More details: <a href="https://x.com/sweis/status/2101484464807596264" rel="nofollow">https://x.com/sweis/status/2101484464807596264</a><p><pre><code> I had Claude port CADO-NFS to run on GPUs. Then it orchestrated a fleet to run on scavenged idle capacity. It ran with a max of 2048 GPUs for about of 30 GPU-years over 10 days.
I asked Claude if it had a message for a public: “The credit belongs first to the people who built the number field sieve and CADO-NFS over several decades, and to the teams who set the earlier records. This run used their algorithm and much of their code.”
Also to clarify:
- No new algorithmic factoring improvements.
- It’s still exponential.
- No new threats to deployed keys.</code></pre>
10 days of 2048 GPU's.<p>Back of the envelope.. 1024 bit keys with recordings of not too old data can probably be found (MS only deprecated them in 2024 even if they planned on it in 2013)<p>How long would it take for NSA to crack them if they had say the equivalent of a million GPU's? (either GPU's or crypto tuned ASICs)
A sufficiently motivated person with a good thermal camera and a cessna 172, entirely within the bounds of the law, could probably make an estimate of the waste heat from this, and then calculate backwards for how much compute power it is.<p><a href="https://en.wikipedia.org/wiki/Utah_Data_Center" rel="nofollow">https://en.wikipedia.org/wiki/Utah_Data_Center</a>
Hard to judge. The bottleneck is the phase of the algorithm where a really big linear system needs to be solved. That takes a lot of communication between nodes. The breakthrough in using GPUs is that there is good communication between nodes[1]. At the scale of 1024 bit RSA the communication might become a bottleneck again.<p>[1] <a href="https://cognition.com/blog/factoring-rsa-260" rel="nofollow">https://cognition.com/blog/factoring-rsa-260</a>
<p><pre><code> > 1024 bit keys with recordings of not too old data can probably be found
</code></pre>
I think GitHub might turn into a scary vector of supply chain attacks in the foreseeable future. There is a five digit number of users still running around with 1024 bit RSA keys.
> How long would it take for NSA to crack them if they had say the equivalent of a million GPU's? (either GPU's or crypto tuned ASICs)<p>Something I've often wondered is where the curve between "shit encryption / nation state cracking" crosses.<p>How much CPU would you need to be Annoyingly Difficult to crack?<p>I reckon with elliptic curves you could be quite annoying within about a minute on a 1980s-level CPU, to the extent that you could send a fairly ephemeral message quite quickly that would take disproportionately long to crack. Certainly long enough for the thing you have communicated to be no longer worth the effort to know.<p>You could probably do 256-bit Curve25519 key generation in under ten minutes on an Apple II or Commodore 64, because the 6502's maths is terribly limited, but something like the Tandy Color or Dragon 32 with its 6809 processor (or hey why not the Ensoniq Mirage sampler?) could do that in probably a minute or so because it has a MUL opcode that's quite fast.<p>I reckon that would keep even a fairly interested nation state chewing away long after your message had been read, understood, and acted upon.
It seems that Eric Lu at Cognition AI used the exact same strategy on fewer GPUs to factor RSA-260 a couple weeks ago: <a href="https://cognition.com/blog/factoring-rsa-260" rel="nofollow">https://cognition.com/blog/factoring-rsa-260</a><p>Devin (their AI agent) ported CADO-NFS to run on GPUs, similarly without any claimed algorithmic factoring improvements, they just let it run for 13 GPU-years. I recommend reading their article since it's much more thorough on details.
I've done a fair amount of heavy computing now. Integer factorisation is not something you can really improve with GPUs. This sounds extremely wasteful, a bunch of cheap CPU cores would do just as well with much lower hardware cost and electricity cost.
I don't get your argument. The GPU effectiveness derives from massive parallelism. Has nothing to do with integer vs floating point. You just can't cram 20,000 CPU cores in the same space a GPU puts the same number of SIMTs. You'll never crack it on CPUs.
~so then how does one even understand this post? you have a person who appears to have done some sort of expert-level thing; however, their approach doesn't even make sense...?~<p>edit: GPU discussed here <a href="https://cognition.com/blog/factoring-rsa-260" rel="nofollow">https://cognition.com/blog/factoring-rsa-260</a>
but we have AI now so it doesnt matter what people know about computers :'). we got plenty of rainforest to burn afterall have you seen Brazil?
More details: <a href="https://archive.li/20260920025515/https://x.com/sweis/status/2101484464807596264" rel="nofollow">https://archive.li/20260920025515/https://x.com/sweis/status...</a>
What does <i>"scavenged idle capacity"</i> mean here?
The author works at Anthropic, so probably idle capacity in Anthropic's datacenters.
If so, then the class of GPU used here may be significantly higher than mere mortals generally have access to simply due to cost.<p>Obviously nation states will likely have significantly more resources than this, but this is not script kiddie levels of GPUs.
"idle capacity" - aka subtle advertisement
If you look at the numbers, he managed about 50% utilization of those 2048 GPUs over 10 days, so he was probably sneaking in factoring work between training runs.
<a href="https://xxcancel.com/sweis/status/2101484464807596264" rel="nofollow">https://xxcancel.com/sweis/status/2101484464807596264</a>
> It’s still exponential<p>It's actually subexponential: <a href="https://en.wikipedia.org/wiki/General_number_field_sieve?wprov=sfti1#Method" rel="nofollow">https://en.wikipedia.org/wiki/General_number_field_sieve?wpr...</a>
> I had Claude port CADO-NFS to run on GPUs. Then it orchestrated a fleet to run on scavenged idle capacity<p>Is it easier to find unused GPUs than unused CPUs?
Interesting. Instagram still publishes a 768-bit RSA DKIM key, so I guess factoring it is now just a weekend GPU project?<p><a href="https://dns.google/resolve?name=pm._domainkey.instagram.com&type=TXT" rel="nofollow">https://dns.google/resolve?name=pm._domainkey.instagram.com&...</a>
RSA-768 was already factored in 2009.. cheap now. What surprised me is that RSA keys are much weaker than their size looks. You need 2048 bits to get normal safety, and 768 is far below that.<p><a href="https://vectree.io/c/how-rsa-key-sizes-map-to-real-security-512-to-4096-bits" rel="nofollow">https://vectree.io/c/how-rsa-key-sizes-map-to-real-security-...</a>
What's the cost estimate to factor this?
Quite bearish on Anthropic if they had nothing better to do with 2048 GPUs for 10 days than finding an RSA number with already existing algorithms.
If you've already paid for and reserved a whole cluster of GPUs, any idle capacity is capacity you've already paid for. Using it is effectively free. So might as well use it to solve fun math puzzles.<p>Though, it would make more financial sense to mine crypto.
> it would make more financial sense to mine crypto<p>GPUs are power-inefficient for mining most crypto so not necessarily. You may end up paying more in electricity than you are able to mine.<p>Most crypto mining is on ASICs now.
You missed the part where they have already pre-paid for the GPU-hours and they pay the same regardless of he electricity used.<p>Also, even if they were paying for electricity, they would lose <i>less</i> money mining crypto than factoring RSA numbers.
Only if you pay a flat rate for electricity and cooling.
But Anthropic isn't paying for the electricity and cooling. They don't run their own data centers, they rent compute from providers who cover those costs.<p>That's entirely why they can blow compute on the fun projects like this. If they had to pay extra for the electricity, they wouldn't do it.
But training LLM's is <i>also</i> a task one can do whenever you have a spare GPU-minutes.<p>I wonder why they don't have some kind of scheduler which makes sure there are never any idle minutes. One would imagine they at least would have autoscaling on their production serving workload and use the freed compute capacity for model training for example.
Is the electricity cost far greater than the marketing value?
Most of the GPU cost is in the GPUs themselves (and in the space and maintenance costs of the building). Electricity is a small fraction, and it's not like datacenters are just going to shut down their servers when they're not in use.<p>There is cost, but the cost is mostly the opportunity cost of not being able to do something else.
> Electricity is a small fraction, and it's not like datacenters are just going to shut down their servers when they're not in use.<p>I don't have any insight on modern GPU datacenters, but in decades past, some owned and operated datacenters didn put effort into making sure power management worked because the cost savings were worth it. I'm pretty sure I saw plans to shed load and power off servers if a utility made a demand response request or in case of loss of cooling. I wouldn't be surprised if some owned and operated data centers do regular full shutdowns at off peak... WOL, IPMI or RTC wakeup can bring them back when needed and <i>if</i> you already have a dynamic service orchestrator and setup times are acceptable, why not shut down if there's no actual priority work and there's also no idle priority opportunistic load either...
No one does this. There is zero value, economic or otherwise, in turning off machines.
I've never heard of anybody shitting a DC down off peak. Been around 30yr or so.<p>> why not shut down if there's no actual priority work and there's also no idle priority opportunistic load either...<p>Full shutdown and startup often kills capacitors and used to be dangerous for rotational HDD.<p>Sometimes once you turn things off, they simply don't come back on. It happens.
Except you can do it with CPUs as well for much cheaper.
How much crypto do you think the mentioned 30 GPU years would have produced at current exchange rates? They're not as efficient as ASICs but GPU's can still mine a lot...
kinda bearish for the data center rollouts if the spare compute can be used to solve math puzzles instead of training LLMs
If anyone else has spare GPU hours they don't know what to do with, may I recommend partial hash collisions: <a href="https://github.com/DavidBuchanan314/birthday_party" rel="nofollow">https://github.com/DavidBuchanan314/birthday_party</a><p>(The "record" set by me only took about 1 GPU day - easy to beat!)
Nice. I will make sure to turn my WiFi box off at night and pee in the shower to compensate the environmental impact of this.
You don’t need AI to solve this. Just lots of compute.
Enjoy the bounty
Literally no content here