If you were told a few years ago that a supercomputer with chips based on the ISA ARM they were going to occupy the first position of the TOP500 (list of the 500 most powerful supercomputers in the world), the laughter and laughter would have been loud. Nobody imagined that an architecture that was practically not used would get there.
Little by little, ARM chips have been gaining ground, conquering the field of mobile devices for their efficiency and performance, as well as many other embedded products. But a few years ago they began to create some servers with ARM low consumption, and began to flirt with this ISA in the HPC (High Performance Computing) sector.
The news that jumped recently about that Apple abandons Intel To create its own ARM-based chips was important, for all that it entailed, but it went almost unnoticed by other, much more important news. And it is that an ARM supercomputer could beat the performance of the IBM Summit and conquer the first position of the Top500 list. For the first time an ARM reaches so high, and that has meant a before and after ...
Going back to Apple, it seemed strange that an ARM chip could surpass Intel in efficiency and performance, but Apple has proposed it and will be able to make very interesting designs. Intel is getting more and more complicated, and not just because of the competition from AMD ...
But what you would never imagine a long time ago is that can also be crowned at the HPC. Do you want to know more details? Well, on the Top500 list for June 2020, the top performance spot is held by the Japanese supercomputer Fugaku. A supercomputer based on 64Ghz Fujitsu A48FX 2.2C chips, of which it incorporates 7.299.072 processing cores to add a beast of floating point computing performance.
Specifically reaches 415,5 PFLOPS (that is, 415.500.000.000.000.000 calculation operations with decimals per second) and will be used for research against SARS-CoV-2, among other things.
It is installed at the RIKEN Computing Center in Kobe, Japan. In this data center their more than 150K nodes of which it is composed, interconnected through a high-speed Tofu Interconect D network to link the ARMv8.2-A SVE chips of 52 cores per node.
Also use the memory high-bandwidth HBM2 with a capacity of 32 GiB per node. Storage wise, it has 1.6 TB NVMe shared per 16 nodes, as well as 150 PB Shared FS, and additional cloud storage service.
As operating system use linux, specifically the RHEL 8 distribution, as well as an IHK / McKernel simultaneously. All performance simulations were measured under McKernel, although Linux is present to provide the rest of POSIX services.
The processing beast that has yielded those figures is quite "humble." It is a chip created by Fujitsu. It's called A64FX and it is a microprocessor based on the ARM 8.2A architecture, also adopting SVE (Scalable Vector Extensions), additional extensions to that base ISA to achieve better calculation results.
The A64FX that has Fujitsu designed it thus replaces its previous SPARC-based HPC chips. And they have not only set a milestone for taking the Fugaku to the top of the Top500, but also for being the first to support 512-bit SIMD EVS.
These chips have been manufactured in TSMC factories, the same where they manufacture the Zen of AMD, and the same where they will manufacture the future chip of Europe. The technology they have used 7nm technology to build their 8.786.000.000 transistors. All of that in a small chip that only needs 594 pins.
Additionally, each processor uses 32GB of HBM2 memory with a 1TB / s bandwidth, with 16 lanes or PCIx lanes per processor to connect them with accelerators, such as GPGPUs and FPGAs.
Finally, works at 2.2 Ghz and enough packages have been added to complete that figure of almost 7.3 million cores and almost 5 PB of memory.