Chapter 213: Luna
A few days later.
Ahsung Motors officially announced that Dojin Frontier had been selected as the final acquirer of its electric vehicle division.
At the same time,
Kang Tae-joon’s feed featured a new post.
[KTalk Feed] - @Kang Tae-joon | 113,172,592 followers
Restructuring the electric vehicle division
Accidents like EV-Beta will never happen again
EV-Nova delivery timeline shortened
End of 2019 → July 2018
#Cheon Ho-tae #Ahsung Motors #Infrastructure #Nice #WillUseWell
└ㅋㅋㅋㅋㅋㅋㅋㅋ
└Are those hashtags making fun of Chairman Cheon Ho-tae?
└Not making fun, just teasing, I think.
└But Chairman Cheon is in prison, he can’t even see this.
└Why not? He uses his lawyer visitation time to the limit every day.
└Anyway, it’s great that the EV-Nova launch is speeding up, right?
└Of course it’s faster—they acquired the entire research team and engineers along with the factory.
└Production infrastructure was Dojin Frontier’s biggest weakness.
└If only Cheon Jun-young hadn’t caused that accident, Ahsung Motors wouldn’t have collapsed.
└Ugh… thinking about it makes me angry. What about my stocks?
└ㅋㅋㅋㅋㅋ Ahsung stocks are down more than half. What now?
└Meanwhile, Dojin Frontier’s stock price keeps rising.
└Current market cap: $175 billion ㄷㄷㄷ.
＊＊＊
Dojn Tech Research Building, 3rd Floor.
Guided by Park Sang-bae, Head of Equipment Innovation, I entered the lab housing the LUNA-EVX prototype equipment.
The lab was filled with a strange tension.
As I stopped walking, the researchers all looked up.
“I’ll start with a brief update on our progress.”
With everyone’s attention focused, Park Sang-bae began his explanation.
“What you see here is the LUNA-EVX prototype. Both the equipment and processes have stabilized to the point where we can perfectly implement the ‘Black Robin microarchitecture,’ which was difficult to achieve in the previous beta version.”
At Park Sang-bae’s gesture, a researcher displayed data on the monitor.
The screen showed complex circuit structures and various metrics.
“Compared to the previous process, gate density has improved by more than double, and circuit line width can now be stably implemented at 2.7nm.”
Chilbong added to the explanation.
“At this rate, we can produce the Black Robin chip with sufficient stability.”
Scanning the numbers on the screen, I asked in a low voice,
“What about neuromorphic implementation possibilities?”
The entire lab fell silent.
After a brief pause, Chilbong cautiously shook his head.
“While the LUNA-EVX prototype theoretically supports neuromorphic chip design, it’s impossible to implement with the current structure.”
“Be more specific.”
Chilbong continued calmly,
“Neuromorphic architecture differs fundamentally from parallel processing structures like the Black Robin. Black Robin processes large volumes of simple computations in parallel, like a traditional CPU, but neuromorphic architecture is based on an ultra-parallel network mimicking the human brain, requiring a completely different physical chip structure.”
“So what’s lacking in the LUNA-EVX prototype?”
This time, Park Sang-bae answered.
“The key feature of neuromorphic chips is the precision of ultra-fine interconnects between synapses and neurons. Our LUNA-EVX process can stably implement 2.7nm line widths, but neuromorphic chips require interconnect density and precision below 1nm.”
Chilbong took over,
“Additionally, neuromorphic chips require non-standard analog-digital mixed-signal processing, unlike traditional semiconductors. Instead of simple digital circuits, we need to form circuit networks that include analog signals. With the current EVX equipment, we need to develop additional processes to minimize interference and crosstalk between analog and digital signals.”
I nodded.
They had a better grasp of the issues than I expected.
The lab fell silent again as everyone held their breath, waiting for my next question.
“So how much more time is needed to mass-produce neuromorphic chips?”
“Prototype development alone will take at least three years. With the need to completely overhaul processes and equipment, full-scale mass production could take five years or more.”
“Then… what about Black Robin?”
This time, the answer came without hesitation.
“As Team Leader Park Sang-bae mentioned, Black Robin can be mass-produced with the prototype equipment.”
Park Sang-bae added,
“The LUNA-EVX 2.7nm process is perfectly optimized for Black Robin’s structure. No process changes are needed, and with minor adjustments, we can ensure stable production.”
“Then… for now, Luna will have to be based on Black Robin.”
If we could implement the AI platform Luna using neuromorphic technology, we wouldn’t just be ‘mimicking’ AI-level performance.
We’d be ‘replicating’ the learning and cognitive systems of the human brain.
But that goal is still far off.
‘Trying to fly from the start is greedy.’
Right now, we need to focus on learning to run.
Being able to move faster than walking is already significant.
“Assuming we start mass-producing Black Robin tomorrow, how long will it take to complete Luna?”
Chilbong looked slightly taken aback at my question.
“To fully implement the AI platform Luna, chip production alone isn’t enough.”
“What else is needed?”
Though I knew better than anyone,
this meeting included most of the core researchers working on Luna’s ‘hardware.’
I wanted them to explain the overall structure.
Seeming to understand my intent, Chilbong displayed internal documents on the screen.
The screen showed a diagram titled ‘LUNA Platform Components,’ with each element appearing in sequence.
“As I mentioned earlier, the computational module Black Robin is the core… it’s essentially the ‘brain.’ For this brain to function, we need a stable input/output structure, operating system, memory interface, API layer, and model launcher.”
Chilbong went through each detail meticulously.
“First, high-speed data I/O. We need a modular interface board with HBM memory and a PCIe Gen5 communication device. If data transmission bottlenecks occur during AI model training, even the best chip won’t perform well.”
“Isn’t this also being developed by the Luna Task Force?”
Chilbong nodded at my question.
“Correct. The I/O module currently under development by the Luna Task Force is designed around HBM memory. Specifically, we’re creating an expandable module based on an ultra-low-power HBM2e derivative originally developed for Quantum smartphones.”
“So you’re applying that to Luna too?”
“Yes, the HBM for Quantum was originally designed for mobile devices, focusing on minimizing power consumption. This significantly reduced heat and signal interference. By introducing HBM2e into Luna’s AI computation platform, we expect stable I/O processing and heat suppression even during high-speed computations.”
“What about the operating frequency?”
“The current prototype achieves a stable 3.6Gbps, with a bandwidth of 460GB/s per package. By improving the TSV (Through-Silicon Via) structure to match the LUNA-EVX process, we aim for over 4.2Gbps.”
These numbers already exceed the threshold for bottleneck-free data processing when combined with Black Robin.
“How will this be applied to future Quantum models?”
“We’ve divided it into a dual structure. For Quantum, heat and optimization are key, so we’re limiting it to a 2-channel structure. For Luna, we’re designing an expandable structure with 4 or more channels. The BGA packaging method will also differ.”
I slowly nodded.
“Alright. We have the computation chip and memory I/O. What about the OS?”
“For the OS… instead of me explaining, why don’t you hear it directly from Han Yu-ju, the Director at Dojin Soft?”
He suggested I hear it from the person in charge of development.
Not a bad idea.
As Chilbong reached for the phone to call Han Yu-ju, I stopped him.
“It’s been a while since I visited Dojin Soft. Might be nice to drop by.”
“I’ll have a car ready.”
“Do that.”
…
…
Dojin Soft.
Not far from my top-floor office,
but it had been a long time since I’d visited in person.
‘It’s still the same.’
Desks cluttered with documents.
Monitors filled with terminal windows and debugging code.
The intense energy in the air brought a smile to my face.
When I first set this up,
we were only using one floor—the 12th floor of the Dojin Industries building.
The entire staff consisted of about twenty members from Han Yu-ju’s alma mater, Korea University’s coding club ‘Lunar.’
But now, Dojin Soft employees occupied more than half of the 20-story Dojin Industries building.
The company had grown so much that a new headquarters was being built in Seoul’s Seocho District.
“Chairman.”
As I looked around, Han Yu-ju, who had been contacted, approached.
Since the last organizational restructuring, she had been promoted to Assistant Executive Vice President and took on the role of General Manager at Dojin Soft. Her demeanor had changed noticeably.
“These days, you must be going home regularly, right? Even washing your hair?”
I joked with a laugh.
Han Yu-ju rolled her eyes.
“When was that? These days, I meet so many people daily. If I lived like before, it would immediately damage the company’s image.”
No one could deny that Dojin Soft’s scale and presence had changed dramatically.
Centering on the smartphone OS ‘Lunar,’ which held over 80% market share,
the ecosystem included KTalk, KTalk Feed, mobile games, the App Store, an ad platform, Dojin Pay, fintech, and Pang Pang Mall.
Add to that the Lunar Subscription Service, based on cloud storage.
Just looking at the scale, they had already surpassed domestic platform giant Naver.
And evaluations suggested they would soon overtake global giants like Google, Amazon, and Facebook.
So her words weren’t surprising.
Han Yu-ju was essentially the face and core of Dojin Soft.
‘Which makes me even more curious.’
I decided to check firsthand how far the AI platform ‘Luna,’ which Han Yu-ju was responsible for, had come.