随着六大超大规模计算公司今年计划投入超过700亿美元用于计算基础设施——约为2025年纪录的两倍——业界对计算能力的需求达到新高度1。但传统的技术进步路径正在放缓：摩尔定律已放缓，丹纳德定律已结束，库梅定律开始下滑1。尽管晶体管密度在过去80年增长了10个数量级，达到每平方毫米1亿多个晶体管1，计算机芯片仍然远未接近由量子力学、广义相对论和信息论决定的理论物理极限1。
计算技术面临四项根本性限制。在计算速度方面，理论极限比当今芯片快35个数量级1。能量消耗受制于Landauer最小值，该原理规定擦除一比特必须释放最少2.9×10^-21焦耳的热量1。存储密度方面，Bekenstein极限表明某区域可存储约10^42比特信息1。通信速率则由Holevo极限决定，光子流传输信息的最大速率约为10^17比特每秒1。
为突破这些限制，研究者正在探索多个方向。Vaire Computing公司在2025年测试了首款可逆逻辑芯片1，该方法有望降低能量消耗。DNA存储技术也展现潜力，以约30个原子存储一比特信息，远优于费曼早期估计的125原子1。量子计算等新型计算范式也被视为突破方向1。
The limits of computing technology remain far from exhaustion despite decades of incremental progress in chip design. While traditional transistor scaling has slowed considerably, with Moore's Law decelerating and Dennard scaling having ended, computer chips are still orders of magnitude away from the fundamental physical boundaries set by quantum mechanics, general relativity, and information theory.1 Six major hyperscale computing companies plan to invest over 70 billion dollars in computing infrastructure this year—approximately double the 2025 record—suggesting continued confidence in the industry's ability to drive performance gains.1
Current semiconductor density has advanced dramatically over eight decades, reaching more than 100 million transistors per square millimeter.1 Yet the theoretical speed limit for computation exceeds today's chips by 35 orders of magnitude.1 Several hard physical constraints define the frontier of computational possibility. The Landauer minimum dictates that erasing a single bit of information must release at least 2.9×10⁻²¹ joules of heat.1 The Bekenstein limit indicates that a given region of space can store approximately 10⁴² bits of information.1 The Holevo limit establishes that photons can transmit data at a maximum rate of roughly 10¹⁷ bits per second.1
Emerging technologies are beginning to address these constraints from multiple angles. Vaire Computing tested its first reversible logic chip in 2025, offering a pathway toward reduced energy dissipation in computation.1 DNA storage represents another promising direction, encoding approximately one bit of information per 30 atoms—substantially more efficient than the 125 atoms per bit initially estimated.1
评论
还没有评论，欢迎留下第一条。