The hottest company in the world is NVIDIA, which has achieved the No.1 position in global market capitalization. The biggest trend of 2023 has been the emergence of generative AI such as ChatGPT and the “AI computing” that powers it. The competitiveness of such AI technologies ultimately depends on GPU performance. Key indicators include how much computation GPU-based systems can process and how quickly they can handle it. As the dominant player in the GPU market, NVIDIA has achieved outstanding performance, including ranking first globally in market capitalization and recording KRW 43 trillion in net profit in 2023. In particular, with a 98% share of the data center GPU market, its monopolistic position is expected to continue.
[Fig] NVIDIA’s Performance Trend and Data Center GPU Market Share Outlook
Recently, NVIDIA launched a new GPU, Blackwell, which is set to redefine the AI computing market. The Blackwell GPU integrates 208 billion transistors, more than double that of its predecessor, Hopper. It offers approximately four times faster AI training performance compared to previous GPUs. This technological advancement is expected to shift the paradigm of AI computing once again. Major tech companies are already adopting it, with Meta planning to use it for training its large language models (LLMs), and deployment also expected across AWS (Amazon Web Services), Google Cloud, and Microsoft Azure.
[Fig] NVIDIA’s Latest GPU “Blackwell” (Left) and Previous GPU “Hopper” (Right)
To maintain its technological leadership, NVIDIA continues to file patents across various GPU technologies. Recent patent trends can be broadly categorized into “resource allocation/partitioning” and “GPU tuning.”
Resource Allocation/Partitioning Technology refers to optimizing how GPU resources are distributed and managed for efficient operation. A representative example is EFFICIENT GPU RESOURCE ALLOCATION OPTIMIZATION METHOD AND SYSTEM (US 1 801 1831). Conventional methods allocate unused resources based on memory access (NUMA) packets, leading to inefficiencies and low utilization. This patented technology optimizes GPU resource allocation, improving processing efficiency and increasing NUMA packet utilization.
[Fig] Representative Diagram of Resource Allocation Technology (US 1 801 1831) and Allocation Example
GPU Tuning Technology refers to real-time optimization of GPU performance and power efficiency. A key example is REAL-TIME HARDWARE-ASSISTED GPU TUNING USING MACHINE LEARNING (US 1 793 2539). Traditional GPU processing suffers from pixel bottlenecks during machine learning workloads, causing a drop in clock speed. This technology resolves such bottlenecks by dynamically tuning cache usage and pixel distribution in real time, thereby stabilizing performance.
[Fig] Representative Diagram of GPU Tuning Technology (US 1 793 2539) and Tuning Example
Amid NVIDIA’s strong performance, a counter-movement is emerging, led by AMD. In December 2023, AMD released its latest GPU, MI300X. Compared to NVIDIA’s H100, MI300X features 2.4 times more high-bandwidth memory (HBM) and 1.6 times higher bandwidth. Based on these specifications, AMD claims that its GPU can deliver up to twice the performance of NVIDIA’s GPUs in AI training and inference.
[Fig] AMD’s Latest GPU “MI300X” and the AI Chip “MI300”
NVIDIA’s growth is also benefiting Korean companies. A prime example is the AI data center industry, which is built using NVIDIA’s AI GPUs such as the H100. Recently, NHN Cloud became the first in Asia to adopt the H100 to build an AI data center. The demand for cloud-based AI data centers using cutting-edge GPUs is growing rapidly, not only for NHN but also for major Korean companies such as Naver and Samsung SDS. This expanding market is opening new opportunities for related industries.
[Fig] NHN’s Data Center Utilizing NVIDIA H100 and Market Outlook
In addition, collaboration with Korean semiconductor backend companies is strengthening. Partnerships with domestic OSAT (Outsourced Semiconductor Assembly and Test) companies are becoming increasingly important. In this ecosystem, AI chips designed by U.S. fabless companies are packaged and tested by Korean OSAT firms. Furthermore, supply agreements have been established with SK Hynix for high-bandwidth memory (HBM) used in AI chips. As a result, strong performance is also expected from Korea’s semiconductor-related industries.
[Fig] Overview of NVIDIA’s Collaboration with Korean OSAT Companies and Backend Process
The BLT Patent & Law Firm Research Center stated that “the rapidly evolving AI and GPU markets are creating new opportunities, and if Korean companies secure competitive technologies, they will be well-positioned to lead the next era.”
As of 2024, BLT Patent & Law Firm has been selected as a partner by more than 2,000 innovative startups, supporting corporate growth and success through IP acquisition, strategy development, investment attraction, and IP-based business support such as technology-special listings.
#NVIDIA #GPU #AIComputing #Blackwell #Hopper #H100 #AMD #MI300X #AIDataCenter #HBM
The hottest company in the world is NVIDIA, which has achieved the No.1 position in global market capitalization. The biggest trend of 2023 has been the emergence of generative AI such as ChatGPT and the “AI computing” that powers it. The competitiveness of such AI technologies ultimately depends on GPU performance. Key indicators include how much computation GPU-based systems can process and how quickly they can handle it. As the dominant player in the GPU market, NVIDIA has achieved outstanding performance, including ranking first globally in market capitalization and recording KRW 43 trillion in net profit in 2023. In particular, with a 98% share of the data center GPU market, its monopolistic position is expected to continue.
[Fig] NVIDIA’s Performance Trend and Data Center GPU Market Share Outlook
Recently, NVIDIA launched a new GPU, Blackwell, which is set to redefine the AI computing market. The Blackwell GPU integrates 208 billion transistors, more than double that of its predecessor, Hopper. It offers approximately four times faster AI training performance compared to previous GPUs. This technological advancement is expected to shift the paradigm of AI computing once again. Major tech companies are already adopting it, with Meta planning to use it for training its large language models (LLMs), and deployment also expected across AWS (Amazon Web Services), Google Cloud, and Microsoft Azure.
[Fig] NVIDIA’s Latest GPU “Blackwell” (Left) and Previous GPU “Hopper” (Right)
To maintain its technological leadership, NVIDIA continues to file patents across various GPU technologies. Recent patent trends can be broadly categorized into “resource allocation/partitioning” and “GPU tuning.”
Resource Allocation/Partitioning Technology refers to optimizing how GPU resources are distributed and managed for efficient operation. A representative example is EFFICIENT GPU RESOURCE ALLOCATION OPTIMIZATION METHOD AND SYSTEM (US 1 801 1831). Conventional methods allocate unused resources based on memory access (NUMA) packets, leading to inefficiencies and low utilization. This patented technology optimizes GPU resource allocation, improving processing efficiency and increasing NUMA packet utilization.
[Fig] Representative Diagram of Resource Allocation Technology (US 1 801 1831) and Allocation Example
GPU Tuning Technology refers to real-time optimization of GPU performance and power efficiency. A key example is REAL-TIME HARDWARE-ASSISTED GPU TUNING USING MACHINE LEARNING (US 1 793 2539). Traditional GPU processing suffers from pixel bottlenecks during machine learning workloads, causing a drop in clock speed. This technology resolves such bottlenecks by dynamically tuning cache usage and pixel distribution in real time, thereby stabilizing performance.
[Fig] Representative Diagram of GPU Tuning Technology (US 1 793 2539) and Tuning Example
Amid NVIDIA’s strong performance, a counter-movement is emerging, led by AMD. In December 2023, AMD released its latest GPU, MI300X. Compared to NVIDIA’s H100, MI300X features 2.4 times more high-bandwidth memory (HBM) and 1.6 times higher bandwidth. Based on these specifications, AMD claims that its GPU can deliver up to twice the performance of NVIDIA’s GPUs in AI training and inference.
[Fig] AMD’s Latest GPU “MI300X” and the AI Chip “MI300”
NVIDIA’s growth is also benefiting Korean companies. A prime example is the AI data center industry, which is built using NVIDIA’s AI GPUs such as the H100. Recently, NHN Cloud became the first in Asia to adopt the H100 to build an AI data center. The demand for cloud-based AI data centers using cutting-edge GPUs is growing rapidly, not only for NHN but also for major Korean companies such as Naver and Samsung SDS. This expanding market is opening new opportunities for related industries.
[Fig] NHN’s Data Center Utilizing NVIDIA H100 and Market Outlook
In addition, collaboration with Korean semiconductor backend companies is strengthening. Partnerships with domestic OSAT (Outsourced Semiconductor Assembly and Test) companies are becoming increasingly important. In this ecosystem, AI chips designed by U.S. fabless companies are packaged and tested by Korean OSAT firms. Furthermore, supply agreements have been established with SK Hynix for high-bandwidth memory (HBM) used in AI chips. As a result, strong performance is also expected from Korea’s semiconductor-related industries.
[Fig] Overview of NVIDIA’s Collaboration with Korean OSAT Companies and Backend Process
The BLT Patent & Law Firm Research Center stated that “the rapidly evolving AI and GPU markets are creating new opportunities, and if Korean companies secure competitive technologies, they will be well-positioned to lead the next era.”
As of 2024, BLT Patent & Law Firm has been selected as a partner by more than 2,000 innovative startups, supporting corporate growth and success through IP acquisition, strategy development, investment attraction, and IP-based business support such as technology-special listings.
#NVIDIA #GPU #AIComputing #Blackwell #Hopper #H100 #AMD #MI300X #AIDataCenter #HBM