[Editor's Note] With Korea recording a low market share of around 1% in the global fabless market, the growth of fabless is becoming increasingly important for the growth of system semiconductors. In the midst of this, Sapion is attracting attention as it continues to develop its own semiconductor design capabilities to compete with global big tech companies such as Apple. Sapion is an independent corporation launched by SK Telecom earlier this year to challenge the global AI semiconductor market. Sapion's 'X220' was also evaluated as having superior performance in areas such as data center inference compared to third-party GPUs in the MLPerf test. We met with Ryu Su-jeong, CEO of Sapion, to hear about AI semiconductor trends and Sapion’s future plans. ■ It has been over a year since Sapion was established. Do you have a customer base at this point?

There are still some areas that are difficult to comment on as they are still in the expansion phase, but there are areas that are being continuously discussed.

Of course, it is mainly on the 'hyperscaler' side, and there are areas where we need to strategically cooperate, so we are continuing to discuss them.
*Hyperscaler: A company that operates large-scale data centers.

■ What is your strategy for expanding the market while operating the company simultaneously with the US corporation?

Sapion's workforce is not that large yet. So for now, we plan to secure the business area first and then grow it with a focus on software. We plan to focus on building a business ecosystem as well as research personnel.

Although our country's industry has grown significantly compared to before, its domestic size is very small in terms of market size. Ultimately, the reason for establishing the US corporation can be seen as an attempt to conquer the global market.

■ I heard that Sapion started out in collaboration with SK and plans to grow into a company with a corporate value of 10 trillion won by 2027. I am curious about the current level of achievement.

It's a weak start, but so far it doesn't seem too bad.

Sapion has recently been focusing on commercializing the x220, which has shown more than twice the performance of Nvidia GPUs in MLPerf tests. Based on this performance competitiveness, the company expects to continue to expand the market until next year.

Also, the next version of the x330 is scheduled to be released after that. It is expected to have a competitive edge with four times the current performance.

I think this x330 product will be a great help in expanding the market in the future. We are also considering entering the mass market after the x330, so we will continue to discuss with customers and continue technical exchanges.

■ If the x330 is released in the first half of next year, when do you think it will be commercialized?

I guess we'll have to wait and see for a while. First of all, we'll have to continue updating it after the release, so it might take some more time until it's commercialized.

After the chip is released, we plan to verify it through a POC (Proof of Concept) and implement the requirements one by one in this order. It is expected to be used internally first.

Also, since there are differences depending on the workload, I think it is premature to determine the target figure at this point. It depends on what kind of applications you will be running in the future, but if you migrate what you were running before, the period may be shortened. It is difficult to say for sure at the moment because you have to consider the characteristics of the network, etc.

■ The biggest difference between the x220 and x330 is

Sapion places the highest priority on power efficiency.

From this perspective, it can be said that this is a differentiating point in itself, as it improves by approximately 4 times. Improved learning performance can also be seen as a strategic point.

In particular, Sapion is specialized in the data center inference area, and this x330 has improved efficiency by changing the data type in areas where accuracy was previously insufficient. This is expected to also reduce costs.

■ The areas where research and projects to be specialized will be carried out next year

We believe that vision-based AI services will become more widespread in practice, so we will continue to focus our research on this area as before.

And starting this year, the natural language processing sector based on Transformer has also begun to receive attention, and we are continuing research and development along with the next product to be released in line with trends such as large models.

Ultimately, one is to strengthen the vision-related application service and continue research so that many services can be implemented together based on it, and the other is to prepare for a new network based on transformers.

■ In the future, Sapion’s business partnership will expand in the following direction:

Sapion is conducting a server verification project in collaboration with NHN. As of now, we are collaborating with SK Telecom, SK Shields for security such as CCTV, and externally with a medical company called Health Hub.

NHN Cloud has built a 5.22 Peta OPS level Sapion AI infrastructure in the NHN data center, and has further expanded and built a 9.22 Peta OPS performance Sapion AI infrastructure this year.

As such, partnerships are extremely important to Sapion. In addition to NHN, we plan to expand our support to small and medium-sized businesses in the future based on this and actively participate in creating a domestic AI ecosystem.