RU

DRAM stacked on logic: Korea and Japan take a swing at HBM

Published: 2026-11-19 · Author: AI Release · @ai_releaseorg
DRAM stacked on logic: Korea and Japan take a swing at HBM

Memory sitting next to the chip is a compromise everyone accepted as normal. SemiFive and Preferred Networks want it on top instead. If it works, more than a chip changes — the whole supply chain shifts.

What they are actually proposing

According to DigiTimes, Korea's SemiFive and Japan's Preferred Networks are developing an inference chip that stacks DRAM directly above a logic die. Conventional 2.5D packaging puts the processor and HBM side by side on a shared interposer. Here memory moves upstairs: shorter path, fewer transitions, fewer watts.

Sounds obvious. In practice it hits two walls — heat and yields. A hot stack either throttles or ends up in the bin. That is exactly why 2.5D survived this long.

Who is playing

SemiFive is a Korean design house that builds custom silicon. Preferred Networks is a Japanese AI lab that has spent years building its own stack: models, frameworks, and the hardware underneath. One side knows how to get a design into silicon, the other knows where inference chokes without memory bandwidth.

An odd pairing, and a sensible one. Each brings what the other lacks.

Why it matters outside Asia

HBM is the scarcest part of AI hardware. Only a handful of companies make it, the queue is booked out, and its price sets inference economics for everyone from startups to cloud giants. Any working alternative cools the market and reshuffles the memory game.

Second layer: packaging. 2.5D stopped being a secondary step long ago — it decides how many chips you can ship at all. Whoever offers a real alternative gets leverage comparable to the memory makers.

Third: geography. While everyone stares at Taiwan, Korea and Japan are assembling their own vertical. TSMC is not the lead actor here. That alone is news.

Where the risk is

The report has no bandwidth figures, no timeline, no named customers. This is development, not a shipping product. Stacking DRAM on logic has been attempted before — the same two problems every time: heat and cost. If SemiFive and PFN don't solve thermal, the result is a handsome demo chip for a keynote, not an HBM replacement.

And one more: a single source. DigiTimes is an industry-grade leak outlet, but no second independent publication covers this work. For now it is a signal, not a fact.

💬 What to argue about

  • Is DRAM-on-logic a genuine HBM replacement, or a niche for inference workloads that can live with less bandwidth?
  • If packaging stops being the bottleneck, who loses first: memory makers or the companies that assemble 2.5D?
  • Will heat and yields kill this idea again — or does the Korea-Japan pair have an answer the others didn't?

Prediction: If the pair cracks thermal, the first public chip demo lands before HBM gets any cheaper.

Disclaimer: supply chain leak, not officially confirmed

DigiTimes

🤖 AI summary
  • Date: 2026-11-19
  • Source: www.digitimes.com
  • Tags: SemiFive, HBM, DRAM
  • Key facts:
    • Memory sitting next to the chip is a compromise everyone accepted as normal.
    • SemiFive and Preferred Networks want it on top instead.
    • If it works, more than a chip changes — the whole supply chain shifts.
    • According to DigiTimes, Korea's SemiFive and Japan's Preferred Networks are developing an inference chip that stacks DRAM directly above a logic die.
    • Conventional 2.5D packaging puts the processor and HBM side by side on a shared interposer.
    • Here memory moves upstairs: shorter path, fewer transitions, fewer watts.
    • In practice it hits two walls — heat and yields.
    • A hot stack either throttles or ends up in the bin.
#SemiFive#HBM#DRAM
📖
Read the guide on this topic
Read →
← PreviousAI Digest 19.11.2026: From Commands to Context: OlloNi SS1's… and 4 moreNext →Deep dive: what's behind the wave of tech outages — and what PHP and AI agents have to do with it

Source: www.digitimes.com · post in Telegram