|
Imagine you want to know whether a given statement is true or false, and you have a friend you can ask. He can only answer once, yes or no. You ask, and he gives you your answer. But let’s say you have a more complicated question, that isn’t just yes or no. It has a follow-up component: “Is X true? And if not, then does Y matter?” Your friend can only answer one question, yes or no. So you need a second friend like that first guy, to answer the follow-up part.
You: “Is X true?”
Friend 1: “No.”
You: “Then does Y matter?”
Friend 1: “...”
Friend 2: “No.”
This is a very stupid way of drawing an analogy to how the very basics of chip computing work, but if you want smarter analogies you can just ask ChatGPT.
Chips are made of transistors, and each transistor is basically like one of those guys – it can provide a binary answer, so to speak, to a question. It can act as a tiny yes-or-no gate. But instead of answering a question, it is controlling the flow of electricity by deciding whether it can pass through the gate.
To enable the completion of complicated tasks, huge numbers of transistors are arrayed to form circuits, which together can process logical operations quickly, with those operations adding up to computation on the larger scale.
Naturally, if you’re making a chip for computation, you want to get as many transistors onboard as possible.
In like the 1960s, when integrated circuits first emerged, engineers were jamming transistors into chips at the micrometer level – so one transistor might have been 10 micrometers, or “microns”, in size. The goal was always to get them smaller so that more could be packed onto a chip, and by the late 1990s, engineers had reached the nanometer level. A nanometer is 1,000 times smaller than a micrometer. In the early 2000s, transistors were starting to get down to less than 100 nanometers in size.
In recent years, we’ve reached single-digit nanometer scales, with manufacturing processes reaching about 3-5 nanometers. At these scales, some parts of a given transistor are just a few atoms wide. There is, effectively, just one company in the world today that has the technology that can engineer these components – ASML, in the Netherlands, with its extreme ultraviolet light lithography machine – but that’s getting a bit off-track for our purposes today.
That’s because we’re talking about IBM’s big announcement that it has developed sub-nanometer chip technology, reaching as small as 0.7 nanometers, or 7 angstroms. This enables designs that would cram almost 100 billion transistors onto a chip the size of a human fingernail.
In terms of how they did this, the breakthrough here is IBM’s concept of the “nanostack” – a 3D approach that involves stacking transistors vertically, instead of just arranging them side-by-side as in a conventional chip.
This has some important advantages in terms of computation. IBM says it could provide either a 50 percent improvement in performance or a 70 percent improvement in energy efficiency compared to the current state-of-the-art 2-nanometer tech.
Now, it’s important to pause here and consider that IBM is just saying it has found a way to do this, not that it’s doing it at scale. In its announcement, the company said it “sees a path to production in as early as the next 5 years.” So this amazing breakthrough may actually be several years away.
Still, its implications are considerable. One of the biggest bottlenecks to AI scaling right now is compute infrastructure – huge data centers packed with chips that are using a ton of energy. That is expected to remain a key issue for several years, and IBM’s nanostack tech could help to resolve it.
It’s also important geopolitically. The US and China are considered the frontrunners in the AI arms race, and both are dependent on chipmakers like TSMC, in Taiwan, as well as SK Hynix, in South Korea. IBM isn’t really a major chipmaker, and this new tech doesn’t change that, but it does now have a groundbreaking design. And it’s an American company – one that has signed a Letter of Intent with the US Department of Commerce promising to provide equity stake in one of its quantum computing ventures in exchange for $1 billion in planned funding. That’s not directly related to chips, but it shows how IBM is becoming entwined with the US government itself.
That relationship could deepen in the coming years as the national security implications of advanced AI become even more glaringly obvious. Many actors are going to want to control access to the tiny armies of yes-or-no guys.
|
