Intel's XBM Memory Architecture Proposes Backend-Transistor DRAM with UCIe for AI Workloads

Intel’s patent for "Cross-Batch Memory" (XBM) details a DRAM stack utilizing backend-transistor fabrication and serial UCIe links to bypass traditional HBM interposer constraints. The design aims to reduce packaging complexity and Z-height while incorporating on-die repair logic.

Chips & Climate Desk
Updated 6d ago 2 min read
Intel's XBM Memory Architecture Proposes Backend-Transistor DRAM with UCIe for AI Workloads

Intel has unveiled a patent application for "Cross-Batch Memory" (XBM), a high-bandwidth memory architecture designed to circumvent the packaging and cost limitations inherent in current interposer-based HBM solutions. According to Tom's Hardware, the patent, published July 2, 2026, describes a memory stack that matches the footprint of HBM4 while replacing conventional DRAM interfaces with back-end-of-line (BEOL) transistors and serial Universal Chiplet Interconnect Express (UCIe) links. This shift aims to eliminate the silicon interposer, which is currently a primary driver of assembly costs and manufacturing complexity in high-performance memory subsystems.

The XBM architecture fundamentally alters DRAM cell placement. Conventional DRAM cells are fabricated in the front-end-of-line, but XBM moves the 1T1C (one-transistor one-capacitor) cells into the BEOL, utilizing thin-film transistors. Tom's Hardware reports that the design features dies of approximately 1.5 gigabytes, organized into 768 datablocks within a 32-by-24 grid. These are grouped into eight channels, each containing eight sub-channels. The stack is designed to scale between eight and 16 dies high, with data transmission handled via UCIe I/O bundles operating at 32 GT/s, facilitated by a base die that performs serialization and deserialization.

To address the yield challenges associated with tall memory stacks, Intel has integrated extensive repair capabilities directly into the base die. The architecture includes built-in self-repair (BISR), dedicated spare channels, and debug logic. Furthermore, the design incorporates four sub-channels of redundant memory arrays, which serve as fungible spares to compensate for defects in the stacked dies. This post-assembly repair strategy is intended to improve manufacturing yields, a critical factor given the high defect sensitivity of densely stacked memory arrays compared to traditional monolithic designs.

Beyond the transistor-level changes, the patent focuses on reducing the physical Z-height of the memory package. Intel proposes "reversed overhang" structures and memory-on-package (MoP) configurations to minimize vertical space. According to Tom's Hardware, these methods aim to eliminate the stiffeners typically required to manage warpage in conventional packages, while also allowing for direct power delivery from voltage regulators to the DRAM. By removing the interposer and optimizing the physical stack, Intel intends to create a more compact and cost-effective alternative to the wide parallel interfaces currently dominating the market.

It is important to distinguish XBM from ZAM (Z-Angle Memory), another architecture Intel is co-developing with SAIMEMORY, which focuses on fusion-bonded, nine-layer stacks of conventional DRAM. Tom's Hardware notes that XBM remains an internal Intel filing with no current product roadmap or shipping hardware. The architecture faces significant engineering hurdles, including the unproven nature of BEOL DRAM at manufacturing scale and the fact that the UCIe interface is already operating at its current specification ceiling. Consequently, XBM represents a potential design direction rather than an imminent commercial offering, requiring further validation against competing standards like HBM4E.

Sources

  1. 01 Intel patent reveals new XBM memory architecture that ditches HBM's costly silicon interposer — backend-transistor DRAM stack uses UCIe links and built-in repair to ease AI's memory bottleneck — Tom's Hardware