Huawei Releases First "Tao Chip" Kirin 2026 Benchmark Data
N.R. Finch
Huawei has published the first production-chip test data for the Kirin 2026 — transistor density up 55% and power down 41% without changing the process node. This means Tao's Law has moved from theory paper to fab-verified silicon, marking a concrete step in China's post-Moore path.
What exactly did the Kirin 2026 achieve?
He Tingbo, Huawei director and head of the semiconductor unit, updated the V2 paper *"Temporal Scaling Theory for Multi-Level Electronic Systems"* on ChinaXiv. It is the first time production chip data has been disclosed.
The headline: using LogicFolding — stacking circuits vertically instead of spreading them flat — transistor density rose from 155 MTr/mm² to 238 MTr/mm², a 55% gain on the same process node. This means → a density jump that previously required three years of geometric shrinking (etching finer lines) was achieved by changing how layers are stacked, not by changing the lithography.
Other metrics improved in lockstep: power down 41% at equal performance; clock speed up to 3.1 GHz (+13%); SRAM frequency up over 40%; clock buffers cut by more than 50%; wire length shortened roughly 30%.
In plain terms = no new lithography machine, no new process node — just "fold the circuits upward." That is the core of Huawei's strategy for working around EUV restrictions.
What does the Ascend roadmap signal?
The paper also lays out an AI-chip timeline: around 2030, the Ascend 990 will bring LogicFolding into AI accelerators; by 2035, hardware integration density is projected to rise over 100×.
Near-term cadence is already locked in: the Ascend 950PR has shipped; the Atlas 950 super-node, built on LingQu interconnect and the Ascend 950DT chip, is slated for Q4 2026. This reflects Huawei maintaining a "one generation per year, double the compute" pace on AI silicon.
The paper notes that in large AI clusters, over 80% of energy goes to moving data and over 70% of cost goes to storing it. Huawei's proposed fix: a unified bus architecture, near-package optical I/O — placing optical interfaces right next to the chip package — and 3D folding.
In plain terms = the AI-chip bottleneck is not just raw compute; it is the energy and cost of shuttling data around. Huawei's answer is to shorten and stack the data paths.
From phones to cars — how wide will Tao's Law spread?
The Kirin 2026 is the first mass-produced Tao's Law chip, expanding from a single active layer to two.
Huawei plans to replicate LogicFolding across automotive chips, telecom base-station chips, and industrial-control chips. This means → Tao's Law is not one chip's story; it is a cross-scenario design paradigm.
He Tingbo projects that from 2026 to 2035, transistor density will push toward 400 MTr/mm² and beyond, with each package integrating three, four, or more active layers.
Who in the supply chain benefits first?
Industry sources expect Huawei to accelerate Chinese OSAT (outsourced semiconductor assembly and test) capacity in hybrid bonding — joining two chips face-to-face — along with 2.5D/3D packaging and through-silicon vias (TSV), vertical channels drilled through a chip to connect stacked layers.
Huawei also plans to gradually open LogicFolding design specs and interface standards, pushing domestic EDA (electronic design automation — the software tools used to draw chip circuits) vendors to adapt their 3D IC design tools.
This means → OSAT firms may be entering a capacity-expansion cycle, and utilization rates at China's mature-node foundries are set to rise as well.
What is the real unresolved question?
He Tingbo writes candidly: "The technology framework for the next decade is clear, but many unsolved problems remain — no single company can crack them alone."
The collaboration areas he names: toolchains, industry standards, performance benchmarks, device physics, and business models.
This reflects the fact that whether Tao's Law truly becomes the post-Moore industrial paradigm depends on how fast the entire ecosystem aligns — the technical path exists, but the distance from one company's paper to an industry-wide manufacturing standard is still long.
Content is for reference only, not financial advice.