What Happened to High-Bandwidth Memory (HBM)?
High-Bandwidth Memory (HBM) is a 3D-stacked synchronous dynamic random-access memory (SDRAM) technology that has revolutionized high-performance computing by offering significantly higher bandwidth and lower power consumption than traditional DRAM. Initially developed to meet the demands of graphics processors, HBM has become critical for AI accelerators and data centers, driving a 'memory supercycle' with unprecedented demand and continuous generational advancements up to HBM4 and beyond.
Quick Answer
High-Bandwidth Memory (HBM) has evolved into a cornerstone technology for artificial intelligence (AI), high-performance computing (HPC), and data centers due to its superior bandwidth and power efficiency. As of late 2026, HBM3E remains the dominant standard in mass production, while HBM4 is rapidly ramping up, with major manufacturers like SK Hynix, Samsung, and Micron having sold out their HBM capacity through the year. The insatiable demand from the AI sector has led to a 'memory supercycle,' significantly impacting the broader DRAM market and accelerating innovation in advanced packaging.
📊Key Facts
📅Complete Timeline15 events
AMD Initiates HBM Development
AMD begins internal development of HBM to address power consumption issues in high-performance graphics.
First HBM Chip & JEDEC Standard
SK Hynix produces the first HBM memory chip, and JEDEC officially adopts HBM as an industry standard (JESD235).
AMD Launches First HBM-Powered GPUs
AMD releases the Radeon R9 Fury X, the world's first graphics card to utilize HBM1, demonstrating significant bandwidth and power efficiency improvements.
JEDEC Accepts HBM2 Standard
JEDEC accepts the second generation, HBM2, as an industry standard, doubling pin transfer rates and supporting up to 8GB per package.
Samsung Announces HBM2E 'Flashbolt'
Samsung announces its Flashbolt HBM2E, featuring eight dies per stack, 16GB capacity, and 410 GB/s per stack, targeting HPC and AI/ML workloads.
SK Hynix Begins HBM2E Mass Production
SK Hynix announces the full-scale mass production of its high-speed HBM2E DRAM, offering 16GB capacity and over 460GB/s bandwidth for AI and HPC.
Samsung Develops HBM-PIM
Samsung announces the development of HBM with processing-in-memory (PIM), integrating AI computing capabilities directly into the memory.
JEDEC Announces HBM3 Standard
JEDEC officially announces the HBM3 standard (JESD238), offering significant improvements in density, speed, bank count, and power efficiency over HBM2E.
SK Hynix Unveils HBM3E
SK Hynix unveils its HBM3E memory, boasting 8 Gbit/s/pin data processing speed and targeting production in the first half of 2024.
Micron Begins HBM3E Mass Production for NVIDIA H200
Micron announces mass production of its HBM3E memory, which consumes 30% less power than competitors and will be utilized in NVIDIA's H200 Tensor Core GPUs.
SK Hynix Mass Produces World's First 12-Layer HBM3E
SK Hynix begins mass production of the world's first 12-layer HBM3E product with 36GB capacity, increasing capacity by 50% at the same thickness as 8-layer products.
JEDEC Releases HBM4 Standard
JEDEC releases the official HBM4 specification, supporting transfer speeds up to 8 Gb/s across a 2048-bit interface, with total bandwidth up to 2 TB/s.
HBM-Led Memory Supercycle Dominates Outlook
SK Hynix's 2026 market outlook highlights a 'memory supercycle' driven by HBM demand, with HBM3E as the flagship and HBM4 gradually increasing its share.
SK Hynix Details HBM Packaging at Hot Chips 2026
SK Hynix presents on HBM packaging at Hot Chips 2026, discussing 3D stacking, bonding, and system integration, noting HBM4 targets over 2 TB/s bandwidth and 48GB capacity.
Micron Targets 2x HBM Capacity by End-2026
Micron reportedly prepares to double its HBM capacity by the end of 2026, aiming for 100,000 wafers a month and increasing 12-high HBM4 output for NVIDIA's Vera Rubin AI platform.
Follow this story
Get an email when this timeline gets a major update.
🔍Deep Dive Analysis
High-Bandwidth Memory (HBM) emerged from a critical need to overcome the limitations of traditional GDDR memory, particularly concerning bandwidth and power consumption in performance-intensive applications like graphics processing. The concept was initially driven by AMD in 2008 to address severe power consumption issues, leading to the development of a 3D-stacked memory architecture. SK Hynix produced the first HBM memory chip in 2013, and the technology was adopted as an industry standard by JEDEC in October 2013.
The first commercial adoption of HBM came with AMD's Fiji GPUs in June 2015, showcasing its potential for unprecedented memory bandwidth while significantly reducing power consumption. This marked a key turning point, demonstrating the viability of stacking multiple DRAM dies vertically and connecting them with Through-Silicon Vias (TSVs) and microbumps, a fundamental departure from planar memory designs. Subsequent generations, including HBM2 (standardized in 2016), HBM2E (announced in 2019, mass-produced in 2020), and HBM3 (standardized in January 2022), progressively increased bandwidth, capacity, and power efficiency, solidifying HBM's role in HPC and data centers.
The most significant catalyst for HBM's explosive growth has been the rise of artificial intelligence, particularly large language models (LLMs) and deep learning. AI accelerators, such as NVIDIA's GPUs and custom ASICs from hyperscalers like Google, require immense memory bandwidth to process vast datasets efficiently. This demand has propelled HBM into a 'memory supercycle' since 2024, with market forecasts suggesting the HBM market size in 2028 could surpass the entire DRAM market of 2024.
As of October 2026, HBM3E is the current production-grade architecture, delivering over 1.2 TB/s per stack and widely deployed in platforms like NVIDIA's H200 and AMD's MI300 series. However, the industry is rapidly transitioning to HBM4, which was officially standardized by JEDEC in April 2025. HBM4 doubles the interface width to 2048 bits, targeting over 2.0 TB/s per stack and is expected to reach volume production in late 2025 to 2026, powering next-generation AI platforms like NVIDIA's Vera Rubin.
The intense demand has led to a highly constrained supply chain, particularly for advanced packaging technologies like TSMC's CoWoS, which is essential for integrating HBM with host chips. This bottleneck is expected to persist through 2026 and potentially into 2027-2028. Major HBM manufacturers—SK Hynix, Samsung, and Micron—have reported their HBM capacity as sold out through 2026, securing long-term agreements with key customers. SK Hynix currently holds a dominant market share, with Micron reportedly overtaking Samsung for second place in some allocations.
Looking ahead, research and development continue at a rapid pace. Samsung announced zHBM (Z-axis HBM) in August 2026, and JEDEC published the Standard Package High Bandwidth Memory (SPHBM4) specification in July 2026, which retains HBM4 DRAM stacks but uses a 512-bit external interface with higher transfer rates. The industry is also exploring concepts like HBM-PIM (processing-in-memory) and custom HBM (cHBM) to further integrate compute capabilities directly into memory, pushing the boundaries of AI hardware.
What If...?
Explore alternate histories. What if High-Bandwidth Memory (HBM) made different choices?