Intel Corp. today unveiled its most advanced central processing unit so far, targeting artificial intelligence networks and other data center applications.

Announced at the MWC telecom conference in Barcelona, the Xeon 6+ CPU, codenamed “Clearwater Forest,” is targeted at networks and cloud infrastructures and is based on a complex multi-chiplet design that combines 12 compute tiles in an extremely dense core arrangement.

Intel is positioning the Xeon 6+ processor in its AI-ready networking lineup, targeting applications such as edge and on-site AI inference at a time when networks are transitioning to 6G. The company has employed various new technologies, including its high-bandwidth, on-chip fabric 2.5D EMIB links and Foveros Direct 3D die stacking innovations to prime it for the intense demands of the new network standard.

Those technologies are what allow Intel to combine 12 compute tiles manufactured on its most advanced 18A process node (1.8 nanometer) into a single package, together with three active base tiles made on the Intel 3 process (7nm) and two I/O tiles made on the Intel 7 process (10nm). Each tile can be broken down further into six modules that contain four Darkmont E-cores, adding up to 24 compute cores per tile. That means a total of 288 cores per CPU, and 576 cores on dual-socket platforms.

The company said each of those cores benefits from improved parallelism, vector throughput and a beefier instruction cache.

The I/O tiles are made up of two chiplets containing eight accelerators each, plus 48 PCIe Gen 5 lanes, 32 CXL 2.0 lanes, and 96 UPI 2.0 lanes, the company said, referring to its point-to-point interconnects for linking Xeon CPUs in large clusters. That means a single Xeon 6+ CPU features a total of 16 accelerators, including four Intel Dynamic Load Balancers, four Intel QuickAssist Technology, four Intel Data Streaming Accelerators and four Intel In-Memory Analytics Accelerators. In terms of memory, Intel said the Xeon 6+ CPUs can support 12 memory channels and up to DDR5-8000 modules.

Intel said the Xeon 6+ chips deliver double the core count of its previous Xeon 6700E CPUs, resulting in a 17% greater increase in instructions per clock per core, five-times more last-level cache and 20% faster memory speed.

Getting away from the nitty-gritty, Intel said it’s planning to launch the Xeon 6+ processors in the first half of the year to its primary target market of network providers and data center operators. The company said cloud providers will be able to support dozens if not hundreds of virtual machines on a single Xeon 6+ CPU. In networking environments, the chips are targeted at Radio Access Network, 5G core and edge workloads, it said.

Kevork Kechichian, who is executive vice president and general manager of Intel’s Data Center Group, said the new chip will support real-time inference in virtualized RAN deployments. That allows data to be processed by AI models where it lives instead of shifting it back and forth between cloud-based servers.

He said the chipmaker is planning to expand its existing partnership with the Swedish telecommunications giant Telefonaktiebolaget LM Ericsson to jointly develop and market “AI-native 6G solutions.” The details of this are scarce, but Kechichian described the collaboration as “advancing future high-performance and energy-efficient compute architectures for both AI for networks and networks for AI.”

According to him, AI-native 6G combines intelligent and programmable networks with advanced compute and real-time sensing. These capabilities will “underpin more responsive, efficient and capable services, and ultimately result in closer integration between sensing and compute.”

Images: Intel