[자막뉴스] 엔비디아 능가한 '초저전력 AI 반도체'...한국 세계 최초로 개발 / YTN

Summary

South Korean researchers have developed the world's first neuromorphic AI chip that operates at ultra-low power, mimicking the human brain. This breakthrough chip, only 4.5mm in size, can run complex AI models like GPT without external servers, significantly reducing computation time and power consumption. It integrates two AI methods for complementary efficiency, allowing it to perform tasks such as summarizing text, translating languages, and generating responses from learned data. The chip's potential applications are vast, with ongoing research to expand its capabilities beyond language models.

Takeaways

• 📱 The script discusses the development of next-generation IT devices that can run AI on-device without external servers, emphasizing the challenges due to the high power and size requirements of AI chips.
• 🧠 South Korean researchers have developed the world's first neuromorphic AI chip that mimics the human brain, capable of running AI with ultra-low power consumption.
• 🔋 The chip is remarkably small, measuring only 4.5mm in width, and has been successfully tested by running GPT on a computer, demonstrating its ability to perform tasks such as summarizing text and translating languages without external assistance.
• ⏱️ The chip can process computations that would take a regular computer 45 minutes in just 3 minutes, showcasing its efficiency.
• 🔍 When compared to Nvidia's GPU A00 graphics card, the new chip consumes only 1/6 of the power and has 1/10 of the area.
• 📲 The chip has been implemented in a smartphone model called 'Galaxy 24', enabling it to perform language tasks as demonstrated.
• 🔄 The chip's design incorporates two AI methodologies, deep neural networks for high-accuracy computation with large datasets and spiking neural networks for efficient power usage with smaller datasets, creating a synergistic effect.
• 💡 The chip operates similarly to the human brain, adjusting its power consumption based on the amount of data it needs to process—using more power for larger tasks and less for smaller ones.
• 🌐 The achievement of running large language models like GPT on this neuromorphic chip is a world-first, indicating a significant breakthrough in AI and chip technology.
• 🔬 The research team is currently working on integrating the next version of GPT, GPT 3.0, into the chip and plans to expand their research to various application fields beyond language models.
• 🎥 The script is from YTN Science, providing a news update on this technological advancement.

Q & A

What is the main challenge in implementing on-device AI for devices like smartphones?

The main challenge is that AI requires complex computations, which in turn demand a lot of power and can result in larger chip sizes.

What has the Korean research team developed that is a world first?

The team has developed a neuromimetic AI chip that operates artificial intelligence at ultra-low power consumption, mimicking the human brain.

What is the size of the developed chip?

The chip is only 4.5mm in width.

How did the team demonstrate the chip's capabilities?

They mounted the chip on a computer and ran GPT to perform tasks such as summarizing text and translating languages without external server assistance.

What is the performance improvement when the chip is used in comparison to a regular computer?

The chip can process computations that would take a regular computer 45 minutes in just 3 minutes.

How does the power consumption of the developed chip compare to Nvidia's GPU A00 graphics card?

The power consumption of the developed chip is only 1/61 of that of the Nvidia GPU A00 graphics card.

What smartphone model was used to implement the chip for language tasks?

The Samsung Galaxy 24 smartphone model was used to connect the board and implement the language tasks.

What are the two AI methods incorporated in the chip that allow for ultra-low power and small size?

The chip incorporates deep neural networks for high computation accuracy with large data inputs and spiking neural networks for power efficiency with small data inputs.

How do the two AI methods work together in the chip?

They work complementarily, with the deep neural network taking over for high-accuracy computations and the spiking neural network handling power-efficient processing for smaller data inputs, creating a synergistic effect.

What is the current focus of the research team in terms of the chip's development?

The team is currently working on integrating the next version of GPT, GPT 3.0, into the chip and expanding the research scope to various application fields.

What is the significance of this development in the field of AI and semiconductors?

This development is significant as it marks the world's first successful operation of a large language model like GPT using a neuromimetic, ultra-low power semiconductor, paving the way for more efficient and compact AI solutions.