Impressive. I wonder the methodology. Algorithmic improvements? More probably just an implementational optimisation. Last RSA record was due to special q sieving methods if I recall well, some 3k core hours.
I hope there’s a theoretical improvement behind the result.
So RSA 260 is about 2-3 times harder than RSA 250, which was solved in 2700 core hours in 2020, so it’s probably no algorithmic improvements, just a tweak here and there plus faster hardware.
> cuda-sieve is an experimental, standalone CUDA implementation of the lattice-sieving relation-collection pipeline used by the Number Field Sieve. It builds factor bases, sieves both sides of a special-q lattice, performs trial division and GPU cofactorisation, and emits relations for msieve.
