Crowdsourcing an attack on ECC2K-130, to find out what it really costs
Read the paper on IACR ePrint 2026/2469
We are running a distributed attack on ECC2K-130, an elliptic curve challenge published in 1997 that nobody has solved. We need machines. If you have one sitting idle, it can walk the curve: ecc-challenges.com.
The experiment is not whether we break it. It is what the attack costs. Certicom priced this problem at 2.7 billion machine days, in 1997 computers. That number has been cited ever since, and nobody appears to have checked it against hardware people actually own.
One machine measures a benchmark. A fleet measures an attack.
What we changed
Not the algorithm. That is decades old and untouched. We rebuilt the layer underneath: how a field element is stored, how the arithmetic is scheduled against the way a processor issues instructions, where the work actually goes rather than where the operation count says it goes.
That is where estimates rot. Our client does far more work per second than anything available when the 1997 figure was written, with no change to the mathematics. The measurements and the method are published, linked at the end.
What transfers
ECC2K-130 protects nothing. Solving it leaves every real system exactly where it was.
What makes it worth the electricity is that none of the hard parts are specific to it. Storage layout, instruction scheduling, finding the real bottleneck rather than the one the operation count predicts: that work carries. The same engine already runs against the rest of the challenge catalog and against the NIST curves that are in everything you use, and the estimates attached to those are stale for the same reasons.
So this is the measurement that calibrates the next one. Attack one curve carefully enough and you learn what to trust about the ones that matter. Estimates age; the mathematics usually does not, and telling those apart takes measurements rather than citations.
What your machine does
It walks the curve, discards almost everything it computes, and reports the rare points worth keeping. When two walks anywhere in the world reach the same rare point, that collision gives up the key.
Every point you report can be recomputed from the seed it started at, so nothing has to be taken on faith, including ours.
What runs it
- NVIDIA, Turing through Blackwell, on Linux
- Apple silicon, via Metal, on macOS
- Any CPU on Linux, macOS or Windows, x86-64 or ARM
AMD and Intel GPUs do not work. There is no ROCm, OpenCL or SYCL backend, so a Radeon or an Arc contributes nothing. The CPU in that machine still counts.
Setup instructions are here. Use the campaign identifier shown on the site.
What it costs you
Not a background task. On a CPU it takes every core unless told otherwise. On a GPU it takes a card and works it.
Stopping is clean. Send a stop signal, give it two minutes, and it finishes its round, uploads, and saves its place. Kill it outright and you lose at most that round.
We have not published power draw or bandwidth figures. We would rather say so than estimate them. If you are weighing this against an electricity bill, you are weighing it against a number we have not given you.
The odds
A random search has no finish line. Half the expected work gives roughly an 18% chance the collision has already appeared. All of it, about 54%. One search in ten needs 1.7 times the expected work.
Nobody can give you a date. More machines buy a shorter wait, not a certainty. Live counts are on the status page.
The prize
Certicom's terms put $20,000 on this challenge, divided evenly among the group that solves it. If it pays, we split it with everyone who contributed.
The terms are from 2009, Certicom can change or cancel them without notice, claiming would require every contributor's real name, postal address and phone number, and nobody has confirmed anyone still administers it. Assume there is no money.
How big this is
The early challenges fell fast, one person or a handful at a time. ECC2-79 six weeks after publication, ECC2-89 in February 1998, ECC2K-95 that May, ECC2-97 in September 1999.
Then they outgrew single machines. About 1,300 people pooled their computers to take ECC2K-108 in April 2000. ECC2-109 needed around 2,600 people and seventeen months, finishing in April 2004.
Nothing on the list has been solved since.
ECC2K-130 is the least work of the ones still standing, and Certicom's own estimate puts it at 2.7 billion machine days against 1.3 million for ECC2K-108. Two thousand times the largest problem a crowd has ever finished, and twenty-two years since anyone finished one at all.
That is the size of the thing, and it is why this page exists.
Isogeny Labs wrote the client and runs the campaign. The code and the measurements are public.
Take part · How the method works · The speed measurements
Read the paper on IACR ePrint 2026/2469