Qualcomm unveiled HBC architecture stacking compute beneath DRAM, claiming 6x bandwidth-per-watt advantage over HBM.
Qualcomm is unveiling HBC (High-Bandwidth Compute) as a breakthrough solution for the AI data center market to address the memory wall problem. Announced at its Investors Day 2026 under the Dragonfly brand, HBC represents an innovative near-memory technology that bonds compute with boosted memory bandwidth in a 3D stacked chip design, targeting the memory bottlenecks that have constrained the sector.
Currently, HBM dominates AI compute accelerators, but it faces efficiency challenges as higher token costs drive increased power consumption and total cost of ownership. Qualcomm claims HBC offers lower energy input per token, increased memory bandwidth, and reduced TCO. The architecture is built on four fundamentals: 3D integration leadership, system-level design, LPDDR leadership, and power-efficiency expertise.
The HBC design places the HBC accelerator beneath an LPDDR stack, with LPDDR selected for its larger capacity. The LPDDR stack connects to the HBC accelerator via Through-Silicon Vias (TSVs). In HBC's first generation, the solution integrates into Qualcomm's upcoming AI250 accelerator chip, with the HBC-boosted LPDDR stack sitting on the same 2D organic substrate. Each AI250 accelerator features 133 TB/s of bandwidth per card, an 18x boost over the AI200 with LPDDR5X.
Competitively, HBC promises 6x the bandwidth per watt versus HBM and 200x the capacity per watt versus SRAM. Qualcomm is partnering with strategic supply chain partners to address the primary bottlenecks in AI today: memory capacity, memory bandwidth, and cost of ownership.
The first-generation HBC Gen1 solution with AI250 accelerators is expected to launch by mid-2027. Qualcomm is also advancing a roadmap toward second-generation HBC Gen2, scheduled for 2028 and accompanied by the AI300 accelerator. HBC Gen2 will deliver up to 54x speed-up in effective bandwidth versus the AI200, along with a 7x jump in bandwidth per watt versus HBM.