華為「韜定律」：晶片技術從空間縮微轉向時間優化

🔑 增強重點摘要

•The Tao Law, also referred to as τ (tau) Scaling Law, redefines chip optimization by compressing characteristic time (τ) across the entire computing stack, from transistor switching to datacenter task completion, spanning 12 orders of magnitude.
•LogicFolding is a circuit-level implementation that distributes a chip's internal circuits, down to the gate and flip-flop level, across vertically stacked multiple wafer layers using ultra-precise hybrid bonding and Through-Silicon Vias (TSVs), rather than stacking independent dies.
•This approach was developed by Huawei as a strategic response to US sanctions restricting access to advanced EUV lithography, allowing them to achieve performance gains equivalent to advanced process nodes (e.g., 1.4nm-class by 2031) without relying on geometric scaling.
•Huawei claims LogicFolding has already demonstrated significant performance improvements in mass-produced chips, including a 55% increase in transistor density, 41% improvement in energy efficiency, and over 40% increase in SRAM frequency.
•Beyond LogicFolding, the Tao Law encompasses a broader "Four Foldings" concept (Circuit folding, Chip folding, System folding, Transistor folding) and system-level optimizations like the Unified Bus and Hi-ONE Optical Interconnect for AI data centers, aiming to reduce communication latency from microseconds to nanoseconds.

🛠️ 技術深入

Tau (τ) Scaling Theory: Defines time (τ) as the core optimization metric, aiming to compress characteristic time across four layers: Transistor (switching speed), Circuit (signal propagation delay), Chip (compute and memory access latency), and System (end-to-end communication and synchronization time).
LogicFolding Architecture:
- Mechanism: Physically folds traditional 2D circuit layouts into vertical, stacked structures, distributing a chip's internal circuits (down to gate and flip-flop level) across multiple active wafer layers.
- Interconnects: Replaces long 2D horizontal paths with short 3D vertical interconnects, significantly reducing RC delay and signal travel time.
- Fabrication: Requires manufacturing two complete chip wafers using the same process node, then grinding one wafer down to paper-thinness. Ultra-precise hybrid bonding fuses copper pads directly, and Through-Silicon Vias (TSVs) create vertical connections between layers.
- Performance Gains (claimed): Achieves a 55% increase in transistor density, 41% improvement in energy efficiency, 13% peak frequency increase, and over 40% SRAM frequency increase, without upgrading process nodes.
System-Level Optimization (Four Foldings):
- Circuit Folding: Shortens wires inside a chip.
- Chip Folding: Bonds different chiplets (e.g., compute and memory) face-to-face.
- System Folding: Makes thousands of chips behave as one, reducing communication latency.
- Transistor Folding: The next frontier, aiming to fold the transistor itself.
AI Datacenter Deployment:
- Unified Bus (UB): Eliminates multi-layer protocols, slashing remote-access latency from microseconds to ~100 nanoseconds (about a 500-fold reduction).
- Hi-ONE Optical Interconnect: 8 Tb/s per module, replaces copper with fiber for extended reach and reduced energy consumption in data movement.
Challenges: Existing Electronic Design Automation (EDA) tools are poorly equipped for this new architecture, and vertical stacking creates significant thermal management problems, especially as the technology scales for data centers. Peking University has developed a prototype "true-3D" EDA tool for LogicFolding.

🔮 前景展望基於引用來源的 AI 分析

Huawei will significantly reduce its reliance on advanced Western lithography equipment for high-performance chips.

The Tao Law and LogicFolding provide an alternative path to performance gains by optimizing time and 3D design, circumventing the need for cutting-edge EUV machines.

The semiconductor industry will increasingly shift its focus from purely geometric scaling to holistic system-level optimization, including time-domain and 3D integration.

Huawei's Tao Law highlights the diminishing returns of Moore's Law and proposes a comprehensive framework for performance improvement across the entire computing stack, which other industry players are also exploring in various forms of advanced packaging and 3D integration.

Huawei's Ascend AI processors will achieve competitive performance against global leaders like Nvidia by 2030-2031.

The LogicFolding architecture is planned for application in Ascend AI processors by 2030, with a target of 1.4nm-class equivalent transistor density by 2031, directly challenging current market leaders.

⏳ 時間線

1987

Ren Zhengfei establishes Huawei in Shenzhen.

2019-2020

US government imposes trade restrictions and TSMC ceases foundry services, limiting Huawei's access to advanced chip technologies.

2020-2026

Huawei conducts six years of R&D on the Tao Law and LogicFolding, mass-producing 381 chips based on these principles.

2025-2026

Hybrid bonding technology from Chinese packaging companies gradually matures, supporting advanced integration.

2026-05-25

Huawei officially proposes the Tao (τ) Scaling Law and LogicFolding architecture at the IEEE International Symposium on Circuits and Systems (ISCAS 2026).

2026-Q3/Q4

Huawei plans to launch its Kirin smartphone chips featuring the LogicFolding architecture.