Intel outlines three chip architectures aimed at the next phase of agentic AI

Intel has used Hot Chips 2026 to detail three chip architectures designed to support the growing demands of agentic artificial intelligence, spanning consumer devices, edge systems and data centres.

The company highlighted Wildcat Lake, Crescent Island and Diamond Rapids as key parts of its hardware strategy for AI workloads. Wildcat Lake is already shipping in mainstream client and edge systems, while Crescent Island is being developed as an inference-focused data centre GPU. Diamond Rapids, meanwhile, is Intel’s next-generation enterprise CPU and is expected to scale to as many as 256 cores.

Pushkar Ranade, Intel’s chief technology officer, said the rise of agentic AI is influencing how the company approaches processor development at every level. The architectures reflect Intel’s attempt to address different parts of the AI computing market, from local AI applications on laptops to large-scale inference and enterprise workloads.

Diamond Rapids targets high-end data centre workloads

Diamond Rapids is the most extensively detailed of the three architectures, reflecting the importance of enterprise processors to long-term data centre planning. Intel plans to market the architecture as Xeon 7, with its highest-end configuration capable of supporting up to 256 processing cores.

The flagship configuration is expected to include 1.28G of last-level cache, 16 memory channels and 128 PCIe 6 lanes. It will also support CXL 3.0, giving data centre operators additional options for connecting memory and other devices. According to ServeTheHome, the largest configuration uses 16-core chiplets and is built using Intel’s 18A-P process technology.

Intel’s approach puts Diamond Rapids directly against increasingly powerful server processors from AMD. AMD already has a 256-core EPYC 9996 configuration, which is manufactured using TSMC’s 2nm process. Unlike Intel’s flagship design, AMD’s processor retains simultaneous multithreading, allowing it to support twice as many threads as physical cores.

The move towards very high core counts reflects the increasing computing requirements of AI and other demanding data centre workloads. Rather than relying solely on accelerators, modern data centres still require powerful CPUs to manage data, workloads, and the infrastructure surrounding AI systems. Diamond Rapids is therefore positioned as a central part of Intel’s broader enterprise strategy.

Crescent Island takes a different approach to AI inference

Crescent Island could prove to be the most significant of Intel’s three designs for AI inference, although the company has revealed fewer technical details about the accelerator. The product is planned as a 350W, air-cooled PCIe card based on Intel’s Xe3P GPU architecture.

The design features 32 GPU cores connected to 256 third-generation XMX engines, alongside 32MB of unified L2 cache. One notable choice is LPDDR5X memory rather than the high-bandwidth memory commonly associated with AI accelerators.

Intel appears to be prioritising memory capacity, power consumption and cost with this design. The approach could allow the company to compete for inference workloads where customers may place greater importance on the overall cost of running AI models rather than simply maximising peak performance.

However, there is an important distinction around Crescent Island’s advertised memory capacity. Intel has promoted support for up to 480GB, but that figure does not represent the capacity of the standard Intel-branded cards. Those products will be limited to 160GB, while the 480GB maximum will be available to original design manufacturers producing their own versions.

Intel has also not disclosed the GPU’s memory bandwidth. The company told Chips and Cheese that it would not provide the figure. Memory bandwidth can have a major effect on AI inference performance because it influences how quickly data can be processed during model decoding. That, in turn, can affect how efficiently AI services generate tokens and their operating cost.

Wildcat Lake brings Intel’s AI strategy to PCs

Wildcat Lake is the only one of the three architectures already available to consumers. The platform launched as Intel’s Core Series 3 and is being used in laptops and mini PCs, as well as other mainstream client and edge devices.

Intel has changed the packaging approach compared with Panther Lake. Wildcat Lake uses an organic multi-chip package instead of the Foveros packaging used by its predecessor. The change removes the base die, which Intel says can reduce assembly costs and yield losses during manufacturing.

The redesign also reduces the chip’s physical size. Intel says the compute die area has been reduced by 38 per cent, while the I/O tile area has fallen by 15 per cent. These changes could help the company control manufacturing costs while continuing to integrate CPU, GPU, NPU and connectivity functions into a single platform.

The AI capabilities of Wildcat Lake are more modest than those being targeted by Intel’s data centre products. Its integrated NPU provides 17 TOPS of performance, below Microsoft’s 40 TOPS requirement for Copilot+ certification. However, the 40 TOPS requirement applies to the combined capabilities of a system’s CPU, GPU and NPU rather than the NPU alone, meaning a platform can potentially meet the threshold through its overall AI performance.

The distinction remains relevant as Intel attempts to strengthen its position in AI computing. While its consumer processors provide a route into local AI applications, much of the company’s wider AI strategy has focused on enterprise and hyperscale customers. By developing separate architectures for client systems, inference accelerators and high-end server CPUs, Intel is seeking to cover more of the computing infrastructure needed as agentic AI workloads expand.