Could AMD Be the First Company to Use Custom HBM?, Sorry I'm Using a Translator.

Why It Matters

If AMD succeeds, it could secure a performance lead in AI‑accelerated graphics and reshape the high‑bandwidth memory ecosystem, pressuring rivals and memory suppliers alike.

Key Takeaways

• AMD may adopt custom HBM4e for upcoming GPUs
• Development cost exceeds $100M, raising efficiency concerns
• Base die compatibility issues with three memory vendors
• Samsung and Hynix have dedicated custom HBM teams
• MI455X rumored to use three c‑dies and custom B‑die

Pulse Analysis

High‑bandwidth memory (HBM) has long been a differentiator for premium graphics processors, but its adoption has been limited by cost and supply constraints. AMD’s potential shift to a custom HBM4e stack—leveraging a proprietary base die, often referred to as a b‑die—represents a strategic move to break the current vendor‑centric model. By engineering its own memory interface, AMD could tailor bandwidth, power efficiency, and stack height to the specific needs of its upcoming MI455X GPU, positioning itself ahead of Nvidia and Intel in latency‑critical workloads such as AI inference and ray‑traced gaming.

The technical and financial hurdles, however, are non‑trivial. Forum participants estimate development expenses north of $100 million, a figure that dwarfs typical GPU‑memory integration budgets. Compatibility is another pain point: the custom base die must interoperate with three major memory manufacturers—Samsung, Hynix, and TSMC—each offering distinct process nodes and I/O architectures. Samsung and Hynix have already formed dedicated custom HBM teams, indicating industry readiness, while TSMC’s proposal focuses on larger I/O pads rather than a true custom stack. Decisions between a single‑die B‑die approach, a dual‑die C‑HBM configuration, or a hybrid solution will directly affect yield rates and per‑chip pricing.

Should AMD bring a custom‑HBM GPU to market, the ripple effects could be profound. A successful MI455X launch would validate the economics of bespoke memory stacks, encouraging other silicon designers to pursue similar paths and potentially driving down HBM prices through increased volume. Memory vendors would be compelled to offer more flexible design services, accelerating innovation across the AI and high‑performance computing sectors. In essence, AMD’s gamble on custom HBM could redefine the balance of power between GPU manufacturers and memory suppliers, reshaping the competitive landscape for years to come.