HBM4-Alternative: Qualcomm validates HBC memory for Dragonfly
Qualcomm has announced the completion of the first validation phase for its High Bandwidth Compute (HBC) memory architecture. Designed for the upcoming Dragonfly accelerators, this technology aims to combine high memory bandwidth with improved energy efficiency and lower manufacturing costs.
HBC architecture consists of multiple DRAM layers stacked directly above a small compute chip. Unlike HBM, where the memory stack is connected to the processor via a wide interface, HBC performs bandwidth-intensive computations within its own Compute Unit located directly at the memory level. The results are then forwarded to the central processor or SoC, reducing the burden on the main data bus.
Energy Efficiency and Substrate Savings
First detailed at Investor Day in mid-2026, the proximity of memory and compute units is intended to shorten data paths and save power. This is particularly relevant for AI inference, which demands both high bandwidth and significant capacity. Qualcomm plans to use LPDDR5X-DRAM for these stacks instead of specialized HBM chips. This approach could allow for cheaper connection methods, such as conventional organic 2D substrates, for the main processor interface since the stack handles part of the computational load itself. However, the vertical stacking still requires Through-Silicon Vias (TSVs) to maintain electrical connectivity through the silicon.
Validation Results and Roadmap
During validation, Qualcomm tested an early HBC stack consisting of two DRAM layers, with a target of four layers for production models. The company claims that HBC achieved over five times the energy efficiency (bandwidth per watt) of HBM4, falling slightly short of its original six-fold target. Additionally, Qualcomm reports that the bandwidth is 18 times higher than external LPDDR5X, though specific values were not provided. Thermal management was also a key part of the validation, addressing the heat challenges posed by placing a compute chip at the base of the memory stack.
The Dragonfly AI250 is expected to be the first AI accelerator to implement HBC technology. Qualcomm plans to begin sampling the product in the first half of 2027. While the technology is also expected to eventually reach consumer Snapdragon products, the specific timeline for mobile integration remains unconfirmed.
Гика
