Fujitsu releases details on A64FX ARM CPU for supercomputers
A64FX CPU based on ARM v8-A SVE architecture will debut in new supercomputer scheduled for 2021
Fujitsu has released details of its forthcoming A64FX ARM-based CPU that will debut in its Post-K supercomputer scheduled for 2021.
Fujitsu produced a prototype in June and has already begun initial testing. It has now released more basic details about the CPU.
The A64FX is made up of 8,786 million transistors and will be manufacturered to a 7nm process. Fujitsu claims that it will be the first chip to take advantage of ARM's Scalable Vector Extensions instruction set, developed specifically for high-performance computing. Fujitsu claims that it could also be used for artificial intelligence applications.
Sporting 48 cores with two-four assistant cores to assist with processes such as input/ouput, jitter reduction and asynchronous MPI [multiple protocol interface]. For floating-point calculations, Fujitsu claims the A64FX will be able to achieve 2.7 teraflops for 64-bit (FP64) operations, more than 5.4 teraflops for 32-bit (FP32) operations, and more than 10.8 teraflops for 16-bit (FP16) operations.
One of the key benefits, and one of the main reasons for using ARM in supercomputing, according to Fujitsu, is its extremely high throughput and parallelism, combined with low-power consumption, relative to (for example) the Intel Xeon architecture that currently dominates supercomputing.
Specialist supercomputing website Top500 suggested that the performance of the A64FX is unlikely to blow away any of the top-10 most powerful supercomputers in 2021.
"Although floating point performance is certainly respectable, its only about 35 percent faster than the top-of-the-line Xeon Skylake CPUs and more than 20 percent slower than the now-defunct Xeon Phi CPUs," it suggested.
It continued: "It's not hard to imagine that Intel will be producing floppier CPUs in the same 2021 timeframe for its Aurora exascale supercomputer, either with the Ice Lake Xeon CPUs or the unconfirmed Xeon AP processors.
"On the other hand, by providing only low-end performance estimates for the A64FX, Fujitsu may be hinting that it will boost those numbers by the time the final silicon rolls out in a couple of years."
It will take around 400,000 A64FX CPUs, it added, to achieve an exaflop according to the Linpack benchmark or on an intensive floating-point-intensive application.
Fujitsu is more closely associated with the Oracle (formerly Sun Microsystems) Sparc64 architecture. It has been involved in the development of Sparc since version 9 in 1993, deploying the architecture in servers and supercomputers since then.
According to Top500, the A64FX ought to be about 2.5 times fasterer than the Sparc64 XIfx, Fujitsu's previous high-performance CPU.
The company also develops its own GS-series of CPUs for its Japan-only mainframe computers.