At Computex 2026 in Taipei, Intel announced its new generation strategy for the data center business and officially released the Xeon 6+ server processor series code-named "Clearwater Forest". A single CPU integrates up to 288 E cores, which is obviously optimized for the current rapidly growing AI training and inference workloads.Intel said that compared with the traditional data center form dominated by general-purpose x86 processors, today's infrastructure is accelerating its evolution to the "AI training and inference frontier." It is expected that in the next five years, AI and traditional data center workloads will account for half, and AI inference will become one of the mainstream scenarios.
Intel emphasized at the press conference that the design focus of data centers in the AI era has shifted from purely pursuing computing power to maximizing performance per watt, performance per core, number of cores per rack, and overall memory bandwidth at the rack and entire computer room levels. At the same time, Intel also admitted that the current data center architecture is no longer "CPU dominates the world", but requires deep coupling of accelerators such as CPU and GPU to jointly undertake large model training and high-concurrency inference tasks.
The expanded server CPU product line is based on Clearwater Forest and uses the Xeon 6+ E-core architecture of the 18A process technology. Intel has previously introduced the potential of this platform in 5G and 6G network evolution scenarios. In terms of specifications, Xeon 6+ can provide up to 288 E-cores based on the Darkmont microarchitecture, and is equipped with up to 576MB of "enhanced low latency" LLC (last-level cache) to meet the needs of high-density computing and large-scale concurrency. In terms of memory, this platform supports 12-channel DDR5 memory with a maximum speed of 8000 MT/s. Intel believes that this high frequency exists in parallel with the strict requirements for high stability of the data center, reflecting its confidence in the reliability of the integrated memory controller.
In terms of the internal structure of the chip, each Clearwater Forest computing tile consists of six modules, each module integrates four Darkmont E cores, so a single computing tile contains 24 E cores. By stacking 12 such computing tiles, Intel has created a flagship model with up to 288 cores to achieve extremely high core density and parallel processing capabilities.
In terms of performance, Intel claims that the flagship model Xeon 6990E+ based on the Darkmont E core is about 30% ahead of the AMD EPYC 9965 in average per-thread performance. In terms of energy efficiency performance, under typical load conditions with a CPU utilization of about 40%, the performance per watt of Xeon 6990E+ can be improved by about 30% compared to the same level of EPYC processors. Compared with its previous generation Xeon 6780E, Intel expects to achieve an efficiency improvement of approximately 55% and an overall performance increase of approximately 126% in similar scenarios.
In order to further tap the potential of energy efficiency, Intel has introduced self-developed Application Energy Telemetry (AET) application energy telemetry technology on Xeon 6+. According to the official introduction, AET can conduct fine-grained energy consumption monitoring of different software workloads at the CPU level with extremely low overhead, including applications, virtual machines and other scenarios. This technology obtains telemetry data directly from inside the chip, bypassing software layers such as the operating system and runtime, thereby reducing additional overhead, and can continuously track the load when computing threads migrate between different cores, providing higher monitoring accuracy. AET will support Linux 7.0 and above, and is mainly used to help cloud service providers and large data center operators conduct more refined energy consumption optimization.
In terms of network, Intel simultaneously launched the E835 Ethernet controller and network card for servers, which can provide up to 200Gbps (approximately 25GB/s) data throughput capability to provide support for high-bandwidth AI clusters and high-speed storage access. In the small and medium-sized enterprise market, Intel has also updated the Xeon 6300 series, launching new products with up to 12 cores for entry-level servers. Compared with the previous 8-core mainstream models, the number of cores has increased by about 50% to meet the needs of lightweight virtualization and business system integration.
In addition to the Xeon 6+ platform that has been mass-produced and is about to be commercialized, Intel also previewed the next-generation server processor code-named "Diamond Rapids" for the first time at this exhibition. Official information shows that the product is expected to be officially launched in 2027 and will continue to be oriented to the AI and data center markets, continuing the development route of high density, multi-core and high bandwidth.
