Qualcomm has completed testing on the first chips built with its High Bandwidth Compute (HBC) memory technology, clearing a key hurdle before the design can be used in data center AI hardware. The company announced the results and said HBC will be featured in its Dragonfly AI250 accelerator, a chip meant to run AI models in data centers.
Best known for its Snapdragon smartphone processors, Qualcomm introduced the Dragonfly brand for its data center products in June. HBC is aimed at one of the biggest constraints in running large language models at scale: moving data between memory and processors fast enough to keep the chips busy.
What is HBC memory in Qualcomm's AI chips?
The approach places computing capability close to the memory itself, which shortens the distance data has to travel. Qualcomm says that increases effective memory bandwidth and raises the amount of work a system can perform per watt of power used.
Testing took place at Qualcomm's laboratory in San Diego using standard automated test equipment. Engineers checked the full memory read and write path, along with calculations performed alongside the memory. They ran random data at sizes that reflect commercial products.
Qualcomm said the results confirmed the bandwidth and energy efficiency gains it had expected. The company also acknowledged that its comparisons with competing products are estimates, adjusted for target operating conditions. Actual performance, it said, will depend on the final product and the workload it runs.
The company has also finished the design of a data center product that uses HBC and sent it for manufacturing, a stage known as tape-out. Early bring-up is now underway for a logic chip with layers of DRAM bonded directly on top.
Getting there required close work with foundry and assembly partners on wafer-to-wafer bonding, the process that stacks memory directly onto logic chips. The design also includes spare capacity and remapping capabilities, which allow the system to route around defective memory cells.
The project builds on more than a decade of Qualcomm research into stacking memory and logic. The company has previously shipped products with DRAM bonded directly to logic chips, and it has contributed to the development of LPDDR memory standards used in smartphones.
How much memory bandwidth will the Dragonfly AI250 offer?
Qualcomm expects the AI250 to begin sampling in the first half of 2027. Its product specifications list 133TB/s of effective memory bandwidth per card, which the company says is 18 times the figure for its AI200 accelerator.
The card will be part of Dragonfly Rack, Qualcomm's complete server rack for running AI models. The company is targeting cloud providers, hyperscalers and large enterprises, and it positions the hardware as a way to increase AI computing output while reducing costs and energy consumption.
Este artigo foi escrito com a assistência de IA.
News Factory APP - notícias agênticas para impulsionar seu SEO e AEO.