Zen 5 chiplet architecture — shipping now

Every core matters.
Every link is infinite.

AMD builds processors out of chiplets — small, specialized dies connected by Infinity Fabric — so a laptop, a workstation, and a server rack can all draw from the same silicon, scaled to exactly what the job needs.

5Zen generations since 2017
128Cores in a single EPYC socket
3DV-Cache stacking, in production
4nmprocess on current CCDs
Processors

One design language, split into the right pieces.

Compute chiplets, I/O dies, and cache stacks combine differently depending on the job — a mobile chip built for battery life, a desktop chip built for frame rates, a server chip built to fill a rack with cores.

Ryzen 9 — Desktop

For builders & creators

Two compute chiplets (CCDs) around a central I/O die, one topped with a 3D V-Cache stack — for heavy multitasking, rendering, and competitive frame rates.

Cores16 (32 threads)
Max boost5.7 GHz
3D V-Cache128 MB
Ryzen AI 9 — Mobile

For thin-and-light everyday use

A single monolithic die — no chiplet links to power — tuned for battery life, with an on-die NPU for local AI workloads on the go.

Cores12 (24 threads)
Max boost5.1 GHz
NPU50 TOPS
EPYC 9005 — Data center

For the rack, not the desk

Up to twelve compute chiplets ringed around a central I/O die — density and memory bandwidth built for AI-scale infrastructure.

Coresup to 192
Memory12-channel DDR5
TDP class500W
Architecture roadmap

Five Zen generations, one steady cadence.

Each generation is a real jump in instructions-per-clock and efficiency. This is the sequence, in the order it shipped.

2019

Zen 2

First chiplet design at scale. Shipped in Ryzen 3000.

2022

Zen 4

Move to 5nm and DDR5. Shipped in Ryzen 7000.

2023

Zen 4c

Density-optimized core for EPYC Bergamo's core count.

2024

Zen 5

Rebuilt front end & wider execution — current generation.

2026

Zen 6

Next architecture, in design and validation today.

12chiplets addressable on a single EPYC package
3.9xlower defect cost from splitting large dies into small ones
2 TB/saggregate Infinity Fabric bandwidth per socket
3process nodes mixed on one package (compute + I/O + cache)

Let's build your next platform on AMD.

Talk to our team about processors, EPYC deployments, or volume orders.