Fujitsu A64FX (arm)

[azhad]

Hi,

I read this in my newsfeed today "Little-known Japanese CPU threatens to make Nvidia, Intel and AMD obsolete in HPC market".

The processor has memory on die and according to https://www.fujitsu.com/global/about...8/0822-02.html supports double precision. It's 2.7 TFLOPS and is more energy efficient. Comparatively Intel's latest is 0.7TFLOPS (700GFLOPS) or so.

Shouldn't we be thinking about designing a program for this chip?

[VBCurtis]

Is it for sale yet? Are boards supporting it available yet? How much $$?

Fancy chips are announced a whole lot more often than they are installed in enough systems to be worth considering a port of computation software. But, if by "we" you meant to offer your coding expertise, by all means have at it.

[retina]

Originally Posted by azhad

Comparatively Intel's latest is 0.7GFLOPS or so.

700MFLOPS? Are you sure?

[azhad]

Originally Posted by retina

700MFLOPS? Are you sure?

Typo. 700GFLOPS.

[axn]

Originally Posted by azhad

Shouldn't we be thinking about designing a program for this chip?

ewmayer's mlucas already supports ARM architecture. If you can get your hands on one of these, I'm sure we can get mlucas to run on it.

[ldesnogu]

Originally Posted by axn

ewmayer's mlucas already supports ARM architecture. If you can get your hands on one of these, I'm sure we can get mlucas to run on it.

mlucas will need to be changed to support SVE to use the vector units to their full extent.

Riken should hit the first spot of TOP500 without using any accelerator (such as GPU).

[ldesnogu]

Originally Posted by VBCurtis

Is it for sale yet? Are boards supporting it available yet? How much $$?

Fancy chips are announced a whole lot more often than they are installed in enough systems to be worth considering a port of computation software. But, if by "we" you meant to offer your coding expertise, by all means have at it.

A64fx prototype machines have been in TOP500 since last November.

But it’s unlikely we’ll see workstations with these chips.

[ewmayer]

Originally Posted by ldesnogu

mlucas will need to be changed to support SVE to use the vector units to their full extent.

Riken should hit the first spot of TOP500 without using any accelerator (such as GPU).

Laurent, I did your latest PM re. SVE a couple weeks ago - I'm not ignoring you, just my Mersenne-related time has been full-up busy selfishly upping my own personal GIMPS throughput via different TFLOP-chip means. :)

For the folks around here, though, SVE-enhanced Mlucas will be more interesting in the context of the ARM-based CPUs Apple is moving to, which one hopes will be not only high-performing relative to (say) Raspberry-Pi/Odroid but also support 256-bit-wide vector arithmetic.

[ldesnogu]

Originally Posted by ewmayer

Laurent, I did your latest PM re. SVE a couple weeks ago - I'm not ignoring you, just my Mersenne-related time has been full-up busy selfishly upping my own personal GIMPS throughput via different TFLOP-chip means. :)

Heh no worry, I know how life and other funny pet projects can get in the way