Intel unveils trio of AI chips to rival hyperscalers

Intel’s multi-architecture blueprint for agentic AI introduces three specialized processors—Diamond Rapids CPU, Crescent Island GPU, and Wildcat Lake SoC—each meticulously engineered to address distinct segments of AI workloads, from enterprise-scale orchestration in data centers to cost-sensitive edge and client devices. Diamond Rapids boasts up to 256 cores, a massive 1.28GB last-level cache, and supports cutting-edge I/O standards like PCIe Gen6 and CXL 3.0, making it ideal for complex, large-memory AI tasks. Meanwhile, Crescent Island focuses on efficient AI inference with 32 Xe3P cores, 256 XMX engines, and up to 480GB of low-cost, power-efficient LPDDR5X memory, optimizing throughput and economics in air-cooled data center environments. Wildcat Lake extends AI acceleration to laptops and edge devices by integrating UCIe chiplet technology, Xe3 graphics, and an NPU delivering up to 17 TOPS, marking Intel’s first mainstream client processor to leverage open chiplet interconnects for scalable, cost-effective multi-chip packaging.

Central to Intel’s agentic AI strategy is the integration of advanced fabrication and packaging technologies that enable scalable, heterogeneous AI systems tailored to diverse deployment scenarios. Utilizing the Intel 18A and 18A-P process nodes alongside innovative Foveros Direct 3D packaging and the UCIe open chiplet interconnect standard, Intel achieves modularity and high performance across its architectures. As CTO Pushkar Ranade emphasized, this approach tightly couples general-purpose compute with purpose-built accelerators and open chiplet technologies to meet real-world constraints of power, cost, and deployment flexibility, thereby future-proofing AI infrastructure from data centers to the edge.

Intel’s Crescent Island GPU exemplifies a strategic shift toward optimizing AI inference economics by eschewing traditional graphics blocks in favor of dedicated AI compute resources. Its Xe3P architecture features 32 cores, each with eight Vector Engines and eight XMX engines supporting advanced precision formats like FP8 and FP4, enabling high token throughput and efficient handling of large models with long context windows. By leveraging up to 480GB of LPDDR5X memory—a cost- and power-efficient alternative to HBM—Crescent Island delivers sustained inference performance within a 350-watt air-cooled PCIe card, significantly reducing total cost of ownership for data center AI workloads.

Intel's hardware integration strategy for agentic AI hinges on a unified memory fabric and advanced memory subsystems that significantly boost inference efficiency and scalability. The Diamond Rapids CPU exemplifies this with its 16 memory channels running at 12,800 MT/s and a substantial 1.28 GB last-level cache, enabling large memory pools and high bandwidth critical for orchestrating complex AI workloads across CPUs and accelerators. This unified memory approach reduces stalls and supports seamless coordination of retrieval pipelines, tool calls, and agent scheduling, positioning Diamond Rapids as a robust backbone for enterprise AI applications.

Intel’s Crescent Island GPU innovatively leverages high-capacity, low-power LPDDR5X memory—scaling up to 480 GB—to accommodate larger AI models and longer context windows within air-cooled data center environments. This memory optimization not only enhances token throughput and concurrency for AI agents but also delivers superior performance-per-watt and cost efficiency compared to competitors relying on HBM memory. By integrating a unified memory hierarchy with 512 KB L1$/SLM per Xe core and a 32 MB unified L2 cache across 32 Xe3P cores, Crescent Island achieves a finely balanced architecture that maximizes AI compute density without traditional graphics overhead, streamlining inference workloads for tokens-as-a-service providers.

Advanced packaging and open chiplet interconnect standards form the cornerstone of Intel’s modular AI hardware vision, enabling flexible and scalable multi-chip systems tailored for diverse AI workloads. By adopting Foveros Direct 3D packaging alongside the UCIe open chiplet interconnect standard—first implemented in the Wildcat Lake SoC—Intel facilitates cost-effective multi-chip package designs that integrate compute, memory, I/O, and accelerators seamlessly. This approach not only enhances scalability from data centers to edge devices but also empowers partners to customize solutions with flexible LPDDR5X memory configurations up to 480 GB, fostering a vibrant ecosystem of ODM-branded AI hardware.

Intel’s strategic partnerships with SambaNova and Terafab underscore its commitment to a holistic AI hardware ecosystem that transcends traditional single-chip solutions. By collaborating with SambaNova on an agentic AI blueprint and leveraging Terafab’s advanced silicon fabrication, Intel is crafting a multi-architecture approach that integrates CPUs, GPUs, and SoCs to address the diverse demands of AI workloads from data centers to edge devices. This ecosystem positioning enables Intel to compete robustly against hyperscalers and other chipmakers by offering modular, scalable AI systems tailored for various deployment scenarios.

Central to Intel’s ecosystem strategy is the emphasis on heterogeneous compute and modular hardware integration, supported by cutting-edge foundry and packaging technologies such as Intel 18A process, Foveros Direct 3D packaging, and UCIe interconnect. These innovations facilitate the creation of adaptable AI platforms that can be fine-tuned for different power envelopes and price points, exemplified by Wildcat Lake’s pioneering use of UCIe in a mainstream client-class processor. This modularity not only enhances flexibility across enterprise and client applications but also reinforces Intel’s position as a systems-level AI solutions provider.

Intel’s Diamond Rapids CPU architecture reasserts the CPU’s pivotal role in agentic AI by focusing on enterprise-scale compute with massive memory capacity, flexible I/O, and composable resources designed to orchestrate complex multi-agent workflows. With up to 256 cores, 1.28GB last-level cache, and support for PCIe Gen6 and CXL 3.0, Diamond Rapids is optimized for tasks such as retrieval pipelines, tool invocation, scheduling, and agent coordination, complementing GPU inference workloads and enabling integrated AI orchestration within data center environments.

Complementing the CPU’s orchestration role, Intel’s Crescent Island GPU targets cost-effective, high-throughput AI inference by offering substantial memory capacity—up to 480GB of LPDDR5X—to support larger models and extended context windows. This design addresses a critical production cost factor in AI infrastructure by keeping more concurrent agent states local to the accelerator within air-cooled data centers, thereby enhancing efficiency and scalability for real-world AI deployments.