Pull to refresh
Logo
Samsung stacks memory directly on AI accelerators with zHBM prototype

Samsung stacks memory directly on AI accelerators with zHBM prototype

New Capabilities

The 3D architecture removes the interposer, targeting 8x performance and 3x power efficiency over HBM5

3 days ago: Samsung debuts zHBM prototype

Overview

Updated 1 hour ago

Samsung has shown a working prototype of zHBM, a memory architecture that stacks high-bandwidth memory directly on top of AI accelerators. The design removes the interposer, the silicon bridge that currently carries data between logic and memory chips placed side by side.

Samsung says the vertical layout shortens data travel from millimeters to tens of micrometers, targeting 8x the performance of HBM5 with 3x the power efficiency. The tradeoff is heat. A 700-1000 watt accelerator sits beneath the DRAM stack, and rising temperatures shorten how long DRAM can hold data. Samsung targets four-high stacks, versus the 12-16 high configurations used today.

Why it matters

If zHBM works, it could decide who leads the AI memory market, a position SK hynix currently holds.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

8x
Performance target vs HBM5
Samsung's official target for zHBM interface performance compared with HBM5.
3x
Energy efficiency target vs HBM5
Samsung says zHBM's wafer bonding delivers threefold energy efficiency over HBM5.
70%
I/O power reduction vs HBM4E
Samsung projects zHBM cuts I/O power by about 70% compared with HBM4E.
6 µm
Hybrid bonding pitch
zHBM requires copper-to-copper bonding at pitches of 6 micrometers or less.
4-high
Target DRAM stack height
Samsung targets four-high zHBM stacks because of heat, versus 12-16 high today.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Sign up to generate historical perspectives on this story.

Play

Exploring all sides of a story is often best achieved with Play.

Most of these play right now — no account needed. Sign up to save scores, keep a streak, and unlock Debate and Predict. Log in Sign Up
Predict 3 ways this could play out. Back the one you believe — contrarian picks score more when a scenario has a resolution date. Log in to play

People Involved

Organizations Involved

Timeline

February 2026 September 2026

6 events Latest: 3 days ago
Tap a bar to jump to that date
  1. Samsung debuts zHBM prototype

    Latest Product

    Samsung shows a working zHBM prototype, stacking memory directly on AI accelerators, at proof-of-concept stage.

  2. Samsung showcases zHBM at SEMICON Taiwan

    Announcement

    Samsung presents the CUBE strategy and zHBM targets: 8x HBM4E performance, 3x power efficiency, 75-90% lower thermal resistance.

  3. Hot Chips 2026 details three-phase HBM roadmap

    Presentation

    Samsung's Sangwook Han presents the roadmap from HBM4 base die to zHBM, eliminating the interposer in phase three.

  4. Samsung unveils zHBM concept at FMS 2026

    Announcement

    Samsung shows the first zHBM concept model at the Future of Memory and Storage conference, alongside zNAND-O.

  5. Samsung samples HBM4E to customers

    Product

    Samsung ships HBM4E samples, with mass production targeted for the following year.

  6. Samsung begins HBM4 mass production

    Product

    Samsung starts shipping HBM4, its current high-bandwidth memory product, built on a 4nm logic base die.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

August 2013

Samsung V-NAND (2013)

Samsung introduced V-NAND, the first commercial 3D NAND flash memory, stacking cells vertically instead of shrinking them on a flat plane. The industry had hit scaling limits with planar NAND, and Samsung bet on vertical stacking to keep density growing.

Then

Samsung became the NAND flash leader and held that position for years as rivals scrambled to catch up with 3D stacking.

Now

3D NAND became the industry standard, and Samsung's stacking expertise became a core competitive advantage.

Why this matters now

Samsung is making the same bet with zHBM: vertical stacking to solve a scaling and performance problem. The company explicitly cites its V-NAND and through-silicon via experience as the foundation for zHBM's wafer bonding.

2013-2015

SK hynix HBM debut (2013)

SK hynix developed high-bandwidth memory with AMD, stacking DRAM dies vertically and connecting them with through-silicon vias. HBM was initially seen as a niche product for GPUs, with skeptics questioning its cost and complexity.

Then

HBM found early use in AMD's Fury graphics cards but remained a small market for years.

Now

The AI boom made HBM essential, and SK hynix became the dominant supplier to Nvidia, holding the market lead Samsung now wants to take back.

Why this matters now

zHBM is a similar bet on a new memory architecture that could become essential. The difference: this time Samsung is the one making the radical bet, and SK hynix is the incumbent defending the interposer approach.

July 2015

Intel-Micron 3D XPoint (2015)

Intel and Micron announced 3D XPoint, a new memory technology they claimed was up to 1,000x faster than NAND and 10x denser than DRAM. It was positioned as a revolutionary new class of memory that would bridge the gap between storage and RAM.

Then

3D XPoint shipped as Intel Optane in 2017 but saw limited adoption due to high cost and niche use cases.

Now

Intel discontinued Optane in 2022, and 3D XPoint never achieved mainstream success, a cautionary tale for new memory architectures.

Why this matters now

zHBM faces the same adoption risk: a technically impressive new memory architecture that must prove it can be manufactured at scale and find enough customers to justify the investment.

Sources

(9)