Huawei's τ-Scaling Law: Post-Moore Semiconductor Roadmap to 2035
Huawei's τ-Scaling Law shifts semiconductor scaling from geometric shrinking to time-delay optimization across device, circuit, chip, package, and system layers, detailing Kirin logic folding, Ascend AI clusters, hybrid bonding, unified bus, and system folding with 2023-2035 targets.
Moore's Law Reaches Geometric Limits
The report reviews Moore's Law (1965), Dennard scaling (1974), and FinFET (2000), noting that transistor gate length scaling continues but design and wafer manufacturing costs have skyrocketed. The core dilemma: further geometric scaling (L) yields diminishing returns, making pure process shrink insufficient for performance gains. Huawei proposes a paradigm shift from Geometric Scaling (L) to Time Scaling (τ), using signal propagation delay τ as the core optimization metric across all hierarchy levels.
Mobile Kirin Chips: Logic Folding in Practice
Huawei's Kirin 9000s (2023) → 9020 (2024) → 9030Pro (2025) demonstrate continuous Logic Folding adoption. Results: transistor density increased from 238 MTr/mm² to 355 MTr/mm²; big-core frequency rose from 2.6 GHz to 3.1 GHz. These gains were achieved without relying solely on process shrink, using logic folding, 3D stacking, hybrid bonding (HB), and TSV.
Circuit-Level Innovations Supporting Logic Folding
SkyBridge : Top-metal data highway reducing area by >60%.
SkyClock : Top-down clock distribution with backside clock-skew tuning, delivering >5% performance uplift.
SRAM-Core co-optimization : Shortens critical-path RC delay to lower τ.
Interconnect Technology Evolution (Package Level)
Three generations defined by bump pitch:
Traditional PCB & MCM (1000–100 μm) : C4 bumps + BGA, limited bandwidth density.
Fan-out packaging (50–25 μm) : RDL redistribution, silicon bridges, enabling chiplets and HBM3/E via micro-bumps.
Hybrid Bonding (≤10 μm) : Eliminates micro-bumps; Cu–Cu direct bonding enables cache stacking (SRAM on logic) for extreme I/O density.
AI Compute: Ascend SuperPOD and System-Level τ-Scaling
Product Roadmap
Ascend 910C (2025) → 950 (2026) → 960 (2027) → 970 (2029) → 980, targeting SuperPOD superclusters.
Cluster Scale & Power
NPU count grows from dozens in 2024 to 8,192 by 2027. Single-chip power exceeds 30,000 W by 2030–2031, making liquid cooling mandatory.
UB Unified Peer-to-Peer Bus
Replaces traditional Ethernet + IB/RDMA + PCIe stacks (high protocol overhead, weak reliability, high cost). UB uses memory-semantic low-latency protocol; CPU, NPU, GPU, MEM, DPU, Switch attach directly. UB Fabric enables cross-chip load/store semantics, logically fusing multiple dies into a "single-chip system." UBoE extends reach for long-distance links.
HiONE Optical Interconnect Engine
Silicon-photonics integration: ~10 ns latency, 8 Tbps bandwidth, >100 m reach, integrated light source, suited for intra- and inter-rack SuperPOD links.
2.5D Bottleneck and System Folding
Traditional 2.5D interposer architectures suffer N² scaling of logic die performance/power/area while memory and interconnect bandwidth scale only linearly (N). System Folding vertically stacks Logic, DRAM, NAND, DTC, PMU, I/O, VRM, breaking the N² wall via vertical interconnects.
Hardware Integration Density Targets
Baseline 2026 → 4× by 2028 → 60× by 2030 → >100× by 2035.
Memory-Compute Convergence Path
1980s decoupled → current HBM tight coupling → 2030–2050 native Logic-Memory fusion.
Full-Stack Challenges and 2023–2035 Roadmap
Tooling & Methodology
New multi-die PPA evaluation, DFX testability, physical verification flows; converged PDKs supporting massive HB and TSV simulation.
Thermal Management
Cross-layer optimization covering device transient thermal shock, circuit hotspots, vertical chip thermal conduction, board-level VRM miniaturization, and data-center PUE.
Quantitative Roadmap Targets
Circuit (2031) : Transistor density 400+ MTr/mm²; big-core frequency 5.0 GHz.
AI Systems : Atlas 950 (8 EFLOPS, 2026) → Atlas 960 (60 EFLOPS, 2027) → Atlas NEXT (2030), 125× performance improvement.
Summary
Huawei's τ-Scaling Law establishes time-delay τ as the unified optimization metric from device to cluster. It delivers a concrete, cross-layer execution path spanning Kirin mobile SoCs and Ascend AI superclusters, backed by specific density, frequency, power, and integration targets through 2035.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
