Moore's Law Is Dead. Here's What Replaced It

Hinzugefügt: 2026-05-22

[00:00:00]You bought the $400 chip. The benchmark promised a bigger number than last year, and the bigger number was true. So, you waited for the part where everything felt faster, and it never quite arrived.

[00:00:09]Here's why. That number on the box stopped being a measurement of how fast your computer is. It quietly became something else. Closer to a release date than a speed. And once you see that, you understand why your next processor is going to disappoint you, too. For decades, the story was simple. Every couple of years, the transistors got smaller, the chip got faster, and you got more for the same money. That was the free lunch. You wrote the same code, you waited 2 years, and it ran faster on its own. Nobody had to do anything. That lunch is over, and the industry knows it's over, which is why the marketing quietly changed underneath you. They stopped selling you speed, and they started selling you three other things, hoping you wouldn't notice the switch.

[00:00:48]Here's what this video is going to show you. Exactly where the wall is, why more cores is a magic trick, and not a gift, and what the number on the box actually means now. Stay till the end, because the last part, the thing that replaced gigahertz on the marketing sheet, is the part almost nobody explains honestly.

[00:01:07]Let's start with that number. The 3 nanometers printed on your newest processor doesn't measure anything physical inside it. There's no part of that chip that's 3 nanometers wide.

[00:01:16]There hasn't been for one of those numbers since the late 1990s. For the first few decades of chip making, the node name actually corresponded to something. It roughly tracked the physical length of the transistor gate, the little switch that turns the flow of electrons on and off. Smaller gate, smaller transistor, more of them per chip. The number meant something you could point to. That stopped being true around the late '90s. Somewhere near the.18 and.13 micron generations, the geometry of the transistor changed shape so much that the node name detached from any single physical dimension. The gate stopped shrinking at the same rate as the label. So, when a foundry says 3 n or 2 n today, they're not telling you the size of anything. It's a generational name. This generation is denser and a bit more efficient than the last one. That's the honest translation.

[00:02:00]2 nm means newer than 3 nm.

[00:02:03]It does not mean 2 nm of anything. And here's why this matters to you, not just to a fab in Taiwan. When the number on the box stopped describing physics and started describing a product cycle, the industry got permission to keep the number climbing while the actual gains underneath flattened out. The label keeps going up. The thing it used to measure stopped going up a long time ago. But maybe you don't care about the label. You care whether it's faster. So, let's ask the real question. Why can't they just keep shrinking it? Because at this scale, they're no longer fighting engineering. They're fighting the atom.

[00:02:32]A transistor works by using a barrier to stop electrons from flowing until you want them to. Make that barrier thin enough, and we're talking about a wall that's now only a handful of silicon atoms thick, and the electrons stop respecting it. They just appear on the other side. They don't go over the wall.

[00:02:48]They don't go through a hole in it. They tunnel. This is quantum tunneling, and it isn't a defect you can engineer away.

[00:02:54]It's how reality works at that scale.

[00:02:56]And it gets worse because tunneling isn't binary. As the barrier gets thinner, the leakage doesn't tick up politely. It climbs exponentially. The transistor that's supposed to be off is now quietly leaking current and burning power doing nothing. So, you push past the wall, and your reward is a chip that runs hotter and wastes more energy to do the same work.

[00:03:16]That's the physical floor. You can be clever about the shape of the transistor. You can stack things vertically. You can use new materials.

[00:03:23]And the industry is doing all of it. But none of it repeals quantum mechanics.

[00:03:28]The era of just make it smaller and it gets free speed is physically over. Not slowing down. Over. So, if they can't make a single transistor switch much faster, what have they been selling you for the last 20 years? Let me show you something that should bother you. Look at the clock speed of a high-end desktop processor in 2007. Now, look at one today. It's roughly the same, somewhere in the 3 to 5 GHz range. 20 years.

[00:03:51]Almost no movement on the one number that used to define the whole industry.

[00:03:54]That flattening has a name. It's the breakdown of Dennard scaling around 2005. Dennard scaling was the quiet promise underneath Moore's law. The idea that as transistors shrank, you could keep cranking the frequency without melting the chip. When that promise broke, the frequency stopped climbing.

[00:04:11]They'd hit a power and heat ceiling on a single core, and they've basically been parked under it ever since. So, they pivoted. They couldn't sell you a faster lane, so they sold you more lanes. Two cores, four, eight, 16, 24. Here's the analogy that makes it honest. A single core getting faster is like raising the speed limit on the highway. Every car gets where it's going sooner. Adding cores is like adding more lanes while the speed limit stays exactly the same.

[00:04:36]More lanes is genuinely useful if you have a lot of cars that can all drive independently, but not one car arrives a single second sooner. And most of what you actually do is one car. Opening an app, loading a webpage, the lag you feel when you click something. A huge amount of everyday computing is a sequence of steps where each step depends on the one before it. You can't put step two in another lane because step two needs the answer from step one. Remember those extra lanes? You built them, but the bottleneck just moved. It's not the road anymore. It's the order the steps have to happen in. And there's a law for exactly this. Amdahl's law. And it's brutal. It says the part of your program that has to run in sequence becomes a hard floor on how much you can ever speed it up, no matter how many cores you throw at it. If 10% of a job has to be done in order, then even with infinite cores, you can never make that job more than 10 times faster. The sequential part is the boss, and it doesn't care how many workers you hired.

[00:05:33]That's the core trick. They've been selling you workers for a job that most of the time only one worker can do. So, if frequency is flat and extra cores only help some of the time, what's the new headline number? What's printed in the big font on the box now. TOPS, trillions of operations per second, the AI number, the NPU, the neural processing unit, and how many TOPS it can do. And one thing the box won't put next to it, those operations are counted in INT8, one of the lowest precision formats they can use, not the kind of math your actual model is usually running. TOPS is the new gigahertz. It's the number that's there to make the chart go up and to the right, so the product feels like an upgrade. And just like gigahertz, it's technically real and deeply misleading on its own. So, the benchmark number is broken, but that's just marketing. The real wall is physical. Here's the thing the spec sheet won't tell you. For most modern workloads, the bottleneck isn't how fast the chip can compute, it's how fast you can feed it. Moving data from memory into the processor costs time, and this is the part that matters now, it costs energy, often far more energy than the computation itself. You can have a processor capable of trillions of operations per second sitting there idle, starved, waiting for the data to arrive. That gap between how fast the chip can think and how fast you can feed it is called the memory wall, and it's the real fight of this decade. So, when does all this silicon actually pay off?

[00:06:55]When the work is genuinely independent.

[00:06:57]When you really do have a thousand cars that don't need to talk to each other, rendering a video frame by frame, compiling a large code base, serving thousands of requests on a server, training a model. That's where 24 cores earn their keep, and that work is exploding, which is exactly why server chips look so different from the one in your laptop. I'll admit something. For years I bought the new chip, saw a benchmark go up five or 10%, and told myself I felt the difference. I didn't.

[00:07:24]I was buying the number on the box because the number went up. The honest measure of progress stopped being the chip a while ago, and I kept paying attention to the wrong thing. Because here's what the end of Moore's Law actually means. Speed is no longer something the hardware hands you every two years for free. From here, real progress comes from somewhere less glamorous and far more honest. Software that's actually written well. Code that respects the memory wall instead of pretending it isn't there. We've watched people do extraordinary things inside almost no space at all. And that wasn't a faster chip. That was a sharper mind.

[00:07:57]So, let me leave you with the question that actually matters now. What would you rather have? A 24-core processor you'll never fully use or a single app that knows how to use one core well? I know which one is harder to build and I know which one the box will never advertise. And the next generation is already on its way with a new number on the box climbing the same chart in the same direction. You're going to be standing in front of that box soon, $400 in hand, trying to decide if the bigger number is real this time. There was one company that genuinely beat all of this for about 2 years. Not by breaking physics, but by quietly rerouting the memory wall instead of running into it.

[00:08:34]The hardware actually felt like magic and that magic had an expiration date built in. If you want to read those numbers and understand the one temporal trick that bought them 2 years before you spend the money, that's the next video.