AMD has confirmed Zen 7 and Zen 8, lining up EPYC Florence for 2028 and Ravenna for 2030.
AMD has only just shown off its first EPYC chips based on Zen 6, but it is already waving the next two CPU roadmaps about at Advancing AI 2026.
Chief executive Lisa Su ended the event with a peek at Zen 7 and Zen 8, after a session packed with AI kit from CPUs to GPUs, racks and software stacks.
The AMD EPYC 9006 series server CPUs sit in a portfolio aimed at agentic AI, cloud, enterprise, high-performance and general-purpose computing. AMD used the event to push them as the brains behind the agentic data centre.
The more interesting bit was the arrival of two fresh CPU architectures, Zen 7 and Zen 8, which keep AMD’s EPYC march ticking along.
Zen 7 is aimed at 2028 and will land in the seventh-generation EPYC Florence family. As with previous Zen generations, AMD will offer standard Zen 7 and compute-optimised Zen 7C flavours.
Florence is expected to use a next-generation process node, likely somewhere below 2nm, with TSMC A16 or A14 looking like obvious candidates. AMD did not spell that bit out, which is the usual corporate trouser dance.
The chips will be the first EPYC family to support new memory standards across MRDIMM and LPDDR lines, probably including DDR6. They will run on existing SP7 and SP8 platforms.
The SP7 versions are being pitched for leadership core counts and performance, while SP8 is being aimed at better performance per system dollar.
EPYC Florence will act as the host CPU for AMD’s next AI rack system, codenamed Ferrara. AMD is developing an agentic sandbox based on Ferrara AI systems called Fidenza, which targets large-scale AI deployments with better performance per watt.
Ferrara will not replace Helios. AMD plans an updated Helios 500 rack-scale system in 2027 using EPYC Verano CPUs, Instinct MI500 GPUs and Pensando Como and Monza chips.
Zen 7 EPYC parts will be AMD’s first EPYC chips to support ACE, or AI Compute Extensions. ACE is part of the x86 Ecosystem Advisory Group, the Intel and AMD effort to keep x86 looking useful while Arm vendors keep kicking the tyres.
The idea is to offer standardised matrix acceleration features across the x86 ecosystem. Chipzilla and AMD agreeing on anything useful is rare enough to be worth noting.
AMD gave its Instinct GPU roadmap a shove too. The MI500 series is due in 2027 with CDNA 6, HBM4e memory, a larger GPU scale-up domain and designs ready for copper and optical links.
Those chips are aimed at Nvidia’s Rubin Ultra GPUs, which are expected to appear in Kyber racks. The silicon handbag fight at the top of AI compute is getting expensive.
AMD is working on CDNA-Next for the MI600 series, targeting 2028. MI600 should line up against Nvidia’s Feynman chips and will slot into Helios 600 AI rack-scale systems.
Those racks will pair MI600 with EPYC Ferrara CPUs and Pensando Palma and Levanzo chips. AMD claims its MI500 series can deliver a “>2,000X Increase in 4 Years” in inference throughput, which is the sort of chart line marketing departments frame for the office wall.
Zen 8 is already in development and will power AMD’s eighth-generation EPYC family, codenamed Ravenna. That one is pencilled in for 2030, assuming the roadmap gods and foundry capacity behave.
