AMD Unveils Three-Tier Zen 7 EPYC Roadmap for 2028 Agentic Data Centers
CEO Lisa Su announces Florence, Ferrara, and Fidenza families with ACE matrix acceleration and sub-2nm process technology.
AMD is restructuring its server CPU strategy around specialized AI workloads, with CEO Lisa Su unveiling a three-family Zen 7 EPYC roadmap at the Advancing AI 2026 event in San Francisco. The 2028 lineup splits into Florence (primary CPUs), Ferrara (AI Host Nodes), and Fidenza (Agentic Sandbox), marking the company's shift from general-purpose silicon to targeted agentic infrastructure.
ACE and the x86 Alliance
Zen 7 will be the first EPYC generation to support ACE (AI Compute Extensions), a standardized matrix acceleration feature developed through the x86 EAG partnership between AMD and Intel. The move integrates AI acceleration directly into the x86 instruction set, potentially reducing reliance on discrete accelerators for certain inference and training tasks.
"Florence carries fresh Zen 7 cores, a new set of AI compute extensions, and support for newer memory technologies," Su said during the keynote.
Three Families, One Architecture
The Florence family serves as the primary processor line, available in both classic and compute-optimized (Zen 7c) variants. Ferrara targets AI Host Nodes and rack-scale platforms, while Fidenza functions as an Agentic Sandbox built on Ferrara AI systems, optimized for perf/watt in large-scale AI deployments.
All three families will ship on SP7 and SP8 platforms, supporting next-generation memory standards including MRDIMM and LPDDR series. AMD confirmed the chips will utilize a sub-2nm process node, though the company did not specify whether it will tape out on TSMC A16 or A14.
The Agentic Data Center Bet
The accelerated roadmap reflects industry demand for infrastructure that can handle autonomous AI agents rather than traditional batch workloads. By diversifying its product stack while collaborating with Intel on ecosystem standards, AMD aims to compete more directly with NVIDIA's integrated rack-scale solutions. The ACE specification could prove particularly significant if it gains broad software support, allowing developers to write once for x86 AI acceleration regardless of vendor.
What Comes Next
AMD's transition to specialized families signals that the era of one-size-fits-all server CPUs is ending. The 2028 timeline gives the company roughly two years to finalize the sub-2nm implementation and build out the ACE software stack before Florence ships to data center customers.
The success of this strategy will depend on whether workload-specific optimization delivers measurable efficiency gains over general-purpose designs—and whether the x86 EAG partnership can maintain momentum across competing vendors.