Chapter 2: The "brains" diversified ― CPU, GPU, TPU, quantum, bio

In Chapter 1, we learned that a chip is "a small board where a swarm of switches computes by rearranging 1s and 0s." But — say "chip," and in fact there are several kinds of "brain," each good at different things. And right now, those kinds are increasing all at once. This is the heart of the "Cambrian explosion" we mentioned in the preface. In this chapter, we survey the main "brains" that exist today, and then stretch out as far as entirely new species that work on a completely different principle (quantum, bio). It's a slightly long chapter, but this is the corner's designated stretch for "how wide the world is." Take your time.

2.1 Why "use different brains for different jobs"?

Let's start from a plain question. "If you just build one all-purpose chip and make it really fast, wouldn't that be enough?" A reasonable thought. In fact, for a long time the world went that way. But once you need to do a certain job in a staggering quantity, the story changes. Say you want to do "the same simple calculation, millions of times, all at once." Rather than having the all-rounder do them one by one in turn, rebuilding a body "dedicated" to that task is dozens of times faster — and more energy-efficient.

Let's use cooking as an analogy. An all-purpose chef is first-rate at simmering, grilling, and chopping. But asked to "julienne cabbage for ten thousand servings," a single chef is in for a very long day. Bring out a slicer dedicated to julienning, and it's done in a flash. The same thing is happening in the world of chips. Using "the all-rounder that does anything" and "the specialist honed for one task" according to the job — this is the mainstream way of thinking now. Let me introduce them in turn.

2.2 The all-rounder ― the CPU

CPU is short for "Central Processing Unit." It's the command tower of a PC or smartphone, the most familiar "brain." Its defining trait is, above all, being all-purpose. Writing text, calculating, making a judgment and branching, issuing instructions to devices … it handles every kind of job, one at a time, cleverly, and adaptively.

To put it another way, a superbly capable "single all-round craftsman." A dependable person who can meet any order, but who basically clears jobs one at a time (a few at a time, really), in sequence. Complex, intricate procedures are their forte. But "thousands of simple tasks, simultaneously" is hard for one person. That's where the GPU, next, comes in.

2.3 Massively parallel ― the GPU

GPU is short for "Graphics Processing Unit." Originally, it was a specialist for drawing the picture on the screen. A screen is a gathering of millions of dots (pixels), whose colors need to be computed all at once. So a GPU is built not as "one clever person" but as a structure that lines up thousands of merely-decent calculators and has them do the same task all at once.

To put it another way, "having 1,000 schoolchildren solve times-tables all at once." No single one is a genius, but do simple calculations all together, simultaneously, and it's far faster than one superhuman. In fact, the guts of today's AI (especially training) is an enormous repetition of multiplication and addition. This is exactly the GPU's great strength. That's why the AI boom set off a worldwide scramble for GPUs, and the company that makes GPUs shot up to one of the largest in the world. This is why people say "the AI boom = the GPU boom."

2.4 Honed for AI alone ― the TPU and NPU

The GPU was "a tool originally for drawing pictures, repurposed for AI." So then — if you built it for AI alone from the start, couldn't you make it faster with less waste? — the "brains" born of that thought are the AI-dedicated ones. The representatives are the TPU (Tensor Processing Unit) and the NPU (Neural Processing Unit). Both are designed to handle, with the utmost efficiency, only the massive multiplication and addition that are the guts of AI.

The difference is roughly "where they sit, and their scale." The TPU is the large type, lined up in data centers (giant computer warehouses), handling an extraordinary volume of AI computation in bulk. The NPU, meanwhile, is the energy-saving type, sitting small inside your phone or laptop, running AI at hand on very low power. The reason recent phones can edit photos or transcribe speech even without a network connection is thanks to this little NPU.

※ This isn't about "which is the greatest." Today's AI has a single job walk across several brains. For instance, a giant AI is trained on GPUs, served in bulk on TPUs, and finally run for your local use on your phone's NPU. "For each layer, a different brain fits." That's why you can't narrow it to one kind, and chips with different roles coexist and increase in number. This is the true nature of the diversification.

2.5 Why are the kinds increasing so much, right now?

A little background. Semiconductors long evolved down a single road: "just make them smaller and pack in more, and they get faster" (we'll cover this "packing race" in detail in Chapter 6). But in recent years, that single road has gradually stopped working as well. Merely shrinking things no longer makes them faster and more efficient the way it used to.

So a piece of wisdom kicked in. "If it's hard to speed up the whole thing uniformly, then prepare a 'dedicated body' for each frequently-used job, and make just that part overwhelmingly fast." This is the root of the current trend of "specialist chips increasing." It closely resembles how life's Cambrian explosion diversified body forms to match various environments (niches). Because a giant new "environment" called AI was born, new "species" adapted to it are appearing one after another — see it that way, and the true nature of today's bustle falls into place.

And here begins the second half of this chapter. The CPU, GPU, TPU, and NPU introduced so far, different in form though they are, were all members of the same "silicon switch" family. But the "explosion" is beginning to reach outside that frame too. Let's look at two entirely new species, different in their very principle.

2.6 [New species ①] Computing on a wholly different principle ― the quantum chip

The first new species is the quantum computer. It overturns the very premise we've relied on — "a switch's on-off = 1 or 0." In the quantum world, a very small particle can take an intermediate state that is "both 1 and 0" (superposition). Using this strange property, it's said you can do computations that explore a great many possibilities all at once.

To put it another way, trying "all the paths" of a huge maze at the same time. An ordinary computer checks the maze's paths one by one, exhaustively. A quantum machine, when it works, can evaluate many paths in superposition, all at once. So for certain hard problems — studying how a drug molecule behaves, searching for new materials, finding an optimum among vast combinations — it's expected to show power on a different order of magnitude.

Research is heating up around the world. A defining feature is that companies compete with completely different approaches: Microsoft, for example, announced a prototype chip on a new approach (topological qubits), "Majorana 1," in 2025; IBM has published a long-term roadmap that widens the scale year by year; and companies on other approaches have begun to report surpassing conventional computers on certain problems. "Which approach is the front-runner" is not yet settled.

※ But please don't misunderstand. A quantum computer is not something that replaces your PC. What it's good at is only the "certain hard problems" mentioned above; email and spreadsheets won't suddenly get faster. Moreover, today's quantum chips are error-prone, require special environments such as extreme cold, and are very difficult to handle. Practical use is still ahead. In this corner, we treat this not as "an amazing all-purpose machine has arrived," but as "a 'new species' — a brain on a different principle — has begun to grow in earnest."

2.7 [New species ②] Computing using living things ― the bio-chip

The other new species lies even further outside imagination. It's the idea of computing using living matter (the cells of a living thing) itself. Instead of silicon switches, you give signals to cultured, genuine nerve cells and use their responses as "computation" — so-called "wetware." Literally, a wet (living) computer.

It sounds far-fetched, but the research is real. One Swiss company, for example, has released a computing platform connecting on the order of a hundred thousand living, human-derived nerve cells, and claims its power consumption is orders of magnitude smaller than a conventional data center's (which makes sense, when you consider that our brains do vast information processing on about the power of a light bulb). Elsewhere, research into storing data in DNA (the molecule of a living thing's blueprint) is also advancing.

※ This is the most "distant" new species. Bio-based computing hides dreamlike potential in energy efficiency, while the road to practical use is said to be the longest. The lifespan (on the order of months) and stability of living cells, and the ethics of "using living matter for computation" — there are many walls to clear. So in this corner too, we convey this honestly, as "research that has only just sprouted," not "a practical product." Even so, the very fact that challenges in this direction are actually in motion tells you the amplitude of today's "explosion."

2.8 And the new species yet to be seen

There are still "brains" we couldn't fully introduce. The optical chip (photonics) that computes with light, in-memory computing that unifies memory and computation, neuromorphic chips that mimic the workings of the human brain … each has its strengths and weaknesses, and it wouldn't be strange for any of them to turn into the next lead. It's a Cambrian sea, precisely. We won't chase them all in this corner, but take home just this one fact: a world that had been "one kind of all-purpose silicon chip" has begun to overflow with this many varied "species."

2.9 This chapter's summary, and the bridge to the next

• An era of using different brains for different jobs. The all-purpose CPU, the massively parallel GPU, the AI-dedicated TPU / NPU. A single job walks across several brains.
• The reason the kinds increase is "specialization." As the single road of merely shrinking things stopped working as well, the direction shifted toward building a "dedicated body" for each job. A new environment called AI summoned new species.
• There are new species outside the frame, too. The different-principle quantum, and the living-matter bio. Both are still at the research stage, but they've begun to grow in earnest.

And so the kinds of "brain" exploded. But here, one question remains that we mustn't overlook. All these varied chips — "how do we run them, and give them commands?" In fact, between humans (software) and chips (hardware), an invisible "shared agreement" is exchanged. Without it, even the finest chip is just a board of sand. In the next chapter, Chapter 3, let's go and see this "instruction set (ISA)," an invisible contract — the lead that sits at the deepest place in today's upheaval.