Cognition研究团队通过AI代理Devin开发的高性能GPU晶格筛选器，成功因数分解了RSA-260（260位数），刷新了公开RSA因数分解的纪录1。这一成就超越了2020年2月创造的RSA-250纪录1。
项目从8月13日启动至9月3日完成，历时约三周1。整个因数分解过程耗费约4,900个GPU计算日，相当于13.5个GPU年，按现行市场价格折合约40万美元1。相比之前的公开技术，GPU晶格筛选的成本下降了10倍1。在这个过程中，Devin AI代理通过192个会话发送了82,702条指令，共计502,887个字符1。
分解结果为两个大质数的乘积1。研究团队指出，超级计算机或AI前沿实验室可能能以约3,000万美元的成本因数分解RSA-1024，但RSA-2048的难度仍比RSA-1024高约十亿倍1。此项目展示了GPU优化在现代密码分析中的潜力，涵盖GPU自适应多项式选择、GPU优化晶格筛选器、GPU块Wiedemann实现和GPU加速平方根等多项技术突破1。
Cognition's research team has successfully factored RSA-260, a 260-bit number, establishing a new public record for RSA factorization and surpassing the previous benchmark of RSA-250 set in February 2020.1 The achievement was accomplished through the development of a high-performance GPU lattice sieve, with computational assistance from the AI agent Devin, completing the project in approximately three weeks at an estimated cost of around $400,000.1
The factorization process consumed approximately 4,900 GPU-days, equivalent to 13.5 GPU-years at current market rates.1 This represents a tenfold reduction in cost compared to previously available public techniques for GPU lattice sieve implementation.1 During the three-week effort spanning from August 13 through September 3, the Devin AI agent issued 82,702 words of instructions across 192 sessions to orchestrate the optimization and computational workflow.1 Multiple GPU-accelerated components were developed to support the factorization, including adaptive polynomial selection, lattice sieve optimization, block Wiedemann implementation, and accelerated square root computation.1
The researchers project that supercomputers or AI-capable research laboratories could potentially factor RSA-1024 at approximately $30 million in cost, though RSA-2048 remains roughly one billion times more difficult to factor than RSA-1024.1
评论
还没有评论，欢迎留下第一条。