Intel data centre AI
Intel AI Accelerators: Gaudi, Crescent Island and Xeon
How Intel’s data-centre AI strategy has shifted from the Gaudi accelerators towards programmable Xe GPUs, rack-scale systems and Xeon CPUs, what ships today, what is only planned, and how oneAPI ties it together.
Export controls: AI chips, servers, networking and memory can need an export licence depending on destination, end user and end use under US, UK and EU rules. We screen every project for this. See what applies.
Where Intel fits
- The value and CPU-led option. Xeon is the host CPU in a large share of AI servers, and Intel is pushing it for CPU-based inference and agent workloads, with Gaudi 3 and new GPUs for accelerator needs.
- A strategy in transition. Intel cancelled Falcon Shores as a commercial product and reset its accelerator roadmap in favour of programmable Xe GPUs and system-level AI infrastructure. That makes roadmap timing the main risk to plan around.
- Ethernet-first networking. Gaudi 3 scales over standard Ethernet, which suits teams that want to avoid proprietary fabrics.
Product by product
- Gaudi 3. Intel’s current AI accelerator. It has been available through cloud partners since late 2024, and Intel still sells it and supports Gaudi 2. A reference rack design offers up to 64 accelerators and 8.2 TB of HBM with liquid-cooling options. Analysts note it has not won meaningful share in large training clusters, so it is best evaluated for specific inference and cost-sensitive workloads.
- Crescent Island. A data-centre inference GPU on the Xe3P architecture, in a 350 W air-cooled PCIe design with native FP4 and MXFP4 up to FP64 support. It uses LPDDR5X memory instead of HBM to cut cost and power. Sources disagree on capacity (160 GB or up to 480 GB), and customer sampling is planned for the second half of 2026, with no launch date.
- Xeon 6 and Xeon 6+. Xeon 6 (Sierra Forest and Granite Rapids, on Intel 3) has shipped since 2024. Xeon 6+ (Clearwater Forest) uses Intel 18A, offers up to 288 efficiency cores, 12-channel DDR5 and 96 PCIe Gen5 lanes with CXL, and was reported available through server makers by mid-2026. Diamond Rapids is planned for the second half of 2026 or later.
- Jaguar Shores. Intel’s longer-term rack-scale GPU architecture with a silicon-photonics focus. It is confirmed as a product, but a release before 2027 is considered very unlikely.
- Ethernet adapters. The Intel Ethernet E835 offers up to 200GbE with RoCEv2 and iWARP RDMA for AI and cloud networks.
Side by side
| Gaudi 3AI accelerator | Crescent IslandInference GPU (Xe3P) | Xeon 6 and 6+CPU with AI features | Jaguar ShoresRack-scale (planned) | |
|---|---|---|---|---|
| Status | Shipping, available through cloud partners since late 2024 | Customer sampling planned for 2H 2026 | Xeon 6 shipping since 2024. Xeon 6+ (Clearwater Forest) reported available through OEMs by mid-2026 | Planned, very unlikely before 2027 |
| Role | Training and inference accelerator on Ethernet | Cost and power-efficient inference | Host CPU and CPU-based inference | Rack-level AI system with silicon photonics |
| Memory | HBM, with racks of up to 64 accelerators and 8.2 TB of HBM | LPDDR5X: sources say 160 GB or up to 480 GB | DDR5, 12 channels on Xeon 6+ | Reported HBM4 |
| Power and cooling | PCIe cards to liquid-cooled racks | 350 W air-cooled PCIe card | Up to 288 E-cores on Xeon 6+ | Not published |
| Precision | Not covered in the sources we read | Native FP4 and MXFP4 up to FP64 | Not covered in the sources we read | Not published |
Several Gaudi 3 and Crescent Island specifications were not available in the sources we read, so we leave them out rather than guess. Ask us for the current datasheets for your project.
oneAPI and the Intel software platform
- What it is. oneAPI is Intel’s cross-architecture programming model and libraries for CPUs and GPUs, with PyTorch support and OpenVINO for optimised inference.
- Upstream-first. Intel describes an open, upstream-first stack, using its Arc Pro B-Series as a development platform for the new GPUs.
- What it means for you. Check framework, model and driver support for the exact part you plan to buy, because Intel’s accelerator lines have changed. See the platform comparison.
- Keeping it secure. Firmware, drivers and runtimes belong in the maintain stage of our agent security and defence practice.
What it means for your data centre
- Air-cooled options. Crescent Island’s 350 W air-cooled card could suit existing racks, unlike the liquid-cooled rack-scale systems from others. See the rack power and cooling guide.
- CPU choices. Xeon 6+ and AMD EPYC compete as host CPUs for any accelerator. We compare them in the server design.
- Roadmap risk. Because Intel’s accelerator plans have been reset, we recommend contracts and designs that can change accelerator without rebuilding the facility.
Where this fits in the stack
Common questions
Is Gaudi still a good choice?
Gaudi 3 is shipping and supported, but Intel has shifted its focus to Xe-based GPUs and rack-scale systems. It suits specific cost-sensitive inference or Ethernet-centric designs, and we advise validating your models first.
What is Crescent Island?
An Intel inference GPU on the Xe3P architecture with LPDDR5X memory and a 350 W air-cooled PCIe design, with customer sampling planned for the second half of 2026.
What happened to Falcon Shores?
Intel cancelled it as a commercial product and kept it for internal use, then added Crescent Island and continued Jaguar Shores as its longer-term GPU line.
Can I buy Intel accelerators for any country?
No. Export rules depend on the chip, destination, end user and end use. We screen every project.
Sources and dating
Figures are as of 18 November 2026, from Tom’s Hardware, Jon Peddie Research, StorageReview and Intel’s newsroom. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against current datasheets when we specify a build. Intel, Xeon, Gaudi, Xe, oneAPI and OpenVINO are trademarks of Intel Corporation. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: Intel AI Infrastructure · NVIDIA Data Centre GPUs · AMD Instinct GPUs · Other Chip Providers · Data Centre Infrastructure
Planning an Intel build
Get the numbers for your workload
