Large Language Model (LLM) artificial intelligence (AI) began to be more actively used when the latest information was reflected in responses. It is considered to have started with Microsoft's Copilot, which combines OpenAI's GPT-4 and Bing search, released in March 2023. Previously, when asked, "Who is the current President of the United States?" LLM would often provide incorrect answers based on data learned 1-2 years prior. Copilot applied Retrieval-Augmented Generation (RAG) technology that references external knowledge to generate responses, meaning it supplements responses with the latest searched data.

Dinotisia is a startup with technology that allows Retrieval-Augmented Generation to operate more accurately and quickly. It is a technology that rapidly searches and inputs high-quality data matching the user's question intent to enable LLM to generate good responses. A vector database (DB) is used here, and for the first time in the world, vector DB search technology has been implemented in a chip. Hardware chips are much faster and more economical than software.

Jeong Moo-kyung, CEO of Dinotisia (49), met at the office in Seocho-gu, Seoul, on the 28th of last month, said, "While many vector DB companies in Silicon Valley process data mainly through software, we are the only ones to create a dedicated chip," adding, "The speed is 10 times faster, and data center costs can be reduced by 80%."

● "In the AI era, vector-based data processing is key"

Computers do not understand words or images as they are. Therefore, a representative method used is to convert words and images into coordinates (vectors) for utilization. For example, 'cat' is converted to [0.8, 0.3, 0.5], and 'dog' to [0.7, 0.4, 0.6]. The technology is to arrange them appropriately for the purpose. In this case, cats and dogs are relatively close in three-dimensional space. Computers consider vectors with similar directions and close distances to have similar meanings. The word 'mammal' is also placed relatively close to cats and dogs. Managing data based on vectors is what a vector DB does. In practice, vectors used are not three-dimensional but can rise to thousands of dimensions. Calculating the direction and distance of vectors requires billions of operations.

When a user asks, "Tell me about recently discovered mammals," the Retrieval-Augmented Generation process begins, and the vector DB presents information close to the mammal coordinates as results, and the LLM generates a response based on this content. Good search results lead to good responses.

● Completing the work of 10 servers with 1

Computer devices have role divisions. The Central Processing Unit (CPU) is the general manager, and the Graphics Processing Unit (GPU) and Neural Processing Unit (NPU) are AI inference experts. However, there was no expert in 'vector search.' That was until Dinotisia created the Vector Data Processing Unit (VDPU), a vector DB chip.

According to Dinotisia, the cost-performance ratio of search performance is more than 10 times better than that of CPUs and GPUs. It also consumes less power. The total cost of ownership (TCO) of data centers can be reduced by 80%, and the data center space can be significantly reduced. CEO Jeong stated, "The work that required 10 servers can be completed with almost one," adding, "It will contribute to breaking down the cost barrier that distinguished companies that can properly use AI from those that cannot."

Dinotisia is currently verifying the vector DB chip as a programmable semiconductor (FPGA) in a prototype capacity. It plans to release it as an application-specific integrated circuit (ASIC) chip in the second half of 2026. It has signed a development contract with the semiconductor design house ASICLAND, and mass production will be carried out by TSMC. Cloud companies and data storage companies are customers.

● Semiconductor, DB, and AI experts unite

The event that showed the feasibility of artificial intelligence through deep learning was when the deep learning model 'AlexNet' won first place in the ImageNet image recognition competition in 2012. CEO Jeong, who received a doctorate in semiconductors from KAIST in 2006, was developing GPUs at Samsung Electronics' Advanced Institute of Technology at the time. Seeing the emergence of AlexNet, he felt, "This (deep learning) seems really possible."

He created an opportunity by moving to SK Telecom in 2015, where he could relatively freely choose research projects. He simultaneously researched and developed databases, GPU clouds, and AI semiconductors. In 2017, he became a team leader and started the NPU project 'SAPEON,' and in 2020, he commercialized the first domestic NPU chip for data centers.

Then he felt the importance of vector-based search was rapidly increasing as the scale of data utilized by deep learning models grew. He decided to start a business, judging that a dedicated hardware for vector operations, separate from existing NPUs, was needed.

Dinotisia opened in 2023. About 20 years of experience, including semiconductor design learned at Samsung and databases and AI infrastructure accumulated at SK, were combined. To create a vector DB acceleration chip, knowledge of semiconductors, databases, and AI is required. Dinotisia gathered such talents.

In April this year, Dinotisia was named in CB Insights' 'AI 100' just 22 months after its establishment. It is a list of 100 innovative AI companies worldwide. It was selected in the newly created 'vector database' category this year, along with a company from Germany.

● AI that thinks in Korean, personal assistant on the desk

Dinotisia is targeting various fields based on vector DB technology. In December 2024, it released the AI model 'DNA,' specialized in Korean processing, as open source. This model ranked first in the Korean specialized knowledge evaluation at the time, surpassing LG Exaone. In March this year, it also announced 'DNA-R1,' specialized in Korean logical reasoning. It is the first model in Korea where the entire inference process is conducted in Korean, and the company explains that it understands the subtle context unique to Korean well.

This year, it introduced the vector DB solution 'Seahorse Cloud' in the form of Software as a Service (SaaS). Being browser-based, it allows the construction and operation of vector DB without separate installation, and it is characterized by providing vector DB, RAG, and AI agents integratively in a cloud environment.

In August, it launched the AI device 'Nimos Workstation,' which does not require a cloud connection. It is highly useful in fields with high data security requirements, such as finance, healthcare, and defense, where company secrets cannot be sent outside. The 'Nimos Workstation' integrates storage, GPU, DNA model, vector DB Seahorse, and AI agent functions into one device. Without the need to separately build complex AI infrastructure as before, it has a structure that allows high-performance computation and advanced AI tasks to be processed in real-time based on integrated software in a single workstation.

CEO Jeong stated, "The advancement of AI technology is now shifting from model size to the precision of data search and inference optimization," adding, "If AI can efficiently reference accurate information from data, it will respond like a person with long-term memory, and we will bring this era closer."

Popular News