Qualcomm HBC demonstrates AI inference leadership with validated silicon

Qualcomm Technologies is advancing AI inference with first-generation Qualcomm High Bandwidth Compute (HBC) silicon that successfully exercised the complete memory read-write path and near-memory computations in our San Diego lab. These hardware results demonstrate the core technology operating as designed — a milestone in delivering the memory foundation for our next-generation AI inference platforms.

First-generation Qualcomm HBC silicon validated the full memory read-write path and near-memory computations at production-representative dimensions.

Advancing AI inference through memory innovation

As AI moves into production, inference systems must address a core scaling challenge: how can increasingly capable models be served at scale without allowing power and cost to grow at the same rate?

For large language model inference, memory bandwidth is a central constraint. Generating each token requires model data to be supplied to compute. At data center scale, the energy and latency associated with that movement materially influence throughput, power consumption and total cost of ownership.

Qualcomm HBC addresses this constraint at its source. By placing compute closer to memory, HBC reduces data movement and increases effective bandwidth, improving the performance-per-watt equation for AI inference workloads.

What we validated in hardware

Using industry-standard automated test equipment, our engineering team drove first-generation Qualcomm HBC silicon through production-representative data pathways. The testing exercised the full memory read-write path and computations using randomized weights and activations at production-representative dimensions.

These measured results demonstrate the bandwidth and efficiency gains enabled by our near-memory compute architecture. With the data paths validated in physical silicon, HBC advances toward product integration with evidence supporting its intended role in production AI systems.

Engineering innovation for commercial scale

Qualcomm Technologies is translating memory innovation into a manufacturable platform technology for commercial AI infrastructure. First-generation Qualcomm HBC silicon marks the latest milestone in a staged engineering program spanning architecture, process development, test vehicles, silicon validation and product integration.

The program has established the capabilities required to move from a successful device to a deployable product:

• Validated architecture: The electrical, thermal and physical foundations of near-memory computing have been demonstrated through dedicated test vehicles and working silicon.
• Qualified bonding process: We have worked with leading foundry and assembly partners to develop and characterize wafer-to-wafer bonding across multiple technology generations.
• Designed-in resilience: Redundancy, remapping and validated test methods are built into the approach to identify and work around memory defects and exercise relevant electrical, thermal, bonding and interface failure modes.
• Manufacturing path: The technology is being engineered for repeatability, reliability and production scale, extending silicon validation toward commercial deployment.
• Active product integration: A data center product has taped out, and early bring-up of a logic base with bonded DRAM tiers is yielding today.

Built on more than a decade of memory-on-logic engineering

This milestone builds on more than a decade of investment in memory architecture, advanced packaging and 3D stacking. Qualcomm Technologies has commercialized direct bonding of DRAM onto logic silicon, helped shape JEDEC LPDDR standards and integrated memory into system-in-package products at global scale.

That experience brings together the disciplines essential to successful 3D memory integration: DRAM, logic, thermal behavior, test coverage, defect management, reliability and supply chain execution. Qualcomm Technologies has built these core engineering capabilities over years of development and is applying them through a unified, multi-year HBC program.

From validated Qualcomm HBC Silicon to Qualcomm Dragonfly AI250

The next step is already underway. Qualcomm Dragonfly AI250, expected to sample in the first half of 2027, will integrate HBC into a complete AI inference accelerator platform. It will bring the validated bandwidth and efficiency advantages of HBC into Qualcomm Dragonfly Rack, our full-stack inference infrastructure solution.

For hyperscalers, cloud providers and enterprises, the value proposition is direct: high-performance AI inference designed to improve tokens per dollar per watt, with the flexibility of an open, disaggregated rack architecture.