Intel has launched the Xeon 6+ server processor series. These are the first data center CPUs built on the Intel 18A process. The six-chip lineup tops out at 288 cores. Notably, the announcement is paired with a preview of the Crescent Island GPU with 480 gigabytes of VRAM, a new Ethernet adapter family, and an entry-level server processor.
The lineup of the six Xeon 6+ chips is based on the Darkmont core design and targets AI infrastructure workloads. The flagship Xeon 6990E+ packs 288 cores backed by an L3 cache several times larger than its predecessor. Two of these chips can be combined in a single server for 576 total cores.
The Xeon 6+ launch marks Intel’s first commercial deployment of 18A in the data center. Intel had already previewed the chip, then known as Clearwater Forest, alongside Panther Lake as part of its 18A roadmap in late 2025. Intel is also actively steering OEMs toward the 18A platform by limiting availability of older-generation processors.
RibbonFET puts flexibility at the transistor level
Intel’s 18A process introduces RibbonFET, a transistor design that replaces the conventional channel with several thin stacked sheets. Engineers can adjust the width of those sheets to tune power efficiency and speed, making it possible to tailor individual transistors to specific workloads. That approach underpins Intel’s claim of 30 percent higher power efficiency and performance per thread compared to AMD’s Epyc 9965. The latter’s processors are all built by TSMC.
AMD has been gaining ground in the data center over the past few years, with a quarter of data center CPUs now Epyc-based. Nevertheless, Intel is looking to rebound even coming from a still-dominant market share. A new software tool called Application Energy Telemetry (AET) lets operators monitor per-workload power consumption on Xeon 6+ systems, giving customers data to optimize their applications. More innovations will doubtlessly keep ‘Chipzilla’ competitive as it ramps up its revitalized foundries.
“As AI becomes more agentic, the constraints shift to orchestration, concurrency, and data movement,” said Kevork Kechichan, executive vice president and general manager of Intel’s Data Center Group. “The CPU remains the control plane for the modern AI infrastructure.”
Crescent Island GPU opts for LPDDR5x over HBM
Alongside the CPUs, Intel previewed its next data center GPU series, codenamed Crescent Island. Unlike most data center accelerators that rely on HBM memory, Crescent Island will use up to 480 GB of LPDDR5x. This is a memory type more commonly found in smartphones. That trades some bandwidth for a significantly lower 350 Watt power footprint. Such a figure is rather pedestrian in the days of kilowatt chips from Nvidia. The Intel GPU’s Xe 3P architecture introduces support for MXFP4, a floating-point format that reduces memory requirements for large language models.
The jump to smartphone memory appears to be a smart one, albeit with ramifications for the already dire memory crunch affecting consumer devices. Capacity tends to beat overall speed if a model does not fit in a single memory pool on the GPU. Copying over to another type of storage, commonly between HBM and DDR5, introduces latency that can cripple performance. With 480 GB available, anywhere from 100 to over 400 billion parameters could be loaded. That isn’t quite at the level of the largest open-source models, but it easily fits something like Gemma 4 or GPT-oss.
Intel also introduced the Xeon 6300, a 12-core entry-level server processor, and the Ethernet E835 adapter series. Intel says the most advanced E835 device delivers up to 1.9 times better performance per watt than comparable Nvidia networking silicon.