华为τ定律：中国半导体的又一个“深度求索”时刻

China Semis: Huawei's Tau(τ)Law, another DeepSeek moment Huawei unveiled the"Tau(τ) Qingyuan Lin, Ph.D.+852 2123 2654qingyuan.lin@bernsteinsg.com a shift from geometric scaling (making transistors smaller) to temporal scaling (reducingsignal latency) as a successor to Moore's Law. A-share semis names rallied yesterday(STAR 50 +5.9%). We believe the Tau law demonstrates that China is making impressivetech breakthrough despite export controls, and sets a clear roadmap for continuousimprovement in chip performance without EUV, making it another DeepSeek moment for Francis Ma+852 2123 2626francis.ma@bernsteinsg.com Zheng Cui+852 2123 2694zheng.cui@bernsteinsg.com Tau Law demonstrates breakthrough beyond what’s achieved.The core idea of theTau law is systematically reducing the time constantτby cutting signal latency acrossentire semis stacks. One of Huawei’s tech innovation called ‘LogicFolding’ looks similar to TSMC’s SoIC at first glance, as they both stack two dies on top of each other. But Huaweiis able to achieve impressive sub-2um bonding pitch (Exhibit 2), which allow them to stacklogic on top of logic to reduce latency (cell to cell) rather than just stack up at chip level (dieto die, usually SRAM on logic). This is why they are able to improve not only on density butalso on frequency (Exhibit 1). On top of that, Huawei emphasizes the importance of end- to-end co-optimization for AI, where the improvement across the transistor-circuit-chip-system levels are all leveraged. They target to achieve 125X compute power improvement However, Tau Law comes with constraints and won’t allow China to close the gapwith global leaders yet.First, Tau Law assumes sustained progress in 3DIC advancedpackaging, areas where global leaders like TSMC still hold a meaningful technologyand ecosystem lead. Second, stacking multiple dies boosts transistor density but alsoincreases power density, thermal constraints will be a key issue that require power-delivery innovations. Thirdly, the yield and cost will also be another barrier for adoption ifnot engineered properly. Arguably the innovations from Huawei are the ones that global The upside for China semis is massive if the Tau Law is executed successfully.Justlike DeepSeek innovation on algorithm encouraged China to invest in AI and build a wholeAI-stack locally, Huawei’s Tau Law now will give the industry a lot more confidence to investin China semis and build a whole Semis-stack locally, further enhancing our localizationthesis for the sector. We likeSMIC, NAURA, Piotechthe most under this theme, as SMIC BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS We rate SMIC, Hua Hong, NAURA, AMEC, Piotech, Cambricon, Hygon Outperform benefiting from this new roadmapfor China Semis. On beneficiaries, we see the Tau Law roadmap as structurally positive for China’s leading foundry, advanced-packaging supplychain, and AI chip ecosystem. On foundry,SMICis a clear strategic partner and likely a critical enabler if Huawei is to deliverthe roadmap and eventually reach TSMC A14-equivalent logic density via multi-die integration without EUV, implying sustaineddemand for SMIC’s most advanced DUV based multi-patterning advanced logic nodes.Hua Honglikely will also benefit asthe expected acquisition of Huali who is potentially also working on advanced logic packaging. In semicap, we see upside skewed to domestic tool vendors who has higher exposure for advanced logic and packaging:NAURAshould benefit morethanAMECas it has much higher exposure in advanced logic, whilePiotechhas been working on bonding tools for advanced packaging and thus market will perceive it as one of the core beneficiaries. For AI fabless, Huawei’s framework provides aclearer path to sustain performance scaling under EUV constraints, which should be supportive for domestic accelerator and DETAILS HUAWEI’S ROADMAP ON CIRCUITS AND SYSTEMS In the event, Huawei released its roadmap for chip improvement under the guidance of Tau Scaling Law. Three key highlights: 1.Fast transistor density improvement.Trough advanced packaging (or so-called ‘LogicFolding’ by Huawei), they are ableto improve the transistor density from 155 MTr/mm^2 in 2025 (HiSilicon Kirin 9030, based on SMIC N+3 node, betweenTSMC N5 and N7) to 238 MTr/mm^2 in 2026 (equivalent to TSMC N3 node density. Likely that’s based on two chips stackon top of each other, for example one SMIC N+3 node chip with another SMIC N+2 node chip, so no longer apple to apple 2.Steadily increasing max frequency.The speed of a transistor is generally measured by how fast it can switch, or thefrequency, further improvement on the max frequency is achieved through the tau optimization that reduced the latency, 3.Exponentially growing AI cluster scaling.On top of improvement on chip performance, the Tau Scaling Law alsoemphasizes on the improvement of networking between chips to build exponentially larger AI superPoDs, we haveillustrated that in the Huawei AI chip roadma