power plant for AI

a power plant for AI

1. It seems that 260,000 GPUs for AI are being brought in. Nvidia must have needed government and national sales performance. They talk about whether 0.5GW is enough to use artificial intelligence data centers. Data centers use power 24 hours a day, 365 days a year, so it’s not a small amount of power.
2. The capacity of power facilities in Korea is 150GW, but the average usage is about 54%. This means that half of the power facilities are playing. The addition of 0.5GW is unlikely to be a problem for the country’s power grid.
3. Data centers are like the heart of the artificial intelligence industry, so strong data centers must be attracted to meet the demand for M7 clouds in the United States. The biggest obstacle to data center construction is insufficient power and whether it meets RE100.
4. Since the U.S. M7 data center will not move to China, Korea, which has developed ICT industry and well-connected optical communication network, is most appropriate for outsourcing from the West. The problem is insufficient RE100 power.
5. A typical problem-solving method in engineering is “improvement.” When a problem like this occurs in the previous one, we add one method or another… We solve the problem by dividing it into small pieces like a small ball in baseball. This is the improvement method called “改” in Japanese style. The advantage of “Kaizen” is that it does not touch the existing system as much as possible and changes it step by step.
6. Korea is making some drastic design changes. It’s gone now, but a disk-type recording device called DVP was developed, which was developed from a recording device called VCR in the past. When we disassembled Japanese products at that time, they had all kinds of functions, so they were developed in a very awkward way, but when we analyze the cost, they looked terrible. Why do they design this way? They look good on the outside, but the internal structure is pretty beautiful, too. But the problem seems to be that they are not competitive in cost. Later on, the VCR and DVP markets disappeared, and the OTT market such as NEPLIX changed. New product development engineers are doomed to die if they do not change.
7. The former parent company developed and exported the printer globally. Dot printer, black and white printer, color printer, was made using ink or laser. During technology training, I met a colleague from the printer business department and cited difficulties in developing the printer. It was difficult to reduce the cost of the printer. The motive was the development product improvement team. I asked why it was difficult. … The structure is divided and there are many harnesses and connectors… He said, “There are almost 100 of them. We have to reduce harnesses and connectors somehow, but it is not easy to add continuous functions.”
8. A harness is a kind of parallel wire and cable. I think improving each and every one of these intuitively can’t solve the problem. We should get rid of them altogether, not reducing the harness. I should think of a zero harness design… That’s how he advised me. My motivation ignored my advice thinking it was ridiculous, but a few years later, all the cables in the printer were gone. Whether it was improving the SoC that was using FPCB or changing the intermodule connection assembly method, the harness eventually disappeared and all the cables in the system that seemed to be scattered over the years disappeared.
9. Korea’s strength is not so much about improvement but about redesigning the ground and changing it boldly. It needs a change of thoughts and thinking. A 10 percent improvement can be made by squeezing dry towels, such as early morning stargazing, not drinking water, and the Cheonlima exercise, but 99 percent improvement is impossible by squeezing people.
10. Solar and wind power generation are essential for attracting data centers. If this power is not procured, it cannot be used by foreign companies. In this case, there is no answer if you add 100MW and 1GW in the existing method.
11. We need to think about what we should do to secure 1TW (1,000GW) nationally. How to transmit and store such large amounts of power. Where to use it, how to charge it, how to disperse it. If you make small improvements and add more, you won’t get an answer.
12. It is said that data centers are difficult to operate due to lack of power in the United States, so is it not necessary to operate them in the United States? There is no reason to operate in the expensive United States as long as optical communication is developed, the power grid is excellent, and new and renewable energy is procured. Islands such as Jeju Island should also be connected to two-way HVDCs so that the remaining power can be transferred to each other by land and from land to Jeju.
13. It is necessary to study how much ESS BESS should be installed to ensure that the electrical grid is not dangerous, write a paper, and demonstrate it. The cost required should be reduced through breakthrough improvements… It is not achieved with small improvements, but it is solved only by drawing a big picture from the bottom nationally. You might think it would make sense to procure 1 TW when Korea’s power generation capacity is 0.15 TW, and the average used power is 0.08 TW.
14. It is also possible to relay renewable energy power procurement between China and Japan. There is no reason to use only electricity generated from Korea’s narrow land. Renewable energy generated in Tibet from Inner Mongolia can cross the Yellow Sea to Korea via the HVDC grid. Renewable energy power that is wasted because it cannot be used from Albon Kyushu to Tohoku can be transmitted across the seabed, supplied to the data center, and stored in BESS.
15. This solution can be solved only when you have a grand goal. If you think about how to build a 1MW class power plant to secure 1,000 power for the 1MW class power plant, you will solve a small problem. Who knows if the data center will really use 1TW-10TW power in the future?