Three-dimensional Revolution: China threatens to upend the chip market
The Chinese Huawei Corporation made a statement at the ISCAS technology conference, which many called a revolution in the semiconductor industry. Some analysts consider this to be the breakthrough of the decade, while others consider it an ingenious marketing bluff. We are talking about the presentation of the so-called Tau Scaling Law and LogicFolding technology. In theory, this makes it possible to circumvent the restrictions and sanctions imposed against China (primarily on the supply of photolithographic equipment, without which the production of modern high-tech chips is almost impossible). See the Izvestia article about what this means for the global semiconductor race and the growing AI industry.
The Death of Moore's Law and the Architectural Maneuver
For the last half century, all the world's electronics have been living according to Moore's Law. Its essence boiled down to geometric scaling: every two years, engineers reduced the size of the transistor in order to place as many computing elements as possible on a piece of silicon. The smaller the transistor, the shorter the signal path, which means that the chip works faster, consumes less energy and costs less. The industry was reaching ever smaller distances: 14, 7, 3 nanometers.
The problem is that this path comes to a dead end. At the two-nanometer level, transistors become several tens of atoms in size. To "print" them, extreme ultraviolet lithography (EUV) installations from the Dutch ASML are required, costing over $150 million apiece. The development of one advanced chip began to cost $1 billion. The "Moore's Law" was actually abolished — the chips continued to become more powerful, but they stopped getting cheaper.
Washington blocked China's access to these ultra-expensive EUV machines in the early 2020s. Left with last-generation equipment, Huawei has hit a technological ceiling.
Now, engineers in Shenzhen have proposed shifting the focus from distance (geometry) to time (signal delay). In physics, there is a concept of a time constant — Tau. Roughly speaking, this is the time it takes for the signal to get from point A to point B inside the processor. Huawei suggests abandoning attempts to make the transistor even smaller, and instead redesign the chip architecture itself so that the signal goes faster.
The practical embodiment of this idea was the LogicFolding technology. It can be compared to a giant one-story factory. A worker needs to run hundreds of meters to transfer a part from one workshop to another. It's long and energy-consuming. If we can't supply more workers, we can rebuild the plant — make it two-storied, and cut holes in the floor and ceiling between the workshops so that the parts can be transferred vertically.
Huawei did something like this at the silicon level. In a traditional processor, the circuit elements lie in the same plane. The signal runs along long paths, losing energy. Huawei engineers cut these circuits and stacked them like a sandwich, vertically connecting millions of microscopic contacts using hybrid splicing technology.
The results demonstrated on the new Kirin 2026 processor look impressive. Without using new lithographic machines, Huawei has increased the transistor density by 55% (to 238 million per square millimeter), while remaining on old processors. Energy efficiency increased by 41%, and the clock frequency increased by 13%. The company claims that by 2031, this approach will make it possible to create chips equivalent to Western processors of the 1.4 nanometer class.
Breakthrough or hype
The reaction of engineers and industry analysts was mixed. From the point of view of the theory of electronic circuits, this is an absolute breakthrough. Huawei has solved the most difficult mathematical problem of routing signals in three-dimensional space at the submicron level.
From the point of view of mass commercial production, there are skeptics. LogicFolding technology has two main difficulties in implementation.
The first is warmth. A "two-story factory" means that the heat generated by transistors is concentrated inside a silicon sandwich. The inner layers have nowhere to give off heat. If the chip overheats, the system will automatically reset its frequency, and all speed gained through shortcuts will disappear. Without a revolution in materials and cooling systems, the announced figures will remain only a laboratory record. The second problem is the percentage of marriage. Combining two layers of silicon so that millions of contacts match up to fractions of a micron is a task of enormous complexity. Any speck of dust turns the chip into garbage. The cost of such a 3D assembly may be higher than printing a traditional chip on new EUV machines.
The Virtue of necessity
The question naturally arises: why did Huawei come up with this idea, and not Intel, Samsung or TSMC? In short: geopolitics. American and Taiwanese corporations had no incentive to puzzle over the most complex 3D modeling of circuits. They had legal access to the ASML machines. They followed the path of least resistance, buying new machines and reducing nanometers.
However, attempts to find an asymmetric response to the ASML monopoly and the limitations of EUV lithography have been made before. The most notable step was the technology of nanoimprint lithography, introduced by the Japanese corporation Canon in the fall of 2023. Unlike traditional installations that use a system of extremely complex lenses and extreme ultraviolet light to "burn out" the circuit on silicon, the Japanese method works on the principle of a mechanical stamp. A special liquid resin is applied to the plate, into which a 3D template with a relief of the future microcircuit is physically pressed. The abandonment of high—power lasers has made it possible to radically reduce the cost of equipment - Canon machines are several times cheaper than ASML devices and consume 90% less electricity. The developers claimed that the technology is capable of stamping chips according to the standards of five nanometers with the prospect of scaling up to two nanometers.
However, the revolution did not happen because of the physics of the microcosm and the strict requirements of mass production. When a stamp and a silicon wafer come into physical contact, any microscopic speck of dust that gets between them not only rejects a particular chip, but also damages the template itself, making all subsequent prints defective. The second problem was the combination of layers. A modern processor consists of dozens of levels, and mechanical indentation of the die microscopically deforms the plate. Achieving an ideal overlay at the level of 2-5 nanometers with this approach turned out to be a task of enormous complexity. That is why NIL technology has so far been used mainly for the production of memory chips (3D NAND), whose regular architecture forgives small errors that are unacceptable in the complex chaotic logic of central processors.
China has been cornered by American sanctions. The US export control was designed to block access to EUV lithography, and Washington would freeze Beijing's technological development at seven nanometers forever. Unable to physically expand, Chinese engineers were forced to move deeper. The sanctions have worked as a powerful catalyst. With their own hands, the United States forced China to finance and create an alternative, post-Soviet branch of microelectronics development.
The Future of AI and the threat of Nvidia's Monopoly
The main threat to American dominance does not lie in smartphones. As SemiAnalysis analyst Dylan Patel correctly points out, Tau technology can have an impact on the artificial intelligence market.
Modern AI is trained on more than one processor. Large models require giant clusters consisting of tens of thousands of accelerators interconnected. Nvidia is currently the monopolist in this market with its chips and the NVLink superfast communication ecosystem. Huawei introduced the UnifiedBus bus and Hi-ONE optical interfaces as part of the "Tau Law". Their goal is to reduce the latency of data transfer between different chips in a cluster by 500 times.
If this architecture proves its effectiveness in the series, Huawei will accomplish the impossible. The company will no longer need to create a single super chip capable of surpassing the conventional Nvidia Blackwell in computing power. The Chinese will be able to take a dozen medium-sized, low-cost processors of the old generation and connect them with 3D packaging and optical interfaces so quickly that this assembly will work as one supercomputer at the output.
This could mean a major upheaval for the global industry. The artificial intelligence economy will receive an alternative hardware base. The United States will lose the ability to control the development of AI in countries that they consider questionable by issuing licenses for the export of AMD and Nvidia products. The "Tau Law" has not yet abolished the American monopoly, but it has shown that the wall Washington has built around advanced technologies may not be so impenetrable.
This concept is still far from becoming an industry standard. Huawei will have to prove that LogicFolding is amenable to mass production, UnifiedBus provides real growth in AI clusters, and Hi-ONE optical connections can be economically implemented in large data centers. The next stage will be determined by specific products, benchmarks, and customer recognition. For now, Huawei has made a move, and the global industry is watching closely. Given that the Chinese have managed to almost catch up with the Americans in software in a matter of years, it cannot be ruled out that a similar trick will be possible in hardware. In this case, the United States has every chance of becoming the second most important technological power in the world, just as it has already become the second most important industrial power.
Переведено сервисом «Яндекс Переводчик»
