New Xeon 6+, Ethernet E835, and Crescent Island GPU: Intel Updates AI Platform for Datacenters

At Computex 2026, Intel announced a series of updates for the datacenter market, which include the new Xeon 6+ processors, an expansion of the Intel Ethernet 800 Series networking solutions with the E835 controllers and adapters, a new model called Xeon 6300 with 12 cores aimed at entry-level servers, and further details on the upcoming datacenter GPU known by the codename Crescent Island.

The company links these developments to the growing spread of agentic AI, an approach involving AI systems capable of autonomously coordinating multiple activities, services, and models. According to Intel, in this scenario, the role of the CPU is once again central as an orchestrating element, managing concurrency and moving data between accelerators, memory, and network.

"With the evolution of AI towards increasingly agentic models, constraints are shifting towards orchestration, concurrency, and data transfer," said Kevork Kechichan, Executive Vice President and General Manager of Intel's Data Center Group. "The CPU remains the coordinator of modern AI infrastructure."

Xeon 6+ Debuts on Intel 18A Manufacturing Process

The main news revolves around the arrival of the Intel Xeon 6+ processors, codenamed Clearwater Forest, which expand the Xeon 6 family with a particular emphasis on computational density, energy efficiency, and cloud-native and AI workloads.

This is the first datacenter processor from the company built using the Intel 18A manufacturing process. Intel claims that the Xeon 6+ is designed to maintain high performance even within stringent energy limits, a feature deemed essential in modern datacenters.

Among the main specifications are:

Up to 288 Efficient cores;
Up to 12 channels of DDR5 memory;
96 PCI Express 5.0 lanes;
Support for Compute Express Link (CXL);
Integrated security technologies such as Intel SGX and Intel TDX.

According to company data, the Xeon 6+ can offer up to 2.5 times the performance of the previous generation and up to 45% higher thread performance per watt compared to competing solutions. Intel also claims a consolidation capacity of up to 9:1 compared to servers based on second-generation Xeon processors. Essentially, if 9 servers were previously needed to achieve a certain performance level, now only one is required.

Another feature is Intel Application Energy Telemetry (AET), a technology that enables real-time monitoring of energy consumption and CPU activity at the individual workload level, providing a more detailed view of operational efficiency.

The new Xeon 6+ are already being validated with telecom operators and server ecosystem partners, including ASUS, Dell Technologies, Ericsson, GIGABYTE, HPE, Lenovo, and Supermicro.

Ethernet E835: Up to 200GbE for Cloud, Edge, and AI

Intel has also expanded its networking offering with the new Ethernet E835 family, consisting of controllers and adapters designed for cloud, edge, enterprise, and AI infrastructures.

The series supports speeds of up to 200GbE and a variety of port configurations, including 2x25GbE, 4x25GbE, 2x100GbE, and 1x200GbE. Additional configurations can be enabled via the Intel Ethernet Port Configuration Tool (EPCT).

The company places particular emphasis on energy efficiency. Based on internal benchmarks, the Intel E835-CQDA2 adapter would offer a performance per watt ratio up to 1.9 times higher than NVIDIA ConnectX-6 DX and up to 1.4 times higher than Broadcom BCM957508-P2100G.

Key supported features include:

RDMA over RoCEv2 and iWARP;
Dynamic Device Personalization to optimize packet processing;
Hardware Root of Trust;
Digitally signed SPDM protocol;
Compatibility with Linux, VMware ESXi, and Windows;
Lifecycle exceeding 10 years.

Intel lists Cisco, Dell, HPE, Lenovo, and Supermicro among the partners who will support the new platform.

12-Core Xeon 6300 for Entry-Level Servers

For the segment of servers aimed at small and medium enterprises, Intel announced the general availability of a new 12-core variant of the Xeon 6300 family. This marks the first time the company’s entry-level platform exceeds the 8-core threshold. The new processor is designed to offer greater processing capacity without requiring infrastructure changes, being compatible with existing motherboards and systems.

According to Intel, this feature should allow for quicker and less expensive upgrades for businesses looking to boost their server performance.

Crescent Island Focuses on Memory Capacity and Efficiency

Intel also provided further details on the upcoming Crescent Island datacenter GPU, an accelerator specifically developed for next-generation AI workloads. Based on the Xe3P architecture, Crescent Island is designed to address what Intel identifies as the main bottlenecks of agentic AI: memory capacity, bandwidth, and energy efficiency.

The GPU will integrate LPDDR5X memory with capacities of up to 480 GB and adopt a 350-watt air-cooled PCI Express design. The stated goal is to support large models and workloads characterized by a high token count while maintaining a competitive performance per watt ratio.

On the numerical precision front, Crescent Island will support a wide range of formats, from FP4 and MXFP4 to FP64, along with new features dedicated to AI operations and improved memory scalability. Intel also highlights compatibility with its open software stack for AI and continuity with the existing Xe ecosystem. In this context, the Arc Pro series GPUs are indicated as valuable development platforms for creating, validating, and optimizing applications that can later be deployed on Crescent Island while maintaining full compatibility.

Lastly: Diamond Rapids

Intel also discussed Diamond Rapids: the upcoming family of Intel Xeon processors that will succeed current architectures, positioning itself as a key pillar of the company's strategy to regain absolute leadership in the high-end server and cloud market. The most significant news is the official launch date set for 2027, along with the adoption of the Intel 18A-P manufacturing process, an even more optimized and "Performance-oriented" version of the already known 18A.

Intel promises a flexible yet optimized System-on-Chip architecture to maintain uniform memory latency (UML), reducing typical bottlenecks of complex multi-chiplet solutions. The image clearly shows an advanced chiplet design, consisting of two large central die (presumably the compute die with the cores and associated caches) alongside four modules arranged laterally (dedicated to memory or ultra-high-speed I/O).

As modern applications, especially those related to Artificial Intelligence, High-Performance Computing (HPC), and in-memory databases, are constantly bandwidth-hungry, Diamond Rapids promises a doubling of memory bandwidth and an increase in the number of channels. Support for the PCI Express 6.0 standard is also confirmed, ensuring extreme speeds and unprecedented scalability for managing external accelerators, GPUs, and next-generation NVMe storage, ideal for heavy I/O workloads. Lastly, the new project foresees 50% more cores than the previous generation, with greater performance per thread.