On 24 August 2026 Intel used its slot at Hot Chips, the annual semiconductor architecture conference, to present three chips it says will carry "agentic AI" from the data-centre rack to the laptop. None is a direct rival to NVIDIA's flagship GPUs. That is the point.
Intel's argument, in the words of its chief technology officer Pushkar Ranade, is that agentic AI is changing chip design "from the transistor and package up through the full system architecture." Instead of one giant accelerator, Intel proposes assigning each stage of an agent workflow to different hardware.
Key takeaways:
- Diamond Rapids (next-gen Xeon, to be branded Xeon 7): up to 256 cores, 16 memory channels, 128 PCIe 6.0 lanes, CXL 3.0. Built on Intel's 18A-P process. Now slated for 2027.
- Crescent Island (data-centre inference GPU): up to 480 GB of LPDDR5X memory instead of HBM, 350 W, air-cooled PCIe card, 32 Xe cores and 256 XMX engines, FP4 and FP8 support.
- Wildcat Lake (Intel Core Series 3, already shipping): up to 6 CPU cores, Xe3 graphics, a 17 TOPS NPU; first Intel processor to use the UCIe chiplet interconnect.
- All three are manufactured by Intel Foundry on the 18A family, using Foveros Direct 3D packaging.
Three layers, three jobs
| Chip | Layer | Job | Key spec |
|---|---|---|---|
| Diamond Rapids | Orchestration | The "control plane": coordinating agents, memory-heavy enterprise workloads | Up to 256 Panther Cove P-cores |
| Crescent Island | Inference | Serving many concurrent agents with long contexts, cheaply | Up to 480 GB LPDDR5X at 350 W |
| Wildcat Lake | Client and edge | Local AI on mainstream laptops and edge devices | 17 TOPS NPU, UCIe |
The Crescent Island gamble
The most unconventional decision is memory. Every flagship AI GPU uses High Bandwidth Memory (HBM), which is fast, expensive, power-hungry and in short supply. Crescent Island uses LPDDR5X, the memory found in phones and laptops. It is slower per byte but far cheaper and cooler, which is how Intel fits 480 GB on a single air-cooled card.
The bet is that for many inference workloads, capacity matters more than bandwidth: a bigger model, a longer context window, more agents served at once. Intel has not yet published Crescent Island's memory bandwidth, and reviewers note that the Intel-branded card may be limited to 160 GB, with 480 GB reserved for partner cards. Until independent token-throughput measurements exist, the claim is a hypothesis.
Diamond Rapids: big, late
Diamond Rapids is a chiplet design: 16 core chiplets on 18A-P, each with 16 cores, plus base tiles and fabric hubs, assembled with Foveros Direct 3D. It adds Advanced Performance Extensions (APX) and enhanced Advanced Matrix Extensions (AMX) for AI work. Reviewers point out it matches AMD's EPYC Venice on cores and bandwidth but arrives in 2027 with half the threads.
Wildcat Lake: the quiet milestone
Wildcat Lake is already shipping as Core Series 3. Its 17 TOPS NPU is below the 40+ TOPS threshold Microsoft requires for Copilot+ PC branding, so this is a mainstream part, not a halo product. Its real significance is UCIe, an open chiplet-interconnect standard, appearing in a volume consumer processor for the first time from Intel. That signals chiplet economics reaching ordinary AI PCs.
Concept to take away: heterogeneous computing. No single processor is best at everything. CPUs excel at branching logic and coordination; GPUs at massively parallel maths; NPUs at low-power local inference. Intel's Hot Chips pitch is that an AI agent workflow is really three workloads, and each belongs on different silicon. Whether Intel wins or not, that framing is how systems are increasingly designed.
Note: everything here is architecture disclosure, not benchmark results. Diamond Rapids and Crescent Island are not yet shipping. Treat specs as design targets until independent testing appears.
