Intel and SoftBank Explore HB3DM Memory with Higher Bandwidth Than HBM4

Why It Matters

HB3DM could redefine AI accelerator design by alleviating data‑movement limits, giving Intel a potential edge in next‑gen high‑performance computing. Its success may spur a new class of memory solutions that prioritize throughput over capacity.

Key Takeaways

• HB3DM uses nine-layer stack with 13,700 TSVs per layer
• Bandwidth density reaches 0.25 Tb/s per mm², ~5.3 TB/s per module
• Total capacity ~10 GB, far below HBM4’s 48 GB per stack
• Hybrid bonding promises lower latency and power use, details pending
• HB3DM targets AI inference workloads where bandwidth outweighs memory size

Pulse Analysis

The AI boom has exposed a critical weakness in modern compute platforms: memory bandwidth. While HBM generations have steadily increased capacity, they still struggle to keep pace with the data‑movement demands of large language models and real‑time inference. Intel’s partnership with SoftBank through Saimemory introduces HB3DM, a memory architecture that flips the traditional trade‑off by stacking a logic die beneath eight DRAM layers and employing a Z‑Angle hybrid‑bonding technique. This approach yields a striking 0.25 Tb/s per square millimeter, translating to roughly 5.3 TB/s for a 171 mm² module—far surpassing current HBM4 throughput per area.

Technically, HB3DM’s nine‑layer configuration packs about 13,700 through‑silicon vias per layer, creating dense vertical interconnects that minimize latency and power loss. However, the design sacrifices capacity, offering only about 10 GB per stack versus HBM4’s 48 GB. This makes HB3DM ideal for workloads where moving data quickly is more valuable than storing large datasets, such as edge AI inference, high‑frequency trading, and certain scientific simulations. The hybrid bonding process also promises tighter layer coupling, potentially reducing power consumption, though Intel has yet to publish concrete efficiency figures.

From a business perspective, the venture underscores Intel’s strategy to diversify beyond CPUs into niche memory markets, leveraging SoftBank’s capital and supply‑chain reach. If manufacturing hurdles—like DRAM sourcing and yield optimization—are overcome, HB3DM could become a differentiator for Intel‑based AI accelerators, compelling OEMs to reconsider memory‑centric system designs. The upcoming VLSI 2026 symposium will be a litmus test for industry interest, and successful adoption could trigger a wave of bandwidth‑first memory solutions, reshaping the competitive landscape for high‑performance computing.