Overview of Memory Technologies - Information Sys. and Digital Tech. #8

Overview of Semiconductor Business and Ecosystem

  • The Semiconductor Ecosystem Categories

    • Integrated Device Manufacturers (IDMs): Companies that design, manufacture, and sell their own integrated circuit (IC) products (e.g., Samsung Electronics).

    • Fabless Companies: Companies that design and sell hardware and semiconductor chips but do not have a fabrication plant (fab). They outsource the manufacturing (e.g., NVIDIA).

    • Foundries: Companies that manufacture chips for other companies, primarily for fabless firms (e.g., TSMC, Samsung Foundry).

    • Supporting Roles:

    • Design Houses: Assist in the design phase for fabless or IDM firms.

    • IP (Intellectual Property) Companies: Provide pre-designed layouts or functions that other companies license to build their chips.

  • Broad Classification of Semiconductors

    • Analog Semiconductors: These process continuous signals representing real-world physical properties.

    • Signal Processing: Measures signals like temperature, sound, and light.

    • Power Management: Manages power in electronics, including switches and battery management.

    • Examples: DAC7715DAC7715, TVP5154A.

    • Digital Semiconductors: Focus on discrete data processing.

    • Memory: Divided into Volatile (loses data when power is lost) and Non-Volatile (stores data permanently but is generally slower).

    • Logic: Processes data through mathematical and logical operations.

  • Specific Electronic Chip Architectures

    • SoCs (System on Chips): Combine multiple components like CPU, GPU, and memory onto a single chip.

    • CPUs (Central Processing Units): Give directions to other chips and process tasks one at a time.

    • GPUs (Graphics Processing Units): Process many tasks simultaneously; specialized for graphics and AI.

    • ASICs (Application-Specific Integrated Circuits): Designed for one specific purpose, such as AI accelerators.

    • FPGAs (Field Programmable Gate Arrays): Can be reprogrammed by the user after manufacturing.

Samsung Electronics’ Device Solutions (DS) Division

  • Memory Business: Focuses on groundbreaking memory technology to enhance IT experiences (DRAM, NAND Flash, SSDs).

  • System LSI Business: Designs logic semiconductors such as mobile processors, image sensors, and Display Driver ICs (DDI) for mobile, automotive, and IoT applications.

  • Foundry Business: Provides state-of-the-art manufacturing processes including:

    • EUV-based: 5nm5nm to 2nm2nm GAA (Gate-All-Around).

    • FinFET: 14nm14nm, 10nm10nm, 8nm8nm, 5nm5nm, and 4nm4nm.

    • Other: 28nm28nm FD-SOI, 8-inch, and specialty processes.

Historical Milestones of Samsung Memory Business

  • 1970s - 1980s: Formation

    • 1974: Acquisition of Hankook Semiconductor.

    • 1975: Start of IC mass production for LED wristwatches.

    • 1983: Initiation of VLSI business; development of 64Kb64Kb DRAM; establishment of Giheung Campus.

    • 1984: Opening of Line 1 at Giheung; development of 256Kb256Kb DRAM.

  • 1990s: Market Leadership

    • 1992: Development of the industry’s first 64Mb64Mb DRAM; achieved world’s largest DRAM market share.

    • 1993: Opening of the industry’s first 200mm200mm fabrication line (Line 5); achieved largest memory market share.

    • 1998: Production begins at the Austin, Texas (SAS) plant.

  • 2000s - 2010s: Expansion and Innovation

    • 2000: Establishment of Hwaseong Campus.

    • 2003: World’s largest flash memory market share achieved.

    • 2005: Samsung enters the Foundry business.

    • 2006: Industry first 32GB32GB SSDs introduced.

    • 2013: First mass production of 3D Vertical NAND (V-NAND); introduction of Exynos 5 Octa (first mobile processor with big.LITTLE architecture).

    • 2014: Mass production of 20nm20nm-class 8Gb8Gb LPDDR4; Xi’an, China (SCS) plant begins production.

    • 2016: Mass production of 10nm10nm FinFET SoCs and 10nm10nm-class DRAM.

    • 2017: Pyeongtaek plant begins production; Foundry Business unit separates from System LSI.

  • 2020s: The AI Era

    • 2021: Industry’s first HKMG-based DDR5 developed; 200MP200MP image sensor introduced; announcement of 17 billion USD17\,billion\,USD investment in Taylor, USA.

    • 2022: Introduction of 512GB512GB CXL memory; first production of 3nm3nm GAA process.

    • 2023: Development of LPCAMM2 DRAM form factor and GDDR7 DRAM.

    • 2024: Development of 24Gb24Gb GDDR7 DRAM; mass production of 9th-Gen V-NAND.

Global Network and Infrastructure

  • R&D and Design Centers: Cambridge (UK), Munich (Germany), Tel Aviv (Israel), Bengaluru (India), San Jose (USA), Tokyo (Japan), Hangzhou (China).

  • Manufacturing Centers: Giheung, Hwaseong, Pyeongtaek, and Onyang (South Korea); Austin and Taylor (USA); Xi’an, Tianjin, and Suzhou (China).

  • Campus Highlights:

    • Hwaseong Campus: Home to Line 16 and Device Solutions Research (DSR).

    • Giheung Campus: The site where Samsung’s VLSI business began.

    • Silicon Valley Campus: Headquarters for Samsung Semiconductor in the US, featuring onsite café facilities.

    • Economic Impact: Samsung’s Texas fabs provide significant localized economic stimulation.

Memory Hierarchy and Architecture

  • Von Neumann Architecture (1945): Based on the stored-program concept where the logical design of a computer uses a central processing unit and a memory unit for both instructions and data.

  • Memory Hierarchy Concepts:

    • Faster Access, Higher Cost: Located at the top (e.g., Cache, DRAM).

    • Slower Access, Lower Cost: Located at the bottom (e.g., SSD, HDD).

  • Memory Expander (CMM-D):

    • Uses the Compute Express Link (CXL) interface.

    • Configured as a CXL Memory Module-DRAM.

    • Offers higher capacity than main memory but slower processing speeds.

    • Memory Tiering: A technique used to enhance efficiency by organizing memory into different performance layers.

DRAM Product Portfolio and Technologies

  • DRAM (Dynamic Random-Access Memory): Volatile memory used for temporary data storage.

  • DDR (Double Data Rate): Sends data twice during every clock cycle to increase performance.

  • GDDR (Graphic DDR): Designed for high per-pin speeds required for rendering and AI computation.

  • LPDDR (Low-Power DDR): Optimized for energy efficiency; critical for mobile devices, edge devices, and automotive platforms.

  • HBM (High Bandwidth Memory):

    • Integrates TSV-based (Through Silicon Via) stacking for high-throughput.

    • Primarily used in data centers for AI training and High-Performance Computing (HPC).

    • Form Factor: 2.5D structure which reduces latency.

    • Generational Evolution of HBM:

    • HBM2 (2016): 10241024 I/Os, 2 Gbps2\,Gbps, 256 GB/s256\,GB/s bandwidth.

    • HBM2E (2018): 10241024 I/Os, 3.6 Gbps3.6\,Gbps, 460 GB/s460\,GB/s bandwidth.

    • HBM3 (2022): 10241024 I/Os, 6.4 Gbps6.4\,Gbps, 819 GB/s819\,GB/s bandwidth (not backward compatible with HBM2).

    • HBM3E (2023): 10241024 I/Os, 8+ Gbps8+\,Gbps, 1+ TB/s1+\,TB/s bandwidth.

    • HBM4 (Expected 2025): 20482048 I/Os, 6+ Gbps6+\,Gbps, 1.5+ TB/s1.5+\,TB/s bandwidth (not backward compatible with HBM3).

Storage (NAND) Solutions

  • SSD (Solid State Drive): Uses flash memory for storage, offering superior speed, durability, and lower power consumption than traditional HDDs.

    • PC SSD: Focuses on speed and responsiveness for on-device AI and gaming.

    • Datacenter SSD: Optimized for hyperscale servers and sustainable growth via power/thermal balance.

    • Enterprise SSD: Designed for consistent performance and data integrity in databases.

    • Automotive SSD: Optimized for durability and Infotainment systems.

  • eStorage Technologies:

    • UFS (Universal Flash Storage): Enables high-resolution imaging and responsiveness in mobile devices.

    • eMMC (embedded Multi Media Card): A cost-effective, reliable solution for smart appliances and entry-level mobile devices.

  • MCP (Multi-Chip Package):

    • Combines DRAM and NAND into a single compact package to save space in slim devices.

    • uMCP: Integrates DRAM and UFS for high-density applications in smartphones and edge devices.

    • eMCP: Combines DRAM and eMMC for entry-level and mainstream mobile devices.

Strategic Collaborations and Future Trends

  • Collaboration with NVIDIA (GTC 2026):

    • Focus on HBM4 and HBM4E for the AI era.

    • HBM4E features a Core Die Wafer available in 2025.

    • Up to 13 Gbps13\,Gbps I/O speed.

    • Targeting AI factories and massive scale-up bandwidth (640 TB/s640\,TB/s).

  • Collaboration with AMD:

    • Expanding strategic collaboration for next-generation AI memory.

    • Supplying HBM4 for AMD Instinct MI455X GPUs.

    • DDR5 solutions for AMD EPYC processors and the AMD Helios platform.

Work and Life at Samsung Semiconductor

  • Regional Roles: Roles involve marketing (Silicon Valley/Korea), engineering, and R&D.

  • Specialized Roles Featured:

    • Semiconductor Engineer: Transitioning from Chemistry to STEM roles.

    • Logic Innovation Engineer: R&D focus on solving unsolved architectural problems.

    • Process Design Engineer: Managing fabrication line complexities (e.g., an Italian engineer working in the Korean-based Foundry for 10 years).

    • Graphics Engineer: Acting as a global bridge between different manufacturing/design locations (e.g., Bengaluru to global sites).

  • Educational Partnerships: Highlights from the MIT-Samsung Workshop and participation in global tech events like Dellworld.

Questions & Discussion

  • The session concludes with a dedicated Q&A section where management of the floor and presentation follow-up occurs.