Qualcomm Announces Strategic Entry Into Infrastructure Market With Dragonfly C1000 Server CPU High Bandwidth Compute Architecture And AI300 Inference Platforms To Power Agentic AI At Scale
Qualcomm Technologies disclosed a broad range of new products and a significant entry into the infrastructure market during its Investor Day event. Snapdragon unveiled a new set of silicon and hardware infrastructures that will help accelerate and improve artificial intelligence inference workloads at scale. These new products and silicon platforms include the Dragonfly C1000 CPU, High Bandwidth Compute architecture, the Dragonfly AI300 inference accelerator and next gen of high speed connectivity products. The focus on these new platforms is to increase throughput per watt while dramatically reducing total cost of ownership for hyperscale data centers. The strategic direction revolves around the explosive increase in the demand for agentic AI that needs to have continuous real time reasoning.
"Infrastructure has to offer substantially increased performance levels at a reduced power profile." Cristiano Amon, president and CEO, Qualcomm Incorporated said. "Dragonfly portfolio moves the company's low power computing prowess directly to the data center with multi year/multi generation contract with key industry partners"
At the heart of this expansion is the Dragonfly C1000 processor, a specialized, custom general purpose processor developed for cloud data center general purpose workloads, as well as AI head node functions. The Dragonfly C1000 chip is built using custom Oryon CPU cores running at over 5GHz. Qualcomm was able to employ a port multichiplet design that stacks over 250 cores on a single platform. This design is subsequently expected to provide over 2 times better efficiency by performance per watt, while using existing market server CPU offerings.
To power next generation accelerators and high speed networking Dragonfly C1000 offers 2+TB/s of PCIe Gen 7 connectivity with CXL intergration. The memory subsystem offers low latency and power consumption with optional High Bandwidth Compute attach points. Critical reliability services from error correction code to fault isolation are incorporated into the prodessor to scale across integrated systems.
The CPU offering is oriented towards three enterprise use cases with a high throughput agentic CPU, commodity general purpose CPU with best invert away CPU per virtual CPU, and AI head node CPU to maximize utilization of external accelerators. The commercial date for the processor is 2028.
To solve the data movement bottleneck that affects ordinary memory Qualcomm announced High Bandwidth Compute, a near memory computing architecture. This architecture combines processor and accelerated memory in a 3D stacked silicon architecture. Qualcomm says that the new architecture is more efficient and scalable than regular High Bandwidth Memory.
The new technology will be released as HBC Gen 1 and will be integrated into the AI250 card.
The effective memory bandwidth will be 133 TB/s, 18 times more than the last generation AI200 card. The next HBC Gen 2 will be included in the AI300 card providing 54 times more memory bandwidth than the AI200. This memory architecture promises to deliver 6 times the memory bandwidth per watt than the published specifications of similar memory cards.
The Dragonfly AI300 is said to be the flagship third generation AI inference platform, applying air and direct liquid cooling support in standard open compute server racks. The card also implements HBC Gen 2 to enable the processing of huge multimodal and large language models workloads. Qualcomm specifications for this platform mention a state of the art performance per watt of 4 to 8 times better than GPU based current systems.
For scaling, the AI300 supports UALink and Ethernet for Scale Up Networking, conformed both on copper and optical interconnects.
The commercial sampling of the AI250 with HBC Gen 1 is announced for mid 2027, and the AI300 will be advertised in 2028.
In addition to standard solutions, Qualcomm is broadening customized silicon offerings of custom chips for advanced cloud workloads. Qualcomm provides complete collaborative design (eliminating the need for multiple vendors) for silicon, packaging and software, to fulfill critical integration requirements. The launch is supported with a complete connectivity ecosystem to support 800G & 1.6T carrying speeds over copper and optical lines, for campus reach deployments up to 20KM without any data losses.
In order to back up this infrastructure push, Qualcomm communicated a major multi year partnership with Meta. The deal makes Qualcomm a strategic provider of data center cpus, with the Dragonfly C1000 scheduled to support Meta’s upcoming new servers fleet. Furthermore, over 35 global industry leaders within the hardware and cloud ecosystems have declared the data center vision. These partners encompass major manufacturing and technology companies such as Lenovo, Gigabyte, Supermicro, SK hynix, Micron, and Samsung SDS, further evidencing Qualcomm's alternative server architecture is ready to go.
