ÇÏÀ̺긮µå ¸Þ¸ð¸® Å¥ºê ½ÃÀåÀº ÀΰøÁö´É ¾ÖÇø®ÄÉÀ̼ÇÀÇ º¸±Þ È®´ë·Î ¿¬Æò±Õ ¾à 26.50%ÀÇ ³ôÀº ¼ºÀåÀÌ ¿¹»óµË´Ï´Ù. ¶ÇÇÑ, ÀΰøÁö´ÉÀÇ ´ëÁßȵµ ¼ºÀåÀÇ ÇÑ ¿äÀÎÀ¸·Î ÀÛ¿ëÇϰí ÀÖ½À´Ï´Ù. ÀÌ ±â¼ú ÀÚü´Â Àü·Ê ¾øÀÌ ºü¸¥ ¼Óµµ·Î ¹ßÀüÇϰí äÅÃÀÌ ÁøÇàµÇ°í ÀÖ½À´Ï´Ù. ÀÌ·¯ÇÑ AI·ÎÀÇ ÀüȯÀº °í¼º´É GPU¿¡ ´ëÇÑ ¼ö¿ä¸¦ Áõ°¡½ÃÄÑ HMC¿¡ À¯¸®ÇÑ ¼ö¿ä ȯ°æÀ» Á¶¼ºÇϰí ÀÖ½À´Ï´Ù. ¿¹¸¦ µé¾î, 2024³â 2¿ù ¼¼°è ÃÖ´ë GPU Á¦Á¶¾÷üÀÎ ¿£ºñµð¾Æ(Nvidia)´Â Àü ¼¼°è°¡ ¸ñ°ÝÇϰí ÀÖ´Â AI ºÕÀ¸·Î ÀÎÇØ µ¥ÀÌÅͼ¾ÅÍ Ä¨°ú GPU¿¡ ´ëÇÑ ¼ö¿ä°¡ Áõ°¡ÇÏ¿© ºÐ±âº° ¼öÀÍÀÌ 3¹è Áõ°¡ÇÒ °ÍÀ̶ó´Â ¿¹ÃøÀ» ¹ßÇ¥Çß½À´Ï´Ù. ¶ÇÇÑ, ±â¾÷µéÀº Ŭ¶ó¿ìµå·Î ¾÷¹«¸¦ ÀüȯÇϰí ÀÖ½À´Ï´Ù. ÀÌ·¯ÇÑ Å¬¶ó¿ìµå·ÎÀÇ ÀüȯÀº µ¥ÀÌÅͼ¾ÅÍ ¼ö¿ä¸¦ Áõ°¡½Ã۰í, ±× °á°ú ´õ È¿À²ÀûÀ̰í Àü·Â ¼Ò¸ð°¡ ÀûÀº ½ºÅ丮Áö ±â¼úÀÎ HMCÀÇ Çʿ伺À» Áõ°¡½Ã۰í ÀÖ½À´Ï´Ù. ¶ÇÇÑ, AI ¸ðµ¨ ÇнÀÀ» À§ÇÑ °íÃâ·Â ÄÄÇ»ÆÃÀÇ Çʿ伺ÀÌ Áõ°¡ÇÔ¿¡ µû¶ó °í¼º´É ÄÄÇ»ÅÍÀÇ ÀαⰡ ³ô¾ÆÁö¸é¼ ½ÃÀå¿¡¼ HMC¿¡ ´ëÇÑ ¼ö¿ä°¡ ±ÞÁõÇϰí ÀÖ½À´Ï´Ù.
Á¦Ç° À¯Çü¿¡ µû¶ó ÇÏÀ̺긮µå ¸Þ¸ð¸® Å¥ºê ½ÃÀåÀº GPU, CPU, APU, FPGA, ASIC·Î ºÐ·ùµË´Ï´Ù. ½ÃÀå¿¡¼ GPU ºÎ¹®ÀÌ HMC¿¡ ´ëÇÑ °¡Àå Å« ¼ö¿ä¸¦ âÃâÇϰí ÀÖ½À´Ï´Ù. ÁÖ·Î Á¦³Ê·¹ÀÌÆ¼ºê AIÀÇ º¸±ÞÀÌ ÀÌ·¯ÇÑ È®ÀåÀ» ÁÖµµÇϰí ÀÖ½À´Ï´Ù. ¶ÇÇÑ, AI ÇÁ·ÎÁ§Æ®¸¦ Áö¿øÇÏ´Â ÄÄÇ»ÆÃ ÆÄ¿ö¿¡ ´ëÇÑ ¾öû³ ¼ö¿ä°¡ ÀÖ´Â °¡¿îµ¥, ¾Æ¸¶Á¸À̳ª ¸ÞŸ(Meta)¿Í °°Àº ´ë±â¾÷ÀÇ Å¬¶ó¿ìµå ÀÎÇÁ¶ó °³¹ßÀÌ ºü¸£°Ô ÁøÇàµÇ°í ÀÖ´Â °Íµµ GPU ¼ö¿ä È®´ë¿¡ À¯¸®ÇÑ È¯°æÀ» Á¶¼ºÇÏ¿© HMCÀÇ ¼ö¿ä È®´ë¿¡ ±â¿©Çϰí ÀÖ½À´Ï´Ù. ¿¹¸¦ µé¾î, 2024³â 1¿ù MetaÀÇ CEO´Â 2024³â ¸»±îÁö ȸ»çÀÇ ÄÄÇ»ÆÃ ÀÎÇÁ¶ó¸¦ °³¹ßÇϱâ À§ÇØ ¿£ºñµð¾ÆÀÇ H100 ±×·¡ÇÈ Ä¨ 35¸¸ °³¸¦ ±¸¸ÅÇϱâ À§ÇØ 10¾ï ´Þ·¯¸¦ ÅõÀÚÇÒ °ÍÀ̶ó°í ¹ßÇ¥Çß½À´Ï´Ù.
¿ëµµ¿¡ µû¶ó ±×·¡ÇÈ, AI, °í¼º´É ÄÄÇ»ÆÃ, ³×Æ®¿öÅ·, µ¥ÀÌÅͼ¾ÅÍ·Î ³ª´µ¸ç, AI¿Í HPC°¡ HMC ¼ö¿äÀÇ ´ëºÎºÐÀ» Â÷ÁöÇϰí ÀÖ½À´Ï´Ù. ÀÌ·¯ÇÑ ¼ö¿äÀÇ ±ÞÁõÀº ÁÖ·Î ÃÖ±Ù »ý¼ºÇü AIÀÇ ±Þ°ÝÇÑ ¼ºÀå¿¡ ±âÀÎÇÕ´Ï´Ù. ¹æ´ëÇÑ µ¥ÀÌÅÍ Ã³¸®¸¦ À§ÇØ »ó´çÇÑ ¾çÀÇ ÄÄÇ»ÆÃ ÆÄ¿ö°¡ ÀÌ·¯ÇÑ ¸ðµ¨À» ÇнÀÇÏ´Â µ¥ »ç¿ëµÇ°í ÀÖ½À´Ï´Ù. ÀÌ·¯ÇÑ º¯È·Î ÀÎÇØ ÀÌ·¯ÇÑ ´ë±Ô¸ð ¾ð¾î ¸ðµ¨(LLM)À» Áö¿øÇÏ´Â µ¥ ÇÊ¿äÇÑ ÄÄÇ»ÆÃ ÆÄ¿ö¸¦ Áö¿øÇÒ ¼ö ÀÖ´Â HMC ¹× ÅëÇÕ GPU¿¡ ´ëÇÑ ¼ö¿ä°¡ ±ÞÁõÇϰí ÀÖ½À´Ï´Ù. ¶ÇÇÑ, Áý¾àÀûÀÎ µ¥ÀÌÅÍ Ã³¸®¸¦ À§ÇØ ´Ù¾çÇÑ »ê¾÷ ºÐ¾ß¿¡¼ °í¼º´É ÄÄÇ»ÆÃÀÇ Ã¤ÅÃÀÌ È®´ëµÊ¿¡ µû¶ó °í¼º´É ¸Þ¸ð¸®¿¡ ´ëÇÑ ¼ö¿ä°¡ ´õ¿í Áõ°¡Çϰí ÀÖ½À´Ï´Ù.
ÇÏÀ̺긮µå ¸Þ¸ð¸® Å¥ºê ½ÃÀåÀº ÃÖÁ¾»ç¿ëÀÚº°·Î ¿£ÅÍÇÁ¶óÀÌÁî ½ºÅ丮Áö, Åë½Å ¹× ³×Æ®¿öÅ·, ÀΰøÁö´É °³¹ßÀÚ, ±âŸ·Î ±¸ºÐµÇ¸ç, HCM ¼ö¿ä¸¦ ÁÖµµÇϰí ÀÖ´Â °ÍÀº ÀΰøÁö´É °³¹ßÀÚÀÔ´Ï´Ù. ÀÌ ¼ö¿äÀÇ ÁÖ¿ä ¿øÀÎÀº ƯÈ÷ ÀΰøÁö´É, ½´ÆÛÄÄÇ»ÆÃ µî µ¥ÀÌÅÍ Áý¾àÇü ¾ÖÇø®ÄÉÀ̼ÇÀÇ ±Þ°ÝÇÑ ¼ºÀå¿¡ ±âÀÎÇÕ´Ï´Ù. ¶ÇÇÑ, ChatGPT¿Í °°Àº ´ë±Ô¸ð AI ¸ðµ¨ÀÇ »ç¿ëÀÌ Áõ°¡ÇÏ¸é¼ HBM1¿¡ ´ëÇÑ ¼ö¿ä°¡ ±ÞÁõÇϰí ÀÖ½À´Ï´Ù. ÀÌ·¯ÇÑ ¸ðµ¨µéÀº ºü¸¥ µ¥ÀÌÅÍ Ã³¸®¿Í µ¥ÀÌÅÍ Àü¼ÛÀ» ÇÊ¿ä·Î Çϴµ¥, ÀÌ´Â ±¤´ë¿ªÆø ¸Þ¸ð¸®¿¡¼¸¸ ±¸ÇöÇÒ ¼ö ÀÖ½À´Ï´Ù. ¶ÇÇÑ, HBMÀº ´ë¿ªÆøÀ» ÇÊ¿ä·Î ÇÏ´Â ¾ÖÇø®ÄÉÀ̼ǿ¡¼ ÀαⰡ ³ô½À´Ï´Ù. ½ºÅÃ´ç ¾à 1.2TB/s·Î ´ë¿ªÆø2 Ãø¸é¿¡¼ HBM3E¸¦ ´É°¡ÇÏ´Â ±âÁ¸ ¸Þ¸ð¸®´Â ¾ø½À´Ï´Ù. ÀÌ·¯ÇÑ ³ôÀº ´ë¿ªÆøÀº AI ¾ÖÇø®ÄÉÀ̼ÇÀÌ È¿À²ÀûÀ¸·Î ÀÛµ¿Çϱâ À§ÇØ ¸Å¿ì Áß¿äÇÕ´Ï´Ù.
ÇÏÀ̺긮µå ¸Þ¸ð¸® Å¥ºêÀÇ ½ÃÀå µµÀÔ¿¡ ´ëÇÑ ÀÌÇØ¸¦ µ½±â À§ÇØ ºÏ¹Ì(¹Ì±¹, ij³ª´Ù, ±âŸ ºÏ¹Ì), À¯·´(µ¶ÀÏ, ¿µ±¹, ÇÁ¶û½º, ½ºÆäÀÎ, ÀÌÅ»¸®¾Æ, ±âŸ À¯·´), ¾Æ½Ã¾ÆÅÂÆò¾ç(Áß±¹, ÀϺ», Àεµ, Àεµ, Çѱ¹, ´ë¸¸, ±âŸ ¾Æ½Ã¾ÆÅÂÆò¾ç), Àü ¼¼°è ±âŸ Áö¿ª¿¡¼ÀÇ ¼¼°è ÀÔÁö¸¦ ±âÁØÀ¸·Î ½ÃÀåÀ» ºÐ¼®Çϰí ÀÖ½À´Ï´Ù. ºÏ¹Ì°¡ °¡Àå Å« Á¡À¯À²À» Â÷ÁöÇϰí ÀÖ½À´Ï´Ù. ºÏ¹Ì ÇÏÀ̺긮µå ¸Þ¸ð¸® Å¥ºê ½ÃÀåÀº ÇâÈÄ ¸î ³â µ¿¾È ±Þ¼ºÀåÇÒ °ÍÀ¸·Î ¿¹»óµË´Ï´Ù. ÀÌ·¯ÇÑ ¼ö¿ä Áõ°¡´Â ƯÈ÷ ÀΰøÁö´É(AI), ½´ÆÛÄÄÇ»ÆÃ µî °í¼Ó µ¥ÀÌÅÍ Ã³¸®°¡ ÇÊ¿äÇÑ µ¥ÀÌÅÍ Áý¾àÀû ¾ÖÇø®ÄÉÀ̼ÇÀÇ ±Þ°ÝÇÑ ¼ºÀå¿¡ ±âÀÎÇϸç, À̸¦ À§Çؼ´Â ±¤´ë¿ªÆø ¸Þ¸ð¸®°¡ ÇÊ¿äÇÕ´Ï´Ù. ¶ÇÇÑ, AI ½ºÅ¸Æ®¾÷¿¡ ´ëÇÑ ÀÚ±Ý Áö¿ø Áõ°¡·Î ÀÎÇØ ÀÌ Áö¿ªÀÇ AI ±â¼ú °³¹ß ¹× Àû¿ë ¼Óµµ°¡ »¡¶óÁö°í ÀÖ½À´Ï´Ù. ¶ÇÇÑ, AI °³¹ß¿¡ ´ëÇÑ ÀÌ Áö¿ªÀÇ À¯¸®ÇÑ Á¤ºÎ Á¤Ã¥µµ »ê¾÷°èÀÇ AI ±â¼ú äÅÃÀ» ÃËÁøÇÏ°í ¼ºÀå¿¡ Å« ±â¿©¸¦ Çϰí ÀÖ½À´Ï´Ù. AI ¿öÅ©·Îµå´Â µð¹ÙÀ̽º¿Í ÇÁ·Î¼¼½Ì À¯´Ö °£ÀÇ µ¥ÀÌÅÍ Àü¼Û ¼Óµµ¸¦ Çâ»ó½Ã۱â À§ÇØ ´õ ³ôÀº ´ë¿ªÆøÀÌ ÇÊ¿äÇϸç, ÀÌ·¯ÇÑ »ý¼ºÇü AIÀÇ ±Þ¼ÓÇÑ ¹ßÀüÀº µ¥ÀÌÅͼ¾ÅÍ ½ÃÀå¿¡¼ °í¼Ó HBM ±â¼ú¿¡ ´ëÇÑ ¼ö¿ä¸¦ Áõ°¡½Ã۰í ÀÖ½À´Ï´Ù. ÇÏÀÌÆÛ½ºÄÉÀÏ·¯¿Í OEM(ÁÖ¹®ÀÚ »óÇ¥ ºÎÂø »ý»ê¾÷ü)Àº ¸ðµ¨ ÇнÀ°ú Ãß·ÐÀ» Áö¿øÇϱâ À§ÇØ ¼¹ö ¿ë·®À» ´Ã¸®¸é¼ ´õ ¸¹Àº AI °¡¼Ó±â¸¦ ÇÊ¿ä·Î Çϰí ÀÖ½À´Ï´Ù. ±× °á°ú, ÀÌ·¯ÇÑ ¾×¼¿·¯·¹ÀÌÅÍ¿Í °ü·ÃµÈ HBMÀÇ °·ÂÇÑ ¼ºÀåÀ» ÃËÁøÇϰí ÀÖ½À´Ï´Ù.
ÀÌ ½ÃÀå¿¡¼ »ç¾÷À» Àü°³ÇÏ´Â ÁÖ¿ä ±â¾÷À¸·Î´Â Micron Technology, Inc., Samsung, SK HYNIX INC., Intel Corporation, NVIDIA Corporation, Global Unichip Corp, Cambricon, Huawei Technologies Co.,, IBM, and Advanced Micro Devices, Inc. µîÀÌ ÀÖ½À´Ï´Ù.The Hybrid Memory Cube (HMC) is a high-performance computer random-access memory (RAM) interface for through-silicon via (TSV)-based stacked DRAM memory. It was co-developed by Samsung Electronics and Micron Technology in 2011. HMC uses standard DRAM cells but it has more data banks than classic DRAM memory of the same size. The memory controller is integrated as a separate die. It promised a 15 times speed improvement over DDR3. Though, Micron discontinued the HMC product in 2018 when it failed to achieve market adoption, however shifting market dynamics towards Artificial Intelligence is boosting the demand for Hybrid memory cubes in the market and encouraging manufacturers to mass produce the HMCs, with the demand exceeding the supply.
The Hybrid Memory Cube Market is expected to grow at a strong CAGR of around 26.50% owing to the growing widespread adoption of Artificial Intelligence applications. Furthermore, the growth can be attributed to the rising proliferation of artificial intelligence. The technology itself is experiencing an unprecedented rate of development and adoption. This shift towards AI increased demand for high-performance GPUs, creating a favorable demand environment for HMCs. For instance, in February 2024, Nvidia, the world's largest producer of GPUs, presented a forecast of a three-fold increase in its quarterly revenue with boosted demand for data center chips and GPUs amid the AI boom the world is witnessing. Furthermore, organizations shifting their operations to the cloud. This inclination towards cloud has increased demand for data centers, consequently driving up the need for HMCs, a more efficient and less power-consuming storage technology. Additionally, the popularity of high-performance computers is also rising due to the increased need for high-power computations to train AI models, further leading to a surge in demand for HMCs in the markets.
Based on product type, the hybrid memory cube market is categorized into GPU, CPU, APU, FPGA, and ASIC. The GPU segment is generating maximum demand for HMCs in the markets. The increasing prevalence of generative AI primarily drives this expansion. Furthermore, the rapid advancement in the development of cloud infrastructure, by large corporations such as Amazon, and Meta, amid the enormous demand for computing power to support AI projects, is creating a favorable environment of growth in demand for GPUs, subsequently benefitting the increasing need for HMCs. For instance, in Jan 2024, Meta CEO announced an investment of a billion dollars in purchasing 350,000 units of Nvidia's H100 graphic chips, for developing its computing infrastructure by the end of 2024.
Based on application the market is segmented into graphics, AI and high-performance computing, networking, and data centers. AI and HPC generate most of the demand for HMC. This surge in demand is primarily driven by the exponential growth witnessed by generative AI in the recent past. A considerable amount of computing power for massive data processing goes into training these models. This shift has led to a surge in demand for GPUs integrated with HMCs, that can match the potential necessary for computing power to support these large language models (LLMs). Furthermore, the growing adoption of high-performance computing in various industrial setups for intensive data processing further elevates the need for high-performance memory.
Based on end-users, the hybrid memory cube market is segmented into enterprise storage, telecommunication and networking, artificial intelligence developers, and others. The AI developers are dominating the demand for HCMs. The demand is primarily driven by the rapid growth of data-intensive applications, particularly in artificial intelligence and supercomputing. Furthermore, the increasing use of large AI models, such as ChatGPT, has led to a surge in demand for HBM1. These models require high-speed data processing and transfer, which can only be achieved with high-bandwidth memories. Additionally, HBM is popular among bandwidth-hungry applications. At around 1.2 TB/s per stack, no conventional memory can beat HBM3E in terms of bandwidth2. This high bandwidth is crucial for the efficient functioning of AI applications.
For a better understanding of the market adoption of Hybrid Memory Cubes, the market is analyzed based on its worldwide presence in countries such as North America (The U.S., Canada, and the Rest of North America), Europe (Germany, The U.K., France, Spain, Italy, Rest of Europe), Asia-Pacific (China, Japan, India, South Korea, Taiwan, Rest of Asia-Pacific), Rest of World. North America holds a prominent share of the market. The North American hybrid memory cube market is poised for rapid growth in the coming years. This rise in demand is primarily driven by the rapid growth of data-intensive applications, particularly in artificial intelligence and supercomputing, which require high-speed data processing, which necessitates high-bandwidth memories. Furthermore, The increased funding for AI startups has accelerated the pace of development and application of AI technologies in the region. Moreover, the favorable government policies of the region regarding AI development have also encouraged the industries to adopt AI technology and have significantly contributed towards the growth. This rapid rise of generative AI has boosted the demand for high-speed HBM technologies in the data center market. AI workloads are driving the need for higher bandwidth to increase data transfer rates between devices and processing units. Hyperscalers and original equipment manufacturers (OEMs) are increasing their server capacity to support model training and inference, requiring more AI accelerators. This is in turn driving strong growth in HBMs associated with these accelerators.
Some of the major players operating in the market include Micron Technology, Inc.; Samsung; SK HYNIX INC.; Intel Corporation; NVIDIA Corporation; Global Unichip Corp; Cambricon; Huawei Technologies Co.,; IBM; and Advanced Micro Devices, Inc.