お知らせ • 3h
Qualcomm Technologies, Inc. Unveils Comprehensive Data Center Roadmap with New Qualcomm Dragonfly Portfolio
Qualcomm Technologies Inc. announced new data center solutions, including the Qualcomm Dragonfly C1000 CPU, Qualcomm High Bandwidth Compute (HBC), Qualcomm Dragonfly AI300 inference accelerator, and connectivity products, together with custom silicon solutions, all engineered to maximize performance per watt and token throughput at lower total cost of ownership. The new platforms highlight Qualcomm Technologies’ growing role in building full-stack data center infrastructure optimized for AI, spanning agentic and data-center-class CPUs, AI inference accelerators, high-performance connectivity, and at scale custom silicon solutions. The Qualcomm Dragonfly AI300 joins the previously announced Qualcomm Dragonfly AI200 and AI250 in its data center solutions portfolio with an annual cadence AI accelerator roadmap. Qualcomm Technologies draws on decades of expertise in systems-on-chips (SoCs), low-power design, high-performance processing, and leading IP, combined with experience engineering over 40 billion components, to deliver disaggregated, rack-scale AI infrastructure designed for data-center-grade, agent-intensive AI inference workloads at hyper scale. These innovations enable improved token economics, low latency, simplified integration, scalable deployment, and lower total cost of ownership. As agentic AI dramatically increases token demand, Qualcomm Technologies’ solutions are optimized for tokens-per-watt as the key lever to reduce total cost of ownership (TCO). Qualcomm Dragonfly C1000 CPU is a purpose-built data center CPU designed for leadership performance and utilization for agentic, general-purpose, and AI head node workloads at best-in-class power efficiency and TCO. Custom-designed Qualcomm Oryon CPU cores are optimized for core performance and frequencies greater than 5 GHz to deliver superior performance for agentic workload deployed at scale. The product features a 250+ core count chiplet design for exceptional throughput and scale while delivering exceptional per-core performance. The CPU offers greater than 2x better performance per watt estimate compared to existing product benchmarks for server CPU competitive offerings based on specs. It is architected and designed for best throughput, responsiveness, and infrastructure utilization for critical data center usages and lowering CapEx and OpEx to deliver best-in-class performance per TCO leadership at scale. Multi-chiplet architecture enables modular integration with advanced packaging technologies for performance and IO scaling addressing general-purpose to AI CPUs in the data center domain. The CPU provides greater than 2 TB/s leading-edge PCIe Gen 7 connectivity, plus CXL connectivity, to support next-generation accelerators, high-speed networking & storage and memory disaggregation. The memory sub-system is built to deliver superior bandwidth, capacity, latency and power efficiency using leading-edge low-power memory technology. CPU-based inference with optional HBC attach is available. The product is built with advanced reliability, availability, and serviceability (RAS) features, including ECC, fault isolation, and error recovery to enable resilient operation at scale. Support for both air and liquid cooling enables deployment across diverse data center environments with OCP ORv3 compliant racks and servers. The CPU portfolio includes: agentic CPU designed for high-throughput agentic orchestration and low latency interactive AI use cases; general-purpose CPU designed for optimal performance-per-TCO for first-party workload and performance-per-vCPU for third-party usage elasticity; AI head node CPU designed to maximize XPU utilization of XPU for generative AI compute through low overhead host processing through high-speed CPU. Commercial availability is expected in 2028. Qualcomm High Bandwidth Compute (HBC) is an innovative purpose-built near-memory computing architecture that bonds compute with highly-accelerated memory bandwidth in a 3D-stacked silicon solution to address AI’s fundamental data movement bottleneck. HBC has a multi-generation roadmap to deliver faster, more efficient, and more scalable processing at lower total cost of ownership and higher energy efficiency compared to high bandwidth memory (HBM). With HBC Gen 1, AI250 is designed to enable an industry-leading 133 TB/s per card, an 18x increase in effective memory bandwidth compared to AI200 with LPDDR5X; AI300 with HBC Gen 2 is designed to enable another stepwise improvement with a 54x increase over AI200. HBC is designed to enable a 6x increase in bandwidth per watt versus HBM compared to competing published product specifications normalized at card-level. HBC is designed to enable a 200x increase in capacity per watt versus SRAM compared to competing published product specifications normalized at rack-level. HBC is designed to enable efficient scaling of AI agents to meet the demands of continuous reasoning, memory bandwidth, and real-time responsiveness. Strategic relationships with the supply chain and unique implementation addresses near-memory computing complexity due to 3D integration leadership, system-level design, LPDDR leadership, and power efficiency expertise. Commercial sampling of HBC Gen 1 with AI250 is expected in mid-2027. Qualcomm Dragonfly AI300 (Card and Rack) is a third-generation, air- and direct-liquid-cooled rack-level AI inference platform – following the introduction of the AI200 and AI250 solutions last October. AI300 integrates breakthrough Qualcomm HBC Gen 2 technology for compute acceleration with integrated memory and increased effective memory bandwidth, designed for disaggregated inference deployments (AI250 uses HBC Gen 1). The product enables industry-leading memory capacity and effective bandwidth enabling high-throughput, low-latency performance for large language & multimodal model (LLM, LMM) inference and agentic AI workloads. The product is expecting 4x-8x better performance-per-watt compared to existing GPU-based architectures on memory bandwidth per watt per card. The product can scale up with UALink (Ultra Accelerator Link) and ESUN (Ethernet for Scale-Up Networking); scale out with copper and optical. Commercial sampling is expected in 2028. Custom Silicon provides performance-optimized silicon at scale for next-generation AI and cloud data center infrastructure. Bespoke custom silicon is available for agentic AI and other specialized workloads. End-to-end co-design capabilities across silicon, system, and software address customer-specific performance, power, and integration requirements. Advanced packaging and modular architectures are designed to improve performance, power efficiency, and scalability.