In addition to continuing dominance of the mobile world, ARM architecture has been trying to enter the enterprise realm, including servers, data centers, supercomputers, and more. Amazon, Ampere, Fujitsu, Marvell, Qualcomm, Huawei and so on have developed their own high-performance ARM processors, such as Amazon’s just released 64 core Graviton 2, as well as a variety of coprocessors, accelerators.
Fujitsu A64FX is a typical example, announced last year, and will enter Japan’s Fugaku /Post-K supercomputing, next year put into use, is expected to be at the top of the top of the top of the top of the top OF the top OF the TOP 500 list.
Recently, Fujitsu also made a rare announcement of its wafers, together with the following:
If you know a lot about the chip structure, you can see very clearly a total of 52 cores, of which there are 10 rows at each end, 12 in the middle, eight scattered between them, and a cache with multiple distributions.
Fujitsu A64FX is manufactured using TSMC’s 7nm FinFET process and integrates 8,786 million transistors, but with only 596 signal pins and 52 cores internally, including 48 compute cores, 4 auxiliary cores (all fully consistent), based on arMv8.2-A instruction set, supporting SVE 512-bit width SIMD with peak performance of 2.7TFlops.
All cores are divided into four groups of 13 each, sharing an 8MB secondary cache.
The interconnect bus uses 6D/Torus Tofu, dual links, 10 ports, and bandwidth 28Gbps. The input and output supports 16 PCIe 3.0. The external combination of four groups of 32GB of HBM2 memory, peak read and write bandwidth of 1TB/s.