What is high-bandwidth memory (HBM)?
High-bandwidth memory (HBM) is a type of DRAM made by stacking thinned memory chips into one unit. The unit sits beside a GPU or other AI processor in the same package. This gives the processor a very wide, fast path to its data.
Also known as: HBM, High Bandwidth Memory
Researched and fact-checked by AI, with no human review. 8 sources listed below. How we verify
Last updated
How it works
HBM takes DRAM, the working memory of a computer, and builds it upward. As IEEE Spectrum described it in February 2026, an HBM unit holds as many as 12 thinned memory chips, called dies, in a single stack. Vertical connections called through-silicon vias (TSVs) run through each die. Microscopic solder balls join the layers. The stack rests on a base die that passes data between memory and processor. The finished unit sits within a millimeter of a GPU or other AI accelerator and is packaged with it.
JEDEC, a microelectronics standards body, defines the interface. Its HBM4 standard, published on April 16, 2025, specifies transfer speeds of up to 8 gigabits per second across a 2,048-bit interface. Total bandwidth is up to 2 terabytes per second. The standard doubles the independent channels per stack from 16 in HBM3 to 32. It allows stacks of 4, 8, 12 or 16 dies holding up to 64 gigabytes. A revised version, JESD270-4A, followed in December 2025.
Why it matters
IEEE Spectrum says HBM targets the memory wall: the time and energy needed to feed a processor the terabytes of data per second that large language models require. It reported that Nvidia's B300 uses eight HBM stacks of 12 dies each. It cited a SemiAnalysis estimate that HBM costs about three times as much as other memory and makes up half or more of a packaged GPU's cost. It also named HBM as the main cause of a wider DRAM shortage.
Where things stand in 2026
Three companies dominate supply. Counterpoint Research figures cited by Reuters put SK Hynix at 57% of the market in the fourth quarter of 2025, Samsung at 22% and Micron at 21%. Samsung announced on February 12, 2026 that it had begun mass production of HBM4 and shipped it to customers. It claimed a speed of 11.7 gigabits per second. Reuters reported that SK Hynix had said in January its HBM4 was in volume production. Citing media reports, Reuters said Micron had said the same. In May 2026 Samsung said it had started shipping samples of a faster successor, HBM4E.
In July 2026, analysts cited by Reuters linked SK Hynix's missed profit forecasts to slower-than-expected HBM4 shipments. In September 2026 a Samsung executive said HBM uses about 20% of DRAM makers' wafer capacity and is expected to reach nearly 30% in 2027.
Sources
- JEDEC® and Industry Leaders Collaborate to Release JESD270-4 HBM4 Standard: Advancing Bandwidth, Efficiency, and Capacity for AI and HPC, JEDEC Solid State Technology Association
- High Bandwidth Memory (HBM4) DRAM, JEDEC Solid State Technology Association
- How and When the Memory Chip Shortage Will End, IEEE Spectrum
- Samsung Ships Industry-First Commercial HBM4 With Ultimate Performance for AI Computing, Samsung Electronics (Samsung Global Newsroom)
- Samsung ships latest HBM4 chips to catch-up in AI race, Reuters
- Samsung Electronics ships faster HBM4E chip samples to customers; shares jump, Reuters
- SK Hynix's record profit misses forecasts, shares slump 10% despite robust AI chip demand, Reuters
- Samsung Electronics says HBM to account for nearly 30% of industry DRAM capacity next year, Reuters
