Intel Launches Xeon 6+ Clearwater Forest — First Data Center CPU on Intel 18A, 288 Cores
Intel's Xeon 6+ Clearwater Forest debuts at Computex 2026 as the first server CPU built on Intel 18A, with 288 Darkmont E-cores and 576 MB L3 cache. Intel claims 30% better per-thread performance than AMD's top 192-core EPYC.
Intel launched Xeon 6+ “Clearwater Forest” at Computex 2026 — the first data center processor built on Intel 18A, the company’s most advanced process node. The flagship Xeon 6990E+ packs 288 Darkmont E-cores and 576 MB of L3 cache in a dual-socket configuration that scales to 576 cores per server. TDP hits 450W at the top SKU, with a 330W variant also available.
Intel’s benchmark claim is direct: 30% faster per-thread performance and 30% better performance per thread per watt versus AMD’s 192-core EPYC 9965. Those numbers will get independently tested quickly, but if they hold, Clearwater Forest marks the most significant Intel-vs-AMD server chip reversal in several years. AMD has dominated data center CPU market share through the Zen 4/5 generations; Intel needs a credible response and is betting Intel 18A delivers it.
The architecture is a chiplet design using Foveros Direct 3D stacking — compute chiplets on Intel 18A, base chiplets on Intel 3, I/O chiplets on Intel 7. That heterogeneous integration is how Intel reaches 288 cores without requiring 18A across the entire die, which would make yields prohibitively expensive. It’s a pragmatic approach, though it adds latency complexity between tile types.
Each socket exposes 96 PCIe lanes. Boost clock tops out at 3.2 GHz, which is low relative to AMD’s EPYC Gen 4 boost clocks, but Darkmont E-cores are dense and power-efficient — a different tradeoff than server cores designed for maximum single-thread frequency. The pitch is throughput and density, not peak clock performance.
Clearwater Forest is targeted at agentic AI inference at scale: running many concurrent model serving requests, not training. For training you still want GPU clusters. But inference at data center scale — the actual revenue workload for most cloud AI deployments — is increasingly CPU-adjacent as models get quantized and distilled down to sizes that fit efficiently on fast CPU memory hierarchies.
Intel shipped production units to customers immediately after the Computex announcement. Pricing has not been disclosed publicly, consistent with standard OEM channel practices for server CPUs.
This is Intel 18A’s data center debut. Foundry services customers (including companies considering TSMC alternatives) will be watching yield data and real-world performance numbers closely.