The communication technology AI expects is analog itself

Millimeter-wave communication is essential in the cutting-edge AI era, but the challenge is created by digital efficiency

The world may look entirely digital, but the technical front lines are different.

We make payments with smartphones. AI engages in conversation. The cloud manages data.
Our daily lives appear to be covered in digital technology.
However, a completely different landscape unfolds at the technical front lines.

It is analog communication technology that supports the high-speed, large-capacity data processing of AI, which is considered cutting-edge.
Behind the glamorous stage of digital, analog technology is quietly but surely supporting the AI era.
The idea that the world is entirely digital is only on the surface.
The technical front lines are still operating in an analog world.

Background: What is 60GHz millimeter wave?
What supports AI's high-speed, large-capacity data processing is
analog communication technology using 60GHz millimeter waves.

How high a frequency is 60GHz?
General smartphone communication: 1–3GHz
Wi-Fi that everyone uses: 2.4–5GHz
60GHz millimeter wave: 10 to 25 times that
This frequency band has distinct characteristics.
It has extremely high directivity and does not penetrate walls. It is also absorbed by rain and oxygen in the atmosphere.

It is not suitable for general long-distance communication.
However, there are situations where those characteristics become a strength.
It can send large amounts of data over short distances all at once.
That is precisely why it is optimal for communication between AI chips and between boards.
Inside data centers, chip to chip
are communicating via 60GHz millimeter waves.
That is the reality of the AI era.

However, this design requires deep knowledge of communication technology down to the board level,
and cannot be handled by half-baked techniques.
Here lies a serious problem that no one talks about loudly.
The "reality of communication" that AI is hiding

ChatGPT returns words. Image recognition makes instant judgments.
Autonomous driving reads the road.
What we see is only the result.

But behind the scenes, an unimaginable amount of data is flying between chips and between boards at speeds close to the speed of light.
AI's computational volume continues to increase at a pace of 10 times every two years.
AI chips lined up in data centers
operate while exchanging vast amounts of data with each other.
The demand for communication speed has already exceeded the limits of conventional copper wiring.
The answer is millimeter-wave communication and silicon photonics.
Connecting chips with millimeter waves. Connecting with light.
From electrical wiring to radio and optical communication.
This is the revolution quietly,
but surely happening in data centers in the AI era.

However, a serious problem emerges here.
Millimeter-wave design, where the first step is fundamentally different,
is different from digital circuit design from the very first step.
Digital designers think in the time domain.
How many clock cycles after does the signal arrive?
Timing charts are their map.

Millimeter-wave designers think in the frequency domain.
Which band does this spurious emission affect?
Spectrum analyzers are their map.
This is not just a "difference in habit."
The recognition models for physical phenomena are fundamentally different.

Do you "know" Maxwell's equations, or do you "feel" them?
Do you "refer to" transmission equations, or are they "ingrained in your body"?
This difference decisively determines the quality of the design.
And at the 60GHz band, wiring on the board is no longer a "line." At this frequency where the wavelength is about 5mm,
even a small wiring pattern on the board acts as an antenna,
and a single connector or a single solder joint significantly affects the characteristics.
Chip design, package design, board design.
If there is a problem anywhere, it will malfunction.
Engineers who can grasp all of this as a single electromagnetic field problem
are critically lacking in the current AI era.

Once, there was the same failure.
This is not a new problem.
In one era, such a demand arose in the automotive industry.
"Integrate two communication chips into one for cost reduction."

It was correct as a social demand.
Engineers had to achieve it.
However, as a result of digital designers leading the foray into high-frequency design,
it did not work in the field.

The cause was clear.
Switching noise from digital circuits contaminated the power lines,
and mixed directly into the high-frequency circuits.
Ground design was handled in the digital style.
The concept of impedance matching was missing.
The simulator said it would "work."
It did not work in the field.
The structure of that failure is now about to be reproduced on a larger scale and in a more serious form in the AI era.

Simulators are just hypotheses.
Simulators provide the correct answer for an ideal world.
Ideal element constants. Ideal grounds. Ideal boundary conditions.
Temperature is constant. There is no aging. These are the answers in such a world.

In the real world, there is variation. There is temperature. There is humidity.
There is aging. And above all, there is the human hand.

The subtle temperature and time of soldering.
The tightening torque of connectors. How cables are bent.
Fine-tuning the orientation of the antenna.
These are all artisan knowledge that cannot be quantified.

And there is one more thing, the most important thing. The time spent facing measuring instruments.
The eyes that read the screen of a spectrum analyzer.
The intuition to track "where this spurious emission is coming from."
The sense felt in the body of "is it a contact failure or is it coupling?" when characteristics change just by touching it. These can only be cultivated in the time spent standing in front of the experimental bench.

Spending as much time in front of the experimental bench as you spend facing the simulator. This is the only path to becoming a full-fledged millimeter-wave designer.

How many times have I seen the correct answer on the desk become the wrong answer in the field? What digital efficiency has taken away
Digitization and efficiency have brought us much convenience.
At the same time, quietly but surely, they have taken something away.

The opportunity for thought.

In the era of manual calculation, there was physical meaning in the intermediate equations.
Intuition was cultivated in the process of calculation.
In the era of recording measurement results by hand, the act of drawing diagrams by hand deepened thought.
Tools have automated everything. Answers come out instantly. The process has disappeared.
Digital design can make you a professional in two years.
Millimeter-wave/analog design takes five years.
This asymmetry is not a problem of training costs. It is the reality that time for experience, failure, and struggle is physically necessary.

Efficiency judged that time as "waste."
The world looks entirely digital.
However, at the technical front lines, the bill for that efficiency is now erupting.

The talent needed for millimeter-wave design in the AI era is not being cultivated.
Leave analog thinking behind. The natural world is analog.
Digital is merely an approximation created by humans.

This major premise is being forgotten.
Analog thinking is thinking continuously.
"Why is this value this size?"
"What happens if it changes a little more?"
"Where is the limit?"
"How much margin is left?"
It is an attitude of continuing to pursue questions without answers.

Maxwell's equations are the language of nature. An engineer who feels them can intuitively sense when a simulator is lying. The body reacts to abnormalities in measured values. In the field, you can tell when "something is wrong."
This intuition cannot be replaced by AI.

Cutting-edge AI is waiting for analog thinking
AI is a powerful tool. It supports design, proposes optimizations, and processes vast amounts of data.
However, what AI is learning is past design data. Much of it is simulation results.
In other words, AI is learning a large amount of correct answers for an ideal world.

AI that has learned armchair theories proposes more sophisticated armchair theories.
Who will see through that risk?

Is the setting of boundary conditions correct?
Are near-field and far-field being confused?
Is the ground processing consistent with reality?
Where are the pitfalls in AI's proposals?

The only ones who can judge that are humans who feel Maxwell.
The world looks entirely digital.
However, what cutting-edge AI technology needs right now is,
ironically, people who have analog thinking in their bodies.

The intuition of engineers who did not get swept away by the wave of efficiency and continued to stand in front of the experimental bench.
The sense of designers who have conversed with nature, not simulators.
The field knowledge that 60GHz millimeter-wave design experience holds.
This becomes the only foundation for using AI correctly.

Tools cannot exceed the quality of the humans who use them.
Only when a human who feels millimeter waves uses AI will communication design for the AI era be completed.

Cutting-edge AI is now waiting for humans with analog thinking.