SK hynix HBM4 12-layer

What was announced

On 19 March 2025, SK hynix said it had become the first company to supply 12-layer HBM4 samples to major customers.^[1] A sample shipment is the step before mass production: customers receive working parts and run their own validation on them, and the memory maker uses what it learns to get ready for volume manufacturing.

HBM (high bandwidth memory) is a family of memory built by stacking several DRAM dies on top of one another and connecting them vertically, instead of placing them side by side on a board. The stack sits close to the processor, so data travels a short distance over a very wide connection. The "12-layer" in the name refers to the number of DRAM dies in the stack.

Performance

According to SK hynix, the samples offer the world's highest bandwidth and capacity among 12-layer HBM products.^[1] The part delivers more than 2 TB of data per second, which the company says is more than 60% faster than HBM3E.^[1] To make the figure concrete, SK hynix compares it to processing more than 400 full-HD films of 5 GB each in a single second.^[1] Capacity reaches 36 GB, described as the maximum available for a 12-layer HBM stack at the time.^[1]

Both numbers matter for the accelerators that train and run large AI models. Bandwidth sets how quickly data can be fed to the compute units, and capacity sets how much of a model, or how much working data, can be held next to them without going back to slower memory elsewhere in the system. Each HBM generation is judged largely on how far it moves those two figures.

How the stack is built

Packing twelve dies into one package means each layer has to be thin and the whole stack has to stay flat, which brings two problems: the stack tends to warp, and heat has more material to travel through on its way out. SK hynix used its Advanced MR-MUF process to control die warpage and improve heat dissipation in the stack.^[1] In other words, the process is the part of the manufacturing flow that keeps a tall, thin stack mechanically stable and lets it shed the heat that builds up inside it.

Where it sits in the roadmap

SK hynix said it moved from HBM3 in 2022 to mass production of both 8-layer and 12-layer HBM3E during 2024.^[1] The 12-layer HBM4 samples went out ahead of the original schedule, and the company said it had started customer validation, with mass production preparation to be completed in the second half of 2025.^[1]

The gap between a sample and a shipped product is where most of the work sits. Customers need time to test a new memory generation with their own accelerators and packaging, and the memory maker needs time to bring yields and test flows up. Announcing samples first is therefore a claim about schedule as much as about silicon.

Public showing

At HPE Discover 2025, SK hynix exhibited both 12-layer HBM4 and 12-layer HBM3E, highlighting HBM4's ability to process more than 2 TB of data per second.^[2] Showing the two generations side by side puts the comparison the company is making into a single display: the newer stack against the one it is meant to replace.

Why it matters

AI hardware has run into a memory problem as much as a compute problem. Processors have grown fast enough that the limit on many workloads is how quickly the right data can be brought to them, not how many arithmetic operations they can perform. A memory stack that moves more than 2 TB per second and holds 36 GB in one package is aimed directly at that limit.^[1]

For SK hynix, the announcement is also a position statement. The company presents its HBM line as a continuous sequence — HBM3 in 2022, 8-layer and 12-layer HBM3E in mass production during 2024, and now 12-layer HBM4 in customers' hands — and it wants the next generation associated with its name before volume shipments begin.^[1] Whether that lead holds depends on validation results and on how quickly mass production preparation is finished in the second half of 2025.^[1]