste resources and missing keys

[Tony8989]

ok!! best set up for 4090 and 5090??

[brainless]

ok!! best set up for 4090 and 5090??

-b 170 -t 256 -p 4096
For 5090, if no fault, then add 32 in 4096 and test with every time by adding 32 till it crash
You will reach maxmium speed of you 5090

[benjaniah]

This is my hardware
2 x AMD EPYC  64 cores each
1.5 T RAM
-n 0x1000000000000 -k 16384

I think you might be able to squeeze a little bit more speed using -n 0x4000000000000 -k 8192. If you saved your 1TB bloom filter using -S, you can re-use it with the different -n and -k values. You have another 0.5 TB ram available, have you tried -n 0x4000000000000 -k 12000 so you can max out your RAM? Even then, you'll run out of money or die before 135 is found.

[oddstake]

This is my hardware https://ibb.co/27kkq4yT
2 x AMD EPYC  64 cores each
1.5 T RAM
-n 0x1000000000000 -k 16384

I think you might be able to squeeze a little bit more speed using -n 0x4000000000000 -k 8192. If you saved your 1TB bloom filter using -S, you can re-use it with the different -n and -k values. You have another 0.5 TB ram available, have you tried -n 0x4000000000000 -k 12000 so you can max out your RAM? Even then, you'll run out of money or die before 135 is found.

I will try with -n 0x4000000000000 -k 8192, but for every different command I need to wait a couple of hours for the bloom filter to fill the RAM.
I tried also with all the 1.5T RAM but the speed is considerably slower because on a dual socket system, each CPU typically has its own memory controller and dedicated memory channels. This architecture is known as NUMA. Latency increases and bandwidth decreases when a CPU accesses the memory attached to the other CPU.
So I'm running 2 separate keyhunt apps with these commands : numactl --cpunodebind=0 --membind=0 ./keyhunt0 -m bsgs -f 135.txt etc etc (1T RAM)
and numactl --cpunodebind=1 --membind=1 ./keyhunt1 -m bsgs -f 135 etc... (512GB RAM).
You are right, 135  is very hard but better to try my luck on random than going sequentially specially on these high ranges puzzles.

Just ordered from Ebay China an EPYC 9965 with 192 cores / 384 threads and 1T of DDR5 at 6400Mts/s , the speed will be about 5-6x higher on this new setup. More than 15k euros put into these puzzles without even knowing I will get 1 penny back.

[Virtuose]

This is my hardware
2 x AMD EPYC  64 cores each
1.5 T RAM
-n 0x1000000000000 -k 16384

I think you might be able to squeeze a little bit more speed using -n 0x4000000000000 -k 8192. If you saved your 1TB bloom filter using -S, you can re-use it with the different -n and -k values. You have another 0.5 TB ram available, have you tried -n 0x4000000000000 -k 12000 so you can max out your RAM? Even then, you'll run out of money or die before 135 is found.

I will try with -n 0x4000000000000 -k 8192, but for every different command I need to wait a couple of hours for the bloom filter to fill the RAM.
I tried also with all the 1.5T RAM but the speed is considerably slower because on a dual socket system, each CPU typically has its own memory controller and dedicated memory channels. This architecture is known as NUMA. Latency increases and bandwidth decreases when a CPU accesses the memory attached to the other CPU.
So I'm running 2 separate keyhunt apps with these commands : numactl --cpunodebind=0 --membind=0 ./keyhunt0 -m bsgs -f 135.txt etc etc (1T RAM)
and numactl --cpunodebind=1 --membind=1 ./keyhunt1 -m bsgs -f 135 etc... (512GB RAM).
You are right, 135  is very hard but better to try my luck on random than going sequentially specially on these high ranges puzzles.

Just ordered from Ebay China an EPYC 9965 with 192 cores / 384 threads and 1T of DDR5 at 6400Mts/s , the speed will be about 5-6x higher on this new setup. More than 15k euros put into these puzzles without even knowing I will get 1 penny back.

If you can spend that much, throwing it all on CPU is honestly a waste.
Even the fastest EPYC setups are way behind GPUs in raw keyrate for secp256k1.
A single RTX 5090 will crush any 192-core CPU in keys/sec, and for less power and less cash.
Unless you’re doing something highly specialized on CPU, brute-forcing with GPU is just smarter. 😄