1. The main categories can be divided into two major camps: NAND Flash and DRAM1. NAND Flash manufacturers These chips are used in devices that require long-term data storage, such as SSDs, USB drives, SD cards, and mobile storage. Currently, the global market is mainly dominated by the following companies, often referred to as the “Big Six Flash Manufacturers”: First Tier (absolute leaders):· Samsung – South Korea · The world’s largest NAND flash manufacturer, leading in technology and consistently holding the top market share. · Brands: Samsung’s own SSDs, mobile phones, and other products use self-produced chips.· Kioxia – Japan · Formerly Toshiba Memory, the inventor of flash memory globally. · Brands: Kioxia (Toshiba) SSDs, memory cards, etc.· SK Hynix – South Korea · A leading global memory manufacturer, strong in both DRAM and NAND fields. · Brands: SK Hynix SSDs (such as the brand “Solidigm”) and chips provided to numerous mobile and computer manufacturers.· Western Digital – USA · A long-term deep partner of Kioxia, jointly building factories and developing and producing flash memory chips. · Brands: Western Digital and SanDisk SSDs, memory cards, etc., use chips produced by their joint venture with Kioxia.· Micron – USA · A major global memory supplier with strong technical capabilities. · Brands: Crucial is its consumer brand, using Micron’s self-produced chips, while also providing chips to numerous industry clients. Second Tier (focusing on specific fields):· Yangtze Memory Technologies – China · The largest NAND flash manufacturer in China, rapidly developing and entering the global market through innovative Xtacking technology. · Brands: ZhiTi is its consumer brand, and many domestic SSD brands also use Yangtze Memory’s chips.2. DRAM manufacturers These chips are used for operating memory in computers, mobile phones, etc., characterized by high speed but data loss upon power failure. This market is even more concentrated, essentially monopolized by three giants, often referred to as the “Three Giants of DRAM”:· Samsung – South Korea · The leader in the global DRAM market, with leading market share and technology.· SK Hynix – South Korea · Comparable in strength to Samsung, another major giant in the DRAM market.· Micron – USA · One of the three giants, with stable market share and leading technology. · Brands: Crucial memory modules use Micron’s self-produced chips. Other participants:· ChangXin Memory – China · A major DRAM chip manufacturer in China, striving to expand capacity and catch up technologically, representing China’s hope for self-sufficiency in the DRAM field.Important Concepts: Original Manufacturers vs. Module Manufacturers· Original Manufacturers: Refers to the companies listed above, which design and manufacture memory chips themselves.· Module Manufacturers: Purchase chips from original manufacturers and assemble them into memory modules, SSDs, and other finished products for sale. For example, well-known brands like Kingston, ADATA, and Corsair are module manufacturers (although Kingston also has some self-testing capabilities).2. The following are several core memory chip types used in AI, sorted by their roles in AI workflows:1. Core King: GPU Memory – HBM This is the absolute main player in high-end AI training and inference today.· High Bandwidth Memory. It is not a single memory chip but a 3D stacking technology that stacks multiple DRAM chips like building blocks and packages them together with GPU/ASIC chips through a silicon interposer. · Extreme Bandwidth: Traditional GDDR memory uses “flat” communication, while HBM uses “3D” communication, possessing extremely wide data channels (over 1024 bits), providing several times the bandwidth of GDDR (currently HBM3E can reach over 1TB/s). AI models (especially large language models) have huge parameters and require massive data to be fed to the computing core instantly; high bandwidth is a critical bottleneck. · High Density: Can achieve large capacity in a small physical space (currently a single stack can reach 24GB or even 36GB). · Low Power Consumption: Compared to GDDR solutions providing the same bandwidth, HBM consumes less power. · Major Producers: SK Hynix (technologically and market-leading), Samsung, Micron. · Application Scenarios: NVIDIA H100/A100, AMD MI300 series, Google TPU, and all high-end AI accelerator cards.2. Important Cornerstone: GPU Memory – GDDR6/GDDR6X This is the backbone of mainstream GPU memory.· Overview: Graphics Double Data Rate Synchronous Dynamic Random-Access Memory. Although designed for graphics, its high bandwidth characteristics are also very suitable for AI. · Why Suitable for AI: · Balanced Cost and Performance: Compared to HBM, GDDR is cheaper, with a more mature manufacturing process, yet still provides very high bandwidth (around 1TB/s). · Wide Application: It is the main memory for consumer-grade and some data center-level AI graphics cards. · Major Producers: Samsung, SK Hynix, Micron. · Application Scenarios: NVIDIA GeForce RTX 4090/3090 (gaming cards also used for AI inference and fine-tuning), NVIDIA L40S, AMD Radeon Instinct series, etc.3. Traditional Mainstay: Server Memory – DDR This is the memory in the CPU system. · Overview: The most familiar type of memory, mainly DDR4 or DDR5 in AI servers. · Accommodating the Entire System: AI training/inference tasks are not solely dependent on the GPU. The CPU is responsible for data preprocessing, task scheduling, model management, etc. Massive training data, operating systems, and applications themselves are stored in DDR memory. · Large Capacity Demand: For models with billions or even trillions of parameters, even if GPU memory cannot accommodate them, large system memory is needed as a cache and swap area. AI servers’ system memory often reaches 1TB or even several TB. · Major Producers: Samsung, SK Hynix, Micron. · Application Scenarios: All AI servers and workstations.3. Looking to the Future· The future trend is heterogeneous memory architecture: HBM as the “close high-speed cache” for GPUs, GDDR as the main memory, DDR as system memory, and CXL providing scalable massive memory pools, working together to support increasingly larger and more complex AI models.