Qualcomm is making 64-core chip around custom based cores, probably using Krait cores, and they should be nearly baked by now.
Phytium have announced a processor codenamed Mars that comes with 64 ARM v8 based processors optimised for High Performance Computing (HPC). Eight of the Xiaomi cores make a Panel and Panel connects directly to the DDR3 memory or PCIe interface.
World heavy weight class CPU journalist Andreas Stiller from heise has seen many of the details including the architecture. He said that the Panel shares a 32MB of cache and two Directory Control Unit (DCU) and a routing Cell for internal mesh.
All eight Panels share 128 MB L3 cache and 16 DDR3 1600 channels with 204GB bandwidth available. They share two PCIe 3.0, with 16 lanes with 32GB per second total bandwith.
Given this is a 28nm based processor manufactured at TSMC it is huge. It measures 640 square mm and works at 2GHz. It can produce 512 GFlops with Double Precision.
In single copy SPEC CPU2006, a single core scores 19.2 in an integer test and 17.8 in a Floating point test. With all 64 cores, the Mars processor scores 672 with the Integer and 585 with the Floating point test.
Andreas said that two Xeon E5 2697v3 with 56 logic cores scores only 120 with integer and 857 in Floating point tests.
Mars has 120W TDP which is low for the super compute market. We will see if Qualcomm can do better with its Hydra processor which will end up on a smaller manufacturing node.