AIBRIEF
Florence, Ferrara, and Fidenza split AMD's 2028 server CPU roadmap; no core counts or sockets disclosed, and the 'agents per watt' metric is AMD's own thread count proxy.
AMD is splitting its 2028 Zen 7 server-CPU generation into three separately named EPYC families: Florence, Ferrara, and Fidenza, reversing the single 27-SKU stack it shipped for Zen 5. The split gives agentic AI its own product category rather than a feature on a general-purpose server chip.
Florence carries fresh Zen 7 cores, new AI compute extensions, and newer memory, per Lisa Su's keynote at AMD's Advancing AI 2026 event in San Francisco. Ferrara is the AI host node of the Zen 7 stack and the CPU inside the Helios 600 rack alongside MI600 Series GPUs and Pensando Palma and Levanzo networking. Fidenza is the agentic sandbox family: a category AMD now uses for sandboxed deployments of AI agents, software that takes multi-step actions on its own, alongside GPU head node servers.
No core counts, process node, or socket were disclosed; AMD said only that Zen 7 uses "leading-edge process technology." The "most AI agents per watt, per dollar, per rack" pitch is a new yardstick, but AMD's own release says the agent counts are estimated from CPU thread resources as a proxy, not measured throughput. AMD also pegged its 2030 TAM at roughly $2 trillion, a figure AMD itself supplied.
Sixth-generation Venice is already shipping: the 9006 series on SP7 now, Venice-X in 2027 with 1,152MB of 3D V-Cache, 96 cores, and a 5.15 GHz boost clock. The Florence, Ferrara, and Fidenza specs, and which customers sign on, remain the open question.