Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    High-bandwidth memory (HBM)

    Also known as HBM, HBM3E, HBM4, HBM4E, stacked DRAM

    High-bandwidth memory (HBM) is DRAM built as a vertical stack of dies, linked by thousands of through-silicon vias and placed in the same package as a GPU or other processor, so data reaches the chip over a very wide, short path.[1][2] The current JEDEC standard, HBM4, allows up to 2 TB/s per stack.[3] Counterpoint Research's market-share estimates list three suppliers: SK hynix, Samsung and Micron.[4]

    Editor reviewedUpdated Wiki27 claims · 85% primarySemiconductors, Next-gen computing hardware, Artificial intelligence
    Key facts

    High-bandwidth memory (HBM) is a form of DRAM, the working memory of computers, built as a stack of chips rather than a flat row of them. Several DRAM dies sit on top of each other, often on a base die, connected by thousands of vertical links called through-silicon vias (TSVs) and tiny solder microbumps.[1] HBM is used in AI chips, and the International Energy Agency reported in April 2026 that a shortage had developed.[5]

    This page explains the technology and who makes it. Read the memory wall explainer first for why memory bandwidth limits AI hardware. Speeds quoted from company pages are the makers’ own ratings, and we did no testing.

    How HBM works

    HBM’s TSVs run straight through the silicon, so, as Micron explains, data flows directly through the stack instead of around the edges of chips that sit side by side.[6]

    The other key feature is width. Micron says one of its HBM cubes has a 1,024-bit data interface with 32 independent channels, 16 times wider than a standard DDR5 module.[2] HBM is mounted in the same package as the CPU or GPU, on a silicon interposer, which shortens the distance data has to travel.[2] Because it sits close to the processor and uses shorter connections, Micron says, HBM uses less energy for each bit it moves.[7]

    Packaging is what makes this possible. NVIDIA‘s Blackwell GPU used TSMC’s CoWoS packaging to combine more than three reticles’ worth of silicon with eight HBM stacks.[8] That is an example of the chiplet approach. Hybrid bonding, which joins chips with direct copper connections, is being applied to HBM stacks too.[9]

    Why AI accelerators need it

    AI chips can often calculate faster than memory can feed them. One 2024 analysis found peak server compute grew about 3.0 times every two years, while DRAM bandwidth grew about 1.6 times.[10] The same analysis argued that memory bandwidth, not compute, had become the main bottleneck for AI, especially when serving models.[11] IBM researchers describe the cost of conventional AI inference as the time and energy spent moving model weights between memory and processors.[12]

    Micron describes HBM’s main advantage as keeping processors busy rather than waiting for data.[13] Google’s Ironwood TPU, for example, has 192 GB of HBM per chip at 7.37 TB/s.[14]

    Generations

    JEDEC, the industry standards body, published the HBM4 standard (JESD270-4) on April 16, 2025. It named generative AI, high-performance computing, high-end graphics cards and servers as target uses.[15] The standard allows up to 8 Gb/s per pin across a 2,048-bit interface, for up to 2 TB/s per stack. It doubles the number of channels per stack from 16 to 32, and allows stacks of 4, 8, 12 or 16 dies holding up to 64 GB.[3] HBM4 is backward compatible with HBM3 controllers.[16]

    SK hynix and Micron rate their HBM4 above the standard’s 8 Gb/s.[3][17][18] SK hynix’s product list shows how bandwidth per stack has climbed:[17]

    GenerationSpeed per pinBandwidth per stack
    HBM2E3.2-3.6 Gbps410-460 GB/s
    HBM3up to about 6.4 Gbpsup to about 819 GB/s
    HBM3E9.6 Gbps and higher1.23 TB/s and higher
    HBM411.0 Gbps and higher2.8 TB/s and higher

    Micron likewise rates its HBM4 at more than 11.0 Gbps and more than 2.8 TB/s per stack, more than double its HBM3E.[18] The next step is HBM4E. Samsung says its HBM4E delivers up to 64 GB and 4 TB/s in a single 16-high stack, at up to 16 Gbps per pin.[19]

    Suppliers and supply

    Counterpoint Research’s estimates list three HBM makers. It estimated that in the second quarter of 2026 SK hynix held 50% of the market, Samsung 33% and Micron 18%.[4]

    • SK hynix began mass shipments of HBM4 in the second quarter of 2026 and expects HBM4 to overtake HBM3E in its shipments and revenue in the fourth quarter.[20][21]
    • Samsung began mass production of HBM4 in February 2026, rated at 11.7 Gbps per pin.[22] In July 2026 it said it had shipped the first HBM4E samples to major customers.[23] It significantly scaled up HBM4 sales in the third quarter.[24]
    • Micron had been in volume production of 12-layer HBM4 for NVIDIA’s Rubin since March 2026, according to a supply-chain tracker.[25]

    Supply is tight. The International Energy Agency reported in April 2026 that an HBM shortage had developed over the previous six months and was expected to last at least until the end of 2027.[5] SK hynix said in October 2026 that it expects memory demand to exceed supply through 2030.[26]

    HBM is also subject to US export controls. A BIS rule effective December 2, 2024 added controls on certain high-bandwidth memory.[27] See US AI chip export controls for the rules, and the semiconductor industry tracker for memory makers’ latest results.

    Next in the crash course

    SK hynix

    SK hynix is a South Korean memory chipmaker and the largest supplier of high-bandwidth memory for AI chips. Its history, HBM role and 2026 results.

    More in Semiconductors

    Developing in AI & Computing

    Questions readers ask

    What is high-bandwidth memory?

    HBM is DRAM built as a stack of dies on an optional base die, joined by thousands of through-silicon vias and microbumps. It is placed in the same package as the CPU or GPU on a silicon interposer.[1][2]

    Why do AI chips need HBM?

    AI workloads are often limited by how fast data reaches the processor. One 2024 analysis found compute growing much faster than DRAM bandwidth. Micron says HBM's high throughput keeps processors busy rather than waiting for data.[10][11][13]

    How is HBM4 different from HBM3?

    The JEDEC HBM4 standard uses a 2,048-bit interface at up to 8 Gb/s for up to 2 TB/s per stack, and doubles the channels per stack from 16 to 32. It supports stacks up to 16 dies high and 64 GB.[3]

    Who makes HBM?

    Counterpoint Research's estimates list three suppliers, SK hynix, Samsung and Micron, and put their shares in the second quarter of 2026 at 50%, 33% and 18%.[4]

    Is HBM in short supply?

    The IEA reported in April 2026 that an HBM shortage had developed and was expected to last at least until the end of 2027. SK hynix said in October 2026 that it expects memory demand to exceed supply through 2030.[5][26]

    Sources

    · 27 claims from 15 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      High-bandwidth memory (HBM) is a 3D-stacked DRAM design that stacks several DRAM dies vertically on an optional base die and links them with thousands of through-silicon vias (TSVs) and microbumps. confirmedas of 2026-11-07

    2. [2]

      Micron says one of its HBM cubes has a 1,024-bit data interface with 32 independent channels, 16 times wider than a standard DDR5 module, and that HBM sits in the same package as the CPU or GPU on a silicon interposer. confirmedas of 2026-11-07

    3. [3]

      The JEDEC HBM4 standard allows transfer speeds up to 8 Gb/s across a 2,048-bit interface for up to 2 TB/s per stack, doubles channels per stack from 16 to 32, and supports 4-, 8-, 12- and 16-high stacks of up to 64 GB. confirmedas of 2025-04-16

    4. [4]

      Counterpoint Research estimated that SK hynix held 50% of the HBM market in the second quarter of 2026, down from 64% a year earlier, with Samsung at 33% and Micron at 18%. reportedas of 2026-06-30

    5. [5]

      The IEA reported in April 2026 that a shortage of high-bandwidth memory, used in AI chips, had developed over the previous six months and was expected to last at least until the end of 2027. confirmedas of 2026-04-16

    6. [6]

      Micron explains that because HBM's TSVs run through the silicon, data flows directly through the stack instead of around the edges of chips that sit side by side. confirmedas of 2026-11-07

      • HBM4 · Micron Technology · Primary source · What is HBM4? (retrieved 2026-11-07)
    7. [7]

      Micron says HBM uses less energy per bit transferred because it sits close to the processor and uses shorter connections. confirmedas of 2026-11-07

    8. [8]

      As of 2024, NVIDIA's Blackwell GPU used TSMC's CoWoS packaging to combine more than three reticles' worth of silicon with eight HBM memory stacks. reportedas of 2024-04-30

    9. [9]

      Hybrid bonding is used in AMD processors and AI accelerators, in camera image sensors, and is being applied to high-bandwidth memory stacks. reportedas of 2024-08-11

    10. [10]

      A 2024 analysis found peak server hardware FLOPS grew about 3.0x every two years, while DRAM bandwidth grew about 1.6x and interconnect bandwidth about 1.4x over the same period. confirmedas of 2024-03-21

      • AI and Memory Wall · arXiv (published in IEEE Micro) · Primary source · 2024-03-21 · Abstract (retrieved 2026-10-10)
    11. [11]

      The same analysis argues that memory bandwidth, rather than compute, has become the primary bottleneck for AI workloads, especially when serving models. confirmedas of 2024-03-21

      • AI and Memory Wall · arXiv (published in IEEE Micro) · Primary source · 2024-03-21 · Abstract (retrieved 2026-10-10)
    12. [12]

      IBM researchers describe the cost of conventional AI inference as the time and energy spent moving model weights between memory and processors. confirmedas of 2025-01-29

    13. [13]

      Micron describes HBM's main advantage as moving very large amounts of data quickly, which keeps processors busy rather than waiting for data. confirmedas of 2026-11-07

    14. [14]

      Google said Ironwood scales to pods of 9,216 liquid-cooled chips delivering 42.5 exaflops, with 192 GB of HBM per chip at 7.37 TB/s, and twice the performance per watt of Trillium, its sixth-generation TPU. confirmedas of 2025-04-09

    15. [15]

      JEDEC published the HBM4 standard, JESD270-4, on April 16, 2025, aimed at uses including generative AI, high-performance computing, high-end graphics cards and servers. confirmedas of 2025-04-16

    16. [16]

      JEDEC says the HBM4 interface is backward compatible with HBM3 controllers, so one controller can work with both. confirmedas of 2025-04-16

    17. [17]

      SK hynix lists HBM2E at 3.2-3.6 Gbps (410-460 GB/s), HBM3 at up to about 6.4 Gbps (about 819 GB/s), HBM3E at 9.6 Gbps and higher (1.23 TB/s and higher) in 8- and 12-high stacks, and HBM4 at 11.0 Gbps and higher (2.8 TB/s and higher) in 8-, 12- and 16-high stacks of 24-48 GB. confirmedas of 2026-11-07

      • HBM products · SK hynix · Primary source (retrieved 2026-11-07)
      • HBM products · SK hynix · Primary source · HBM3E product card (text continues "d higher)") (retrieved 2026-11-07)
      • HBM products · SK hynix · Primary source (retrieved 2026-11-07)
      • HBM products · SK hynix · Primary source (retrieved 2026-11-07)
    18. [18]

      Micron says its HBM4 uses a 2,048-pin interface running above 11.0 Gbps for more than 2.8 TB/s per stack, more than double its HBM3E, which delivers more than 1.2 TB/s per stack. confirmedas of 2026-11-07

      • HBM4 · Micron Technology · Primary source (retrieved 2026-11-07)
      • High-bandwidth memory (HBM) · Micron Technology · Primary source (retrieved 2026-11-07)
    19. [19]

      Samsung says its HBM4E delivers up to 64 GB and 4 TB/s in a single 16-high stack, with I/O speeds up to 16 Gbps per pin, more than 20% faster than HBM4. confirmedas of 2026-11-07

      • HBM | DRAM · Samsung Semiconductor · Primary source · HBM4E (retrieved 2026-11-07)
    20. [20]

      SK hynix said it began mass shipments of HBM4 in the second quarter of 2026 and completed HBM4E sample shipments in the first half of the year. confirmedas of 2026-07-29

    21. [21]

      SK hynix said HBM4 shipments began in earnest in the third quarter of 2026, that it would ramp its M15X fab to maximum capacity by year-end to make more HBM4, and that it expected HBM4 to surpass HBM3E in shipments and revenue in the fourth quarter. confirmedas of 2026-10-28· forecast

    22. [22]

      On February 12, 2026 Samsung said it had begun mass production of HBM4 and shipped commercial products to customers, built on its sixth-generation 10nm-class (1c) DRAM process with a 4nm logic base die, and delivers 11.7 Gbps per pin, about 46% above the 8 Gbps industry standard. confirmedas of 2026-02-12

    23. [23]

      Samsung said in July 2026 that it had scaled up HBM4 sales and shipped the industry's first HBM4E samples to major customers. confirmedas of 2026-07-30

    24. [24]

      Samsung's Device Solutions division posted third-quarter 2026 revenue of KRW 147.1 trillion and operating profit of KRW 106.5 trillion, and its Memory Business significantly scaled up HBM4 sales. confirmedas of 2026-09-30

    25. [25]

      Supply-chain tracker Silicon Analysts reported in September 2026 that SK hynix began mass-producing a 48GB, 16-layer HBM4 stack in the third quarter, while Micron had been in volume production of 12-layer HBM4 for NVIDIA's Rubin since March. reportedas of 2026-09-15

    26. [26]

      SK hynix said in October 2026 that it expects memory demand to exceed supply through 2030, citing the difficulty of leading-edge processes and long lead times for new fabs. confirmedas of 2026-10-28· forecast

    27. [27]

      A BIS rule effective December 2, 2024 added controls on certain semiconductor manufacturing equipment and high-bandwidth memory, and created new foreign direct product rules aimed at advanced-node chip production. confirmedas of 2024-12-05

    Revision history (1)
    1. Page created.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Mar 1, 2027.

    Cite this page

    "High-bandwidth memory (HBM)." ContentLora, updated Nov 7, 2026. https://contentlora.com/wiki/high-bandwidth-memory

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.