Samsung Unveils Three-Phase HBM Roadmap at Hot Chips 2026, Aiming to Merge Memory and Compute

Plan culminates in zHBM architecture that stacks DRAM directly atop processors, eliminating interposer

edit
By LineZotpaper
Published
Read Time3 min
Samsung has detailed a three-phase roadmap for high-bandwidth memory (HBM) that progressively integrates logic and compute capabilities into the memory stack, culminating in a radical 'zHBM' design that places DRAM directly above the processor. Presented at Hot Chips 2026, the plan leverages an advanced logic process for the HBM base die to address rising bandwidth and power constraints in AI accelerators.

Samsung outlined its vision for the future of high-bandwidth memory at the Hot Chips 2026 conference, describing a three-phase evolution that moves HBM from a passive data conduit to an active participant in computation. The company's Sangwook Han of the DRAM design team identified the base die as the key enabler, beginning with Samsung's decision to manufacture the HBM4 base die on a 4nm logic process rather than a DRAM process.

In conventional HBM, multiple DRAM core dies are stacked vertically on a base die and connected via thousands of through-silicon vias (TSVs). The stack sits beside an XPU on a 2.5D interposer, with the base die bridging memory and compute. Current HBM4 stacks offer roughly 1 to 5 TB/s of bandwidth using 1,024 to 2,048 data I/Os running at 8–16 Gbps each. However, scaling bandwidth by adding more TSVs or increasing signaling speed consumes area, forces tighter pitches, and drives up power consumption — the total HBM power continues to rise even as energy per bit improves.

Phase 1: Reclaim XPU Area

The first phase focuses on reducing the physical footprint of memory interfaces on the accelerator die. Samsung proposes moving the HBM Physical Interface (PHY) — one of the largest non-compute blocks on the XPU — onto the base die, replacing the traditional interface with a much smaller die-to-die (D2D) link. This frees up valuable silicon area on the XPU for additional compute logic, helping to overcome reticle and interposer limits.

Phase 2: Active Memory

In the second phase, the base die gains more processing capabilities, transitioning from a simple interface to performing data-intensive operations such as near-memory compute. This approach reduces data movement between memory and the XPU, lowering latency and power consumption. The exact functions to be offloaded remain under development, but Samsung envisions the base die handling tasks like data filtering, compression, or simple arithmetic.

Phase 3: zHBM — Stacking DRAM Directly on the Processor

The final phase, called zHBM, eliminates the physical gap between DRAM and the XPU altogether. The entire DRAM stack is placed directly on top of the processor, removing the need for a 2.5D interposer. By stacking memory vertically over compute, zHBM dramatically shortens interconnect distances, potentially offering orders-of-magnitude improvements in bandwidth density and energy efficiency. Samsung debuted zHBM as part of its next-generation memory technologies for AI data centers, relying on advanced wafer bonding techniques.

Samsung refers to the broader strategy as "custom HBM" (cHBM), where the base die is tailored for a specific accelerator while keeping the standard DRAM stack. The roadmap underscores a growing industry push to overcome the "memory wall" by embedding intelligence within the memory hierarchy.

§

Analysis

Why This Matters

  • Samsung's roadmap directly addresses the scaling limits facing AI accelerators, where bandwidth growth is increasingly constrained by power and physical layout.
  • Moving memory interface logic off the XPU die could free up critical area for more compute units, potentially boosting accelerator performance.
  • If successful, zHBM could upend current packaging approaches (2.5D interposers) and redefine memory hierarchy in data centers.

Background

High-bandwidth memory (HBM) has become the standard for AI and high-performance computing workloads, used in GPUs from NVIDIA, AMD, and Intel. HBM stacks multiple DRAM chips vertically, connected to a processor via an interposer. Each generation increases the number of data I/Os and signaling speed, but power and die area constraints are mounting. Samsung's approach to moving the base die to a logic node (first used in HBM4) enables more sophisticated functionality in the base die, setting the stage for deeper integration.

Key Perspectives

Samsung: By migrating the base die to a 4nm logic process, Samsung gains a more capable silicon layer that can offload functions from the XPU and eventually host compute logic. The phased roadmap allows incremental adoption without requiring immediate redesign of the DRAM stack. XPU makers (NVIDIA, AMD, etc.): They stand to benefit from reclaiming die area and reducing memory-related power, but may face integration challenges and dependency on Samsung's packaging capabilities. Custom HBM allows tailoring to specific accelerators, potentially locking in supply chain relationships. Critics/Skeptics: Stacking DRAM directly on top of a processor introduces thermal and reliability challenges, as DRAM operates best at lower temperatures than logic. Manufacturing yields for such dense 3D stacks remain unproven at scale. Additionally, the industry is heavily invested in current 2.5D packaging infrastructure, which may slow adoption.

What to Watch

  • Adoption of Samsung's custom HBM (cHBM) base dies by major accelerator vendors in upcoming products.
  • Thermal management solutions for zHBM's processor-under-DRAM stacking.
  • Competitors' responses — SK Hynix and Micron have their own HBM roadmaps that may differ in integration philosophy.

Sources

newspaper

Zotpaper

Articles published under the Zotpaper byline are synthesized from multiple source publications by our AI editor and reviewed by our editorial process. Each story combines reporting from credible outlets to give readers a balanced, comprehensive view.