Intel explores stacking memory directly above future CPUs to tackle AI performance bottlenecks

2 Sources

Share

Intel CEO Lip-Bu Tan announced the company is developing new Intel memory architecture that places DRAM vertically above processors using 3D stacking. The move targets performance bottlenecks in AI workloads and signals Intel's potential return to memory manufacturing after hiring former SK hynix CEO Seok-Hee Lee.

Intel Repositions Memory as Strategic Priority for AI Workloads

Intel CEO Lip-Bu Tan announced the company is fundamentally rethinking its approach to Intel memory architecture, treating DRAM as a strategic component rather than an interchangeable commodity. The shift comes as AI workloads and data-intensive applications increasingly create performance bottlenecks between processors and storage.

1

Tan explained that memory now determines complete-system performance as compute and memory demand continues growing with multiple Gigawatt-scale AI Factories coming online in the years ahead.

2

The strategic pivot marks a significant departure from Intel's previous stance. Tan admitted he personally avoided investing in memory because it was "kind of commodity business," but market dynamics have changed drastically with growing AI demand.

2

Memory has become strategically important because GPU, CPU and accelerator performance can be limited by bandwidth and capacity rather than arithmetic throughput.

1

Stacking Memory Directly Above Future CPUs Using 3D Stacking

One architecture under active consideration would place memory vertically above a processor through 3D stacking. This approach to CPU-memory integration could shorten data paths and increase interface width compared with conventional DIMMs or package-adjacent memory.

1

The company is working on embedding CPUs with stacked DRAM capabilities to meet growing customer requirements.

2

However, stacking memory directly above future CPUs creates substantial manufacturing and cooling challenges. Stacking DRAM over high-power compute logic can restrict heat removal, while additional bonding steps introduce yield and repairability concerns.

1

Intel has not disclosed the specific memory type, interconnect, capacity or commercialization schedule for this future computing architecture.

XBM Architecture and Advanced Memory Packaging Technologies

A related Intel patent describes an XBM architecture designed to provide high-bandwidth memory without requiring a large silicon interposer. Traditional HBM connects several DRAM stacks and a processor through an interposer carrying thousands of wires. The alternative packaging approach could reduce cost, though performance and manufacturing yield will determine whether it becomes competitive.

1

Intel has been making inroads in the memory segment with partners, developing technologies such as XBM and ZAM. According to current timelines, Intel's ZAM project is scheduled for 2029-2030, while XBM should be unveiled at the start of the next decade.

2

These initiatives represent course-corrections after Intel's past DRAM attempts, including HMC (Hybrid Memory Cube) and MCDRAM, faced various issues and never reached market.

2

Strategic Hiring Signals Intel's Memory Ambitions

Intel recently recruited former SK hynix chief executive Seok-Hee Lee as EVP of the Foundry business and advanced memory and packaging technologies.

1

Tan called Lee "my good friend" and suggested the hire signals something significant: "you kind of know something that I'm thinking about we are not ready to unfold it."

2

The appointment strengthens Intel's expertise in advanced memory and packaging at a critical time when the entire memory segment faces heightened shortages.

Intel's Historical Context and Future Direction

Intel, founded in 1968, started as a semiconductor business with its first products being memory chips, including the 3101 SRAM and the 1103 DRAM.

2

The company later focused on processors, eventually selling its NAND and SSD operation to SK hynix. Intel's previous venture was Optane products based on 3D XPoint architecture co-developed with Micron, but these were discontinued in 2022 as competing technologies like HBM showed better potential.

2

The planned return would likely involve differentiated memory rather than rebuilding a conventional commodity-DRAM business.

1

Owning more of the memory interface could help Intel differentiate future Xeon, Core and accelerator products to address performance bottlenecks.

1

If Intel successfully returns to memory manufacturing, its Foundry Services business could become a leading manufacturer of advanced semiconductors, memory solutions, and bleeding-edge packaging technologies all under one roof.

2

Memory shortages are expected to persist beyond 2030 as compute demand grows.

2

Today's Top Stories

© 2026 TheOutpost.AI All rights reserved