November 22, 2026

Qualcomm Validates High-Bandwidth Computing Architecture to Overcome the AI Memory Wall

In a significant move to reshape the landscape of artificial intelligence infrastructure, Qualcomm executives have begun showcasing the first tangible results of a new architectural approach designed to eliminate one of the most persistent bottlenecks in modern computing. This summer, the company introduced an architecture known as High-Bandwidth Computing (HBC), which is positioned as a direct answer to the "memory wall"—a phenomenon where the speed of data movement, rather than raw processing power, becomes the primary constraint on performance for AI inference and the emerging field of agentic AI.

The memory wall has long been a source of frustration for chip designers and data center operators alike. As AI models grow in complexity and size, the industry has reached a point where the performance of these systems depends less on how many calculations a processor can perform per second and more on how quickly that processor can be fed the necessary data. This imbalance not only throttles the potential of advanced large language models but also significantly inflates the costs associated with running AI workloads at scale.

Qualcomm’s strategy involves a fundamental shift in how compute and memory interact. Traditionally, the industry has relied on two primary methods to manage data flow. The first involves housing the AI model within fast, on-chip memory, which offers incredible speed but is severely limited by physical space and capacity. The second, more common approach in high-end AI hardware involves the use of stacked High-Bandwidth Memory (HBM). While HBM provides the necessary throughput, it comes with a steep premium in terms of cost, power consumption, and manufacturing complexity.

Qualcomm is now advocating for a third path: near-memory computing, or HBC. This architecture effectively moves the compute resources closer to the data they need to process, minimizing the physical distance data must travel and, by extension, reducing the latency and energy required for every operation.

Gerardo Giaretta, vice president and general manager of datacenter CPU and AI at Qualcomm, recently detailed the challenges of the current status quo during a briefing with journalists and analysts. He noted that traditional computer architectures typically rely on DDR memory located off the package. While DDR offers plenty of capacity, it is hampered by a narrow link that moves data at a relatively slow pace. To solve this, the industry turned to HBM, which increases the number of "lanes" and the speed of data movement between the compute and the memory stack.

However, Giaretta pointed out that HBM is not a perfect solution. The premium paid for HBM is not just financial; it includes a significant increase in power consumption and requires highly complex 2.5D packaging and co-host packaging techniques. These requirements make it difficult to scale efficiently, especially as the demand for inference-only workloads grows.

The HBC architecture aims to solve the problem of data movement by fundamentally changing the physical relationship between the components. In Qualcomm’s implementation, a stack of LPDDR-based memory is placed directly over a compute die. This vertical integration means that data no longer has to traverse millimeters of distance across a circuit board or an interposer. Instead, it moves across micrometers.

The physical proximity translates into immediate and measurable gains. According to Giaretta, the reduction in the distance data traverses results in a "very, very high bandwidth" between the memory and the compute. More importantly, it yields substantial energy savings. By reducing the energy spent simply moving bits from one place to another, Qualcomm claims it can achieve an order of magnitude improvement in energy efficiency.

Diving deeper into the technical specifics, Giaretta explained that the first generation of HBC utilizes an LPDDR stack with a dedicated AI accelerator positioned directly beneath it. This accelerator is designed to handle the vast majority of operations required for AI workloads. While some operations may still be offloaded to the central system-on-a-chip (SoC)—specifically Qualcomm’s Dragonfly AI250—the goal is to keep the bulk of the processing within the HBC package itself.

The Dragonfly AI250 platform, which is slated for release next year, boasts significantly higher TFLOPs than its predecessors. However, the interface between the AI250 and standard LPDDR is naturally slower than the internal HBC link. To manage this, Qualcomm treats the HBC component as a co-processor. The system is designed so that once the HBC is instructed to perform a task, the majority of the decode operations are handled internally. The end-to-end performance remains high because the system only needs to send back results across the slower link, rather than the massive raw data sets.

This week, Qualcomm announced a major milestone in this journey: the first generation of HBC silicon has successfully reached the lab for validation. Using industry-standard automated test equipment (ATE), Qualcomm engineers have confirmed that the complete memory read-write path and near-memory computations are functioning at production-level standards.

The results of these tests have been striking. Qualcomm reports that HBC can deliver up to 18 times the memory bandwidth compared to standard LPDDR. Furthermore, in terms of energy efficiency, the architecture achieved a measurement of five times the terabit-per-second-per-watt compared to HBM4, which is the sixth generation of high-bandwidth memory.

For data center operators, these statistics represent a potential shift in the total cost of ownership (TCO). AI inference is inherently memory-bandwidth limited, and by providing that bandwidth at a lower energy cost, Qualcomm is targeting the operational expenditures (OPEX) that currently plague large-scale AI deployments. Giaretta emphasized that the amount of energy spent moving data in this new architecture is at least one order of magnitude less than traditional methods.

The validation of HBC silicon is a critical component of Qualcomm’s broader ambitions in the data center. For over a year, the company has been positioning itself as a serious contender in the AI infrastructure space and a viable alternative to Nvidia, which currently dominates the market. The HBC technology will be integrated into the Dragonfly AI250 to create a comprehensive AI inference accelerator platform. This platform will serve as the foundation for the Qualcomm Dragonfly Rack, a full-stack infrastructure solution specifically designed for inference workloads.

The company’s roadmap for the coming years is ambitious. Following the current validation phase, the AI250 is expected to start sampling in the first half of 2027. This will be followed by the Dragonfly C1000 CPUs, which are also set for next year, and samples of the Dragonfly AI300, which are scheduled for 2028.

Giaretta noted that the validation is a "clear differentiator" for Qualcomm. He stressed that the solution is not just about improving memory bandwidth—it is about the integration of compute directly under the memory stack. This "High-Bandwidth Compute" label is intentional, signaling that the company is looking beyond simple storage and retrieval to a more holistic processing model.

Furthermore, Qualcomm is actively collaborating with various partners to explore the diverse applications of HBC. While the Dragonfly AI250 is the flagship product for this technology, it will not be the only one. The architecture’s versatility suggests potential applications across various segments of the computing market where the memory wall remains a hurdle.

The successful bring-up of the HBC silicon in the lab is also intended to address industry skepticism regarding manufacturing. Because HBC involves a novel way of stacking and connecting components, questions have been raised about whether Qualcomm could achieve the high yields necessary for mass production. Giaretta addressed these concerns directly, stating that the decision to share these early lab results—a rare move for a company that produces a vast volume of chips—was made specifically to prove the maturity of the technology.

"It’s important to understand that this is not a normal chip," Giaretta said. "This is the first time that somebody has done something like this." By validating the silicon now, Qualcomm aims to provide the industry with a "good validation" of the high-yield manufacturing capabilities they expect to demonstrate next year.

As the AI industry shifts its focus from training massive models to the more frequent and cost-sensitive task of running them (inference), the efficiency of the underlying hardware will become the primary battleground. With HBC, Qualcomm is betting that the key to winning that battle lies not in how fast a chip can think, but in how little effort it takes to remember.