This tied in with a story in August of 2015 where Fudzilla wrote about a MCM Multi Chip module that featured a Zeppelin core, a super-fast 100GB/s interconnection via 4 GMI links and Greenland (Vega) high performance GPU with 4+ TFlops of performance. This APU will still happen, it will just be a bit later – the end of 2017.
Now we have a few more details about Zeppelin cluster and this is proving to be another "Fudzilla told you so" moment. Apparently you can put up to four Zeppelin CPU clusters on a single chip and make a 32-core chip. This will be connected via coherent interconnect (coherent data fabric).
Each Zeppelin module has eight Zen cores and each core has 512 KB of L2 cache. Four Zen cores share 8MB of L3 cache, making the total amount of L3 cache per Zeppelin cluster 16 MB.
Each Zeppelin cluster will have PCIe Gen 3, SATA 3, and a 10GbE network connection. A server version of the chip has the server controller hub, DDR4 memory controller and AMD secure processors.
AMD will have at least three pin-compatible versions of the next generation Opteron using Zeppelin clusters of Zen cores. There will be 8-core versions with a single Zeppelin cluster, a dual Zeppelin cluster version and a quad Zeppelin version. That last option is codenamed Naples for enterprise servers and will have 64MB L3 cache. All of these configurations sound like a lot.
We are expecting to see Zen-based Opterons in eight, sixteen and thirty two-core versions for servers in 2017.