Huawei Found a Loophole in Moore's Law

追加: 2026-06-04

[00:00:00]Moore's law, the phenomenon that shrunk computers down from buildings to smartphones, is quietly starting to break. And the timing couldn't be worse because AI is pushing energy demand to all new heights. But there's some good news because instead of shrinking transistors, Huawei is rewriting the rules on the scaling of time. I flew out to Shanghai to attend an ILE E symposium to partner with Huawei to learn how logic folding and time scaling work. I learned so much that my brain is still buzzing. And in this video, I'm going to break it all down and show you why I think this could be an absolute gamecher. I'm Ricky and this is Tubbit Da Vinci.

[00:00:37]What Huawei unveiled the Tao scaling law is a brand new way of thinking when it comes to Moore's law. But to see why it's such a big deal, you have to understand the wall the chip industry is starting to slam into. For 60 years, computers got better through one simple trick. Shrink the transistor. the microscopic on-off switches that does all the thinking and pack in more of them. That's Moore's law and it took us from roomsiz computers to supercomputers in your pocket. But Mo's law had a quiet partner called Dennard scaling that kept those smaller transistors cool and fast on the same heat budget. Around 2005, voltage stopped scaling because leakage became a problem and clock speeds plateaued around 5 GHz. That's why your computer started advertising cores instead of gigahertz. And now even the shrinking is hitting walls of its own.

[00:01:22]The smallest features on a chip are just a handful of atoms across. Go any smaller and the electrons literally leak straight through. 2 a chip's tiniest features. The industry uses light and the smaller the feature, the shorter the wavelength you need. Visible light is around 400 to 700 nm. The deep ultraviolet chip makers leaned on for years was 193. To go smaller still, they had to invent extreme ultraviolet EUV.

[00:01:48]At just 13.5 nanometers, the way they make this is, to use a technical term, bananas, a laser zaps droplets of molten tin 50,000 times a second, vaporizing them into a tiny plasma that glows at exactly that wavelength, then bounces the light off the flattest mirrors humanity has ever built in a vacuum.

[00:02:08]Because EUV is absorbed by glass, by air, by basically anything. The machines that pull this off are arguably the most complex tools humans have ever built.

[00:02:18]And Huawei didn't have access to the latest ones. So, they were forced to ask a completely different question. And their answer kind of changes the entire way we approach computing. Instead of chasing a smaller size, they chased a shorter time. Here's a secret most people never learn. The transistors on a chip are blazingly fast. The slow part is the wires between them. Every signal has to travel down a tiny metal wire and the time that takes depends on how far it has to go. Engineers boil this down to a single number, the time constant, written as the Greek letter toao. There is a tiny equation behind it. Tao equals R * C. The wire's resistance times its capacitance. And both grow with the wire's length. So cutting the distance in half makes the delays four times shorter. So stop optimizing for size and start optimizing for travel time. And the moment you frame it that way, a wild idea emerges. If the problem is distance, the fix is obvious. Make signals travel less far. In a normal chip, all the logic is spread out flat on one layer. So, a signal can have to sprint clear across the silicon.

[00:03:21]Huawei's technique, logic folding, does exactly what it sounds like. It takes the long flat path and folds it into a vertical stack, building the chips across two layers instead of one. A signal that used to run clear across the chip now takes a tiny hop straight up like a sprawling office getting a second story added. So people who work together sit right above each other, one elevator ride apart instead of a long walk across the building. But pulling this off is straight up impressive. You have to fuse two chips together so precisely that individual wire lines up across the seam finer than a 40th of the width of a human hair. And here's the geometric leap that makes folding so much more powerful than the old approach. In a 2.5D layout, two chips sat side by side on a connecting board and talk to each other along their edges. Double the chip's width and you only double the connections between them. Fold them face to face instead. And now every spot across the entire die surface can be a connection. Double the width here and you get four times the connections. They scale with the whole area, not just the edge. For a 20 mm chip, Huawei's number jumps from about 12,000 connections to a million between layers, roughly 100 times more bandwidth, just from going from a single layer to two. On Huawei's next Kieran phone chip launching this fall, Logic Folding crammed in 53% more transistors, all without a smaller, fancier manufacturing process. For me in particular that obsesses about energy, that 41% more energy efficient number is really what caught my eye. So, how do they actually make this then? Two completely separate wafers, each designed in parallel to hold half the chip's logic. Each one gets fabricated the normal way. Transistors built into the silicon. Then layer after layer of copper wire laid down on top. Then one wafer gets ground down to the thickness of a sheet of paper and tiny vertical tunnels called TSVs or through silicon vas are drilled straight through. Then comes the wild part. The two wafers are pressed face to face with submicron precision and the copper pads on one wafer fuse directly to the copper pads on the other. No solder, no microbumps, and no gaps. The two halves become one continuous 3D chip. With millions of those tiny TSVs running straight through the seam like elevator shafts, phones are just a warm-up. The real prize is AI, where thousands of chips work as one. And here the energy math gets staggering. In these giant AI systems, more than 80% of the energy isn't spent computing at all. It's spent just moving data around. So Huawei applied the same shorten the distance idea at scale. A new interconnect called unified bus lets thousands of chips behave like a single giant chip. And because copper wastes enormous energy over distance, they move data between racks using light beamed through a fiber at 8 trillion bits a second. The whole data center behaves less like a warehouse of separate computers and more like one enormous efficient brain. Every move is an energy play. Shorten the distance and cut the wattage. Now, going vertical this way isn't a completely new concept. Your gaming PC might already have a version of it. AMD's 3D Vcash stacks an extra slab of cache memory on top of a finished processor. That's called stacking, and it works great for adding more of the same thing. But Huawei isn't stacking, they're folding. stacking bolts a separate chip on top of another with a coarse interface between them.

[00:06:40]Folding takes one circuit and splits its own inner workings across the layers wire by wire so that the actual logic of the chip lives in 3D. The real key is something called EDA which is electronic design automation. It's the software toolkit that they use to take the design of their chip that they want to build and actually figure it out. Think about it like this. If you were trying to build a house and I told you, okay, you got to plum up the plumbing, sewer, water for a three-bedroom house, you could do that. Now, try a billion-bedroom house. That's what a chip is there. It's beyond the scope of a human mind to even comprehend. So, there was no tooling for this. They had to build their own. That's part of what Huawei has been working on for 6 years.

[00:07:19]And these aren't separate tricks. Huawei has a formal map for a whole approach.

[00:07:23]They call it the four foldings. Circuit folding is the one we walk through, shortening wires inside a chip. Chip folding bonds different chiplets say compute and memory face to face. System folding takes thousands of chips and make them behave as one. And the next frontier folding the transistor itself going vertical at the device level is already on the horizon. Each fold is a new dimension to grow into. That's why it's called a law like the time scaling law that they're approaching not just a individual breakthrough. None of this is easy. The TSVs themselves are deceptively hard to make. Skinny enough not to steal precious chip area, right?

[00:07:57]Because if you have to drill tunnels for the wire communications between two layers, that's air you're taking away from transistors. So you got to make them tiny and deep enough to travel all the way between the two layers. That is a huge engineering problem. It's like digging a well at the atomic scale. If you're anything like me, you have to be thinking about the heat problem a little bit here because my first thought is, you know, CPUs are hot already. How can we possibly now fold them and possibly, you know, hope to cool them? The first is shortening the wires. Right? If you have these TSVs where you're tunneling information and shorting the routes, you make less heat. It turns out a lot of the heat comes from not the transistors flipping on and off, but from the communications, the wires, the voltage that's passed through all this electrical cabling that gets shortened.

[00:08:43]So, you have less heat to begin with.

[00:08:45]The second thing that they do is they optimize the chip design. If you ever look at a CPU under like a heat thermal imaging camera, you'll see hot spots.

[00:08:53]the chip gets hot in certain areas more than others and that's your limiting factor. The heat sink is going to get saturated right around that point. So one of the challenges then is how to lay out in the design the EDA software to put things that get hot separately to make it more uniformly distributed.

[00:09:10]There's a lot of overhead that way. And there's also some proprietary stuff I think that handles heat dissipation within the chip that sound off in the comments and subscribe. Let me know if you want me to dig in deeper because I'd have to ask their engineers and really nerd out and I've already been nerding out to get to this point. The second question you're probably wondering is if you can fold a chip into two layers, can you fold it into three or four or 10? I had the exact same question. So I asked and the answer was beautiful. In the future, climbing into more layers can become the new shrinking the transistor, right? Instead of going from 14 nm to 7 to 5 and shrinking the nanometer scale which is the geometric scaling, maybe the layering becomes the new tow scaling. Just like the industry spent 60 years dropping the node from 10,000 nanometers down to two. The next era could be spent climbing layers two then four and maybe more which affords us compute density in a whole new way.

[00:10:02]Every new layer is harder. The TSVs get nastier and the tolerances stack up.

[00:10:07]Right? If you have to get a line perfectly between two layers and now you add a third layer, it just gets really, really difficult. So, it won't be easy.

[00:10:15]But Huawei isn't waiting. Their next server CPU, the Comp 960, already targets a three die stack running at 4 GHz with 42 metal layers and a 54% speed jump over any other current chip just by folding more. Some of the most surprising leaps in technology come not from the people with the most resources, but from people who are forced to do more with less. Here's some examples.

[00:10:38]Take DeepSeek. Working with limited access to the most powerful AI chips, they got ruthlessly efficient, pushing techniques like quantization, running their AI on lower precision math so far that they trained a world-class model for a tiny fraction of the usual cost.

[00:10:54]It completely shook the industry. But here's the beautiful thing about efficiency breakthroughs. They don't stay locked up. The moment one is proven, everyone adopts it. So, even labs with the best Nvidia GPU hardware, the whole field gets cheaper and greener overnight. Look at this, cuz I've been thinking about summer vacation coming up here. This is the Light Ship L1. Towing a trailer behind an EV can cut your range almost in half. So, most people say you can't tow with EVs. Well, Light Ship did something that nobody else had thought to do. They put motors in a huge battery on the trailer itself with solar panels on the roof to keep it charged.

[00:11:28]And the trailer shrinks down vertically to minimize the frontal area when traveling and then pops back up when you park. A constraint breeding a surprising answer that's better than what came before it. This is the whole idea of innovation from constraint. And here's the image that Huawei left us with. Fold a single sheet of paper in half. Then again, then again. By the 42nd fold, that one sheet is thick enough to reach the moon. That's the math of compounding leaps. And that's what they mean by folding, right? Circuit, chip, system, each folds a new dimension. So, Huawei isn't shrinking anymore. They weren't able to. So, they're climbing. This is going to be huge. Now, you might be wondering, why do I care so much about this? It's because AI is a lot more than silly cat videos on YouTube. There is some seriously impressive benefits that AI can bring to the whole entire world.

[00:12:18]In my trip to China, I saw intersections that were powered by AI where the lights turning red and green is handled by what is actually happening on the road. That would completely remove that feeling of being at a red light at night when there's no one else around and the light just won't change for you. That could remove congestion, idling times and improve on traffic. At CES, I saw a smart toilet that monitors your pee to see how healthy you are to try to catch things. That is AI diagnostic data for health and it could completely revolutionize things. If you could test your health parameters on a daily basis because it was so cheap, you could catch things so early and give yourself a way better chance of survival. And you got to remember, we live in an era where AI computing and data center growth is about to absolutely explode. And a lot of big companies that had climate pledges to be carbon negative at 2030 have thrown those out because now there's a new frontier, a gold rush, and they're going to go after AI with everything they got. So, in this world where AI can be so helpful, and we're going to be building it out, we got to talk about it. And we have to figure out how to increase the performance per watt, how we can do more with a unit of electricity because those are the breakthroughs that are going to be hugely impactful at scale. But what do you think? How big of a deal do you think Logic Folding is? And how impressive is this Huawei announcement?

[00:13:37]I was genuinely impressed and blown away. Now, sound off in the comments if you want us to reach out to see if we could do like a lab tour or get in and see the machines or do something more hands-on. I would love to do that. So, sound off in the comments below. And until next week, I'm Ricky and we'll catch you on the next