Nicholas Carlini在一台1995年发布的200 MHz Intel Pentium Pro商用台式机(配备64MB RAM)上成功运行了现代Stockfish国际象棋引擎1。实验结果显示,该引擎的估计Elo评分达到3300 ± 87,高于加里·卡斯帕罗夫约2800 Elo的历史峰值1。
这项研究旨在探讨AI领域中算法效率与算力规模的关系1。数据显示,从旧版引擎到现代引擎的总Elo增益约为+1335分,其中+720分的提升直接归因于编程与算法改进,而非硬件升级1。作为对比,Deep Blue曾使用30个P2SC处理器和480个定制VLSI协处理器,每秒分析超过2亿个局面1。测试还表明,将1996年的Crafty引擎运行在现代Ryzen 7处理器上仅获得2614 Elo,而将现代算法移植回Pentium Pro并在30秒思考时间内运行时,最高可达到3180 Elo1。
Carlini指出,大部分代码编写工作由Fable 5.1、Opus 5.5和GPT-6等大语言模型辅助完成1。
Nicholas Carlini demonstrated that a modern Stockfish chess engine running on a 1995-era Intel Pentium Pro processor achieved an estimated Elo rating of 3,300 ± 87, significantly surpassing the approximately 2,800 Elo rating associated with Garry Kasparov 1. The test utilized a commercial desktop computer featuring a 200 MHz CPU and 64 MB of RAM, proving that software advancements alone can yield overwhelming competitive advantages over historical hardware configurations 1. This result highlights that algorithmic efficiency contributed +720 Elo to the total performance gain, far outweighing the impact of raw hardware upgrades in this specific comparison 1.
In contrast to the experimental setup, IBM’s Deep Blue from 1996 relied on massive specialized hardware, employing 30 P2SC processors and 480 custom VLSI coprocessors to analyze more than 200 million positions per second 1. To isolate the effects of code versus hardware, Carlini compared the 1996 Crafty engine running on a modern Ryzen 7 processor (scoring 2,614 Elo) against the modern Stockfish algorithm ported to the Pentium Pro (reaching up to 3,180 Elo at 30 seconds per move), confirming that the modern algorithm dominated even on obsolete silicon 1. The complex task of porting the modern code to the legacy architecture was largely assisted by large language models, including Fable 5.1, Opus 5.5, and GPT-6 1.
评论
还没有评论,欢迎留下第一条。