High Bandwidth Memory
Stacked memory interface for high-performance computing and graphics.
High Bandwidth Memory (HBM) is a standard for 3D-stacked dynamic random-access memory. Developed by AMD, Samsung and SK Hynix, it offers wide data paths and low power use. It serves graphics processors, AI accelerators and supercomputers where speed matters most.
- Stacks up to 32 DRAM dies using through-silicon vias.
- Offers up to 2 TB/s bandwidth in HBM4 standard.
- Used in Nvidia H100 and AMD Fiji GPUs.
- Demand from AI sector drives global memory prices.
- JEDEC standardized the interface in October 2013.
What it is
High Bandwidth Memory (HBM) is a type of computer memory that stacks multiple DRAM dies vertically. This design creates a short, wide path for data to travel between the memory and the processor. It was created to handle the heavy data loads of modern graphics cards and AI chips. Unlike traditional memory sticks, HBM sits very close to the processing unit, often on the same package. This proximity reduces delay and saves energy. The technology supports massive bandwidth while keeping the physical footprint small. It is now a key component in high-end servers and supercomputers.
How it came about
AMD started developing this memory format in 2008 to cut power use and size. They worked with SK Hynix, which had experience in stacking chips, and other partners like UMC and Amkor. The first standard arrived in 2013 after work by JEDEC. SK Hynix made the first chips in 2013. AMD launched the first consumer products using it in 2015 with their Fiji GPUs. Since then, the standard has evolved through several generations, with JEDEC updating specs for HBM2, HBM3 and HBM4 to meet growing demand from artificial intelligence applications.
Inside the chip
The core idea is vertical stacking. Up to 32 DRAM layers sit on top of a base die. Tiny holes called through-silicon vias connect these layers electrically. This creates a very wide data bus, often 1024 bits or more per stack. In contrast, standard memory uses narrow buses with many channels. HBM splits its interface into independent channels that do not need to sync with each other. This allows high-speed data transfer at lower voltages. The memory connects to the GPU or CPU via a silicon interposer, which acts as a bridge. This setup keeps signal paths short but adds manufacturing cost compared to standard circuit boards.
How it was made
Manufacturing involves stacking thin DRAM dies, each about 30 micrometres thick. TSMC produces the base die for many of these stacks. Major producers include SK Hynix, Samsung and Micron. The process uses advanced packaging techniques to align the layers precisely. As demand from the AI sector surged in 2025 and 2026, production capacity shifted toward HBM. This shift reduced the supply of standard DDR5 memory, causing prices to rise. Makers are now moving to taller stacks, with 12 and 16 layers becoming common for higher capacities.
Where it went
This memory powers high-performance graphics processors, such as Nvidia’s H100 and Blackwell series. It is also found in FPGAs, network devices and some specialized CPUs like the Fujitsu A64FX. Data centres rely on it for AI training and inference tasks because of its speed. Supercomputers use it to process vast datasets quickly. Consumer applications include virtual reality headsets and high-end gaming PCs. The technology is essential wherever large amounts of data must move rapidly between storage and processing units without bottlenecks.
Versions and family
The first version, HBM1, offered 128 GB/s bandwidth. HBM2 doubled the speed and allowed larger stacks. HBM2E increased this further to over 400 GB/s. HBM3 brought speeds up to 819 GB/s per stack and introduced square chip shapes. HBM3E pushed rates past 1 TB/s. The latest standard, HBM4, doubles the interface width to 2048 bits, allowing up to 2 TB/s bandwidth and 64 GB capacity per stack. A variant called SPHBM4 uses a narrower external interface to avoid expensive interposers. Future versions like HBM5 are already in development.
Why it matters
HBM has become critical for the AI boom. Its ability to move data quickly enables larger machine learning models. The surge in demand has reshaped the memory market, squeezing out standard DRAM production. This has led to significant price increases for general-purpose memory. The technology has forced manufacturers to adopt new packaging methods and invest heavily in 3D stacking capabilities. It has also driven innovation in interposer design and thermal management for dense electronic packages.
Problems and reception
The industry has widely adopted the standard due to its performance benefits. However, the high cost of production and the need for silicon interposers remain challenges. Some critics noted that early HBM3 designs required retrofitting before the final JEDEC standard was ready. Despite this, major players like Nvidia and AMD have integrated it into their flagship products. The transition to HBM4 aims to simplify some aspects by offering backward compatibility and alternative packaging options.




Questions people ask about High Bandwidth Memory
What makes HBM different from standard RAM?
HBM stacks memory chips vertically instead of placing them side by side on a board. This creates a much wider data path and shorter connections, allowing faster transfer speeds and lower power use compared to DDR4 or GDDR5 memory.
Who developed High Bandwidth Memory?
AMD initiated the development in 2008 to solve power and size issues. They partnered with SK Hynix, Samsung and other industry players. JEDEC later adopted it as an official standard in 2013.
Why is HBM important for AI?
AI models require moving massive amounts of data quickly. HBM provides the high bandwidth needed for these tasks, enabling faster training and inference. Its efficiency helps manage the power demands of large-scale AI processors.
How is HBM manufactured?
Manufacturers stack thin DRAM dies on top of a base die. They connect these layers using tiny vertical wires called through-silicon vias. TSMC often makes the base die, while companies like SK Hynix and Samsung assemble the final stacks.
Good reading and sources
- NVIDIA Data Centers for the Era of AI ReasoningNvidia, Article
- Checking your browserAnandTech, Article
- HBM3: Cheaper, up to 64GB on-package, and terabytes-per-second bandwidthArs Technica, Article
The facts on this page come from the Wikipedia article "High Bandwidth Memory", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryChip rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.