Who Will Seize the 75 Trillion Won AI Semiconductor Market? [Juhyung Lim's Tech Talk]

'AI Semiconductor' Designed for AI Computation
Sales Expected to Reach 75 Trillion Won by 2025
Companies from the US, UK, and Others Engage in a 'Silent War'

Artificial Intelligence / Photo by Yonhap News

[Asia Economy Reporter Lim Juhyung] Artificial Intelligence (AI) has long since permeated our daily lives. The video-sharing application 'TikTok' uses AI to display recommended videos tailored to user preferences. AI assistants like Apple's 'Siri' and Samsung Electronics' 'Bixby' have also become familiar to us.

AI has rapidly improved its performance to the point where it can now assist human tasks. This is thanks to deep learning technology, which continuously trains machines by processing vast amounts of data. Through deep learning, AI identifies patterns and rules embedded within data and uses them to perform tasks.

However, deep learning has its limitations. Creating intelligent AI requires enormous amounts of data and highly complex deep learning models to process that data. Especially as industries demand more powerful AI, the size of deep learning models and datasets is growing explosively.

Because of this, traditional semiconductors like Central Processing Units (CPUs) cannot handle AI software effectively. To process massive program sizes and data, new computer chips specialized for 'parallel computing' are needed. These are the 'AI semiconductors' specialized for AI processing.

According to the strategy consulting firm McKinsey, the AI semiconductor market revenue is expected to grow to $67 billion (approximately 75 trillion KRW) by 2025. Currently, companies worldwide are engaged in a 'war without gunfire' to secure dominance in the future AI semiconductor market.

◆NVIDIA Grabs the Center of AI Semiconductors

The company leading AI semiconductor technology today is the U.S.-based GPU manufacturer NVIDIA. In fact, NVIDIA's main product, the GPU, was originally developed not for AI but for handling complex 3D graphics in games.

NVIDIA's latest GPU 'A100' unveiled last July / Photo by NVIDIA

However, it was discovered that the so-called 'parallel computing' architecture, which processes multiple commands simultaneously for graphics processing, is also optimized for AI algorithm computations. As a result, NVIDIA GPUs have become central to AI semiconductors today.

Another strength of NVIDIA is its vast development ecosystem. NVIDIA supports developers with software development kits such as CUDA, cuDNN, TensorRT, and OptiX, and the number of developers creating software using these tools exceeds 2 million.

◆Google Accelerates AI Chip Research Following AlphaGo

U.S. IT giant Google is also accelerating its AI semiconductor development. Google already possesses world-class technology in AI software. In 2014, it acquired the UK AI research company DeepMind, and in 2016, DeepMind's Go-playing AI 'AlphaGo' defeated 9-dan professional Lee Sedol in a match.

AI semiconductor TPU server developed by the US IT company Google. TPU serves as the brain of 'AlphaGo.' / Photo by Google

Google is currently designing the supercomputer semiconductor that serves as the core of AlphaGo. This AI computation-specialized semiconductor is called the 'TPU.' Notably, Google's 4th generation TPU, unveiled last July, demonstrated performance comparable to NVIDIA's latest GPUs in 'MLPerf,' an AI semiconductor performance test hosted by global U.S. companies and universities.

◆Startup Graphcore Invents a New Concept Semiconductor

In the UK, the birthplace of AlphaGo's AI, AI semiconductor companies challenging NVIDIA's dominance are emerging. The most representative company is the startup 'Graphcore,' founded in 2016. Within four years of its establishment, Graphcore has grown into a strong company valued at $2 billion (approximately 2.24 trillion KRW).

The UK's AI semiconductor designer 'Graphcore' unveiled the second-generation intelligent processing unit platform 'IPU-Machine M2000' last July / Photo by Graphcore

Graphcore is famous for designing semiconductors specialized solely for AI program computations. Their computer chip, called the 'Intelligence Processing Unit (IPU),' features over 1,000 cores arranged like blocks within a single semiconductor to enhance parallel computing. The IPU, now in its second generation, is rapidly gaining customers as it is added to cloud services of major IT companies such as Microsoft and Dell.

◆Government Research Support in Korea

Currently, AI semiconductors are a competitive arena dominated by traditional AI powerhouses like the U.S. and the UK, but other countries' companies are likely to introduce their own unique computer chips soon. In this context, Korea's government unveiled a research and development investment plan last month on the 12th to leap forward as a leading AI semiconductor nation.

According to the vision titled 'Realizing a Leading AI Semiconductor Nation and a Comprehensive Semiconductor Power,' jointly announced by the Ministry of Science and ICT and the Ministry of Trade, Industry and Energy, the government plans to invest over 1 trillion KRW over the next ten years from this year through 2030.

In particular, the government plans to invest in a new technology called 'Processor-In-Memory (PIM),' tailored to Korea's status as the world's largest memory semiconductor producer. PIM is a semiconductor that integrates some computational functions within memory storage devices and is considered a next-generation chip that supports the core functions of AI computers.

This content was produced with the assistance of AI translation services.