Intel's Diamond Rapids Server CPU Debuts at Hot Chips, Betting on AI Agentic Computing
Nashnova编辑部
Intel unveiled Diamond Rapids at Hot Chips 2026 — a 256-core server processor aimed at orchestrating enterprise AI agents — but AMD's server share has climbed to 34.5% and Arm hit a record 13.6%, forcing Intel to answer two questions at once: is the chip competitive enough, and can the fab deliver at scale?
What does Diamond Rapids actually bring to the table?
Built on Intel's 18A-P process — an enhanced version of 18A — with up to 256 cores, 1.28 GB of last-level cache, and 16 memory channels running at 12,800 MT/s.
It also packs 128 lanes of PCIe Gen6 and CXL 3.0. This means → it can attach a large number of accelerator cards and memory-expansion devices simultaneously, acting as the "dispatch hub" for AI-agent systems.
18A-P vs. 18A: 9% more performance at the same power, or 18% less power at the same performance, with thermal resistance improved 20–40% — and fully compatible with existing 18A design rules. In plain terms = fabs don't need to learn a new playbook; existing designs carry over with minor tweaks.
Why is Intel betting on inference rather than training?
Gartner forecasts global AI-optimized infrastructure-as-a-service spending at roughly $42.3 billion in 2026, up 96.4% year-on-year, rising to $66.1 billion in 2027.
The key inflection: inference infrastructure is expected to absorb $23.3 billion this year, overtaking training at $19 billion for the first time. Inference's share rises from 55% in 2026 to 59% in 2027.
This means → enterprises are shifting from "building big models" to "keeping big models working." AI agents require multi-step planning, reasoning, and execution — general-purpose CPUs don't shrink in that world; they grow. Diamond Rapids targets exactly that orchestration layer.
What else did Intel show?
Crescent Island — a data-center inference GPU with up to 480 GB LPDDR5X memory, 350 W power draw, in a PCIe card form factor that supports air cooling. Intel positions it on inference economics, not peak training throughput.
However, no independent pricing or total-cost-of-ownership comparison against Nvidia and AMD equivalents has been published. The economic case remains unproven.
Wildcat Lake (Intel Core Series 3) — a laptop and edge processor on the 18A node, with NPU performance up to 17 TOPS. It is Intel's first processor using UCIe — a chiplet interconnect standard that lets different chip modules snap together like building blocks.
Can 18A-P actually reach volume production? That is the deeper test.
Intel announced in June that 18A-P has entered risk production — in plain terms = the fab is running test wafers, but mass shipment hasn't started; yield and stability are still ramping.
Pushing a 256-core server processor to volume directly tests whether yield is high enough, power low enough, and thermals manageable. Any shortfall erodes the product's cost competitiveness.
This reflects a larger point: Diamond Rapids is not just a chip launch — it is a public exam for Intel's process roadmap.
How far ahead have the rivals run?
Mercury Research data: AMD's server-CPU share rose from 27.3% to 34.5% in one year; Arm-based server processors hit a record 13.6%.
Intel posted $16.1 billion in Q2 revenue, up 25% year-on-year — a better commercial starting point than a year ago. But server-processor shipments industry-wide grew 20% over the same period. This means → part of Intel's revenue gain is the tide lifting all boats, not necessarily share recovery.
Whether Diamond Rapids can reclaim server sockets in a market where AMD has scored heavily and Arm keeps penetrating will be the ultimate verdict of this product cycle.
Content is for reference only, not financial advice.