[OT] A64Fx was: ISO: The Eiffel OO programming language and IDE, on VMS

On Wed, 26 Mar 2025 23:53:14 -0000 (UTC)
Lawrence D'Oliveiro <ldo@nz.invalid> wrote:

On Wed, 26 Mar 2025 23:45:54 +0200, Michael S wrote:
 

Top performing Arm CPUs are designed in USA.
1. Cupertino, California (Apple)
2. Raleigh, North Carolina (Qualcomm)
3. Austin, Texas (Arm Inc) 

 
All outclassed by Fujitsu’s A64FX, as used in the Fugaku machine, and
successors -- designed in Japan.

That's very OT in comp.os.vms, but I'd write it anyway. A64FX did not
outclass any of mentioned above.

A64FX is good chips for its intended applications, i.e. massively
parallel supercomputers. Its most important feature is Tofu comm links
that together with Tofu switches allow high-bandwidth communication
between tens of thousands of nodes with the latency that is not exactly
low, but can be called acceptable. It's second important feature is
built-in connection to HBM2 memory. Its third important feature is
rather good floating-point throughput when the application amenable to
vectorizing. Its fourth important feature is that said good throughput
is achieved without consuming too much power.
However the 4th point should be taken in time context. Consumed power
per FLOP was lower relatively to A64FX 2019 contemporaries. It is
significantly higher relatively to the next generation of GPGPUs. It is
true even if we only look at single and double precision calculations.
For lower precision calculations that are so popular nowadays and
becoming more popular with every passing minute, A64FX is hopelessly
behind state of the art.

Those were strong points of A64FX. Its weak points that make it a poor
choice for general purpose computing are:
(a) - integer performance. No benchmarks published, but description in
various papers could lead as to guess that in benchmarks like
SPECInt2017_speed A64FX core would be be at least 3 time slower than
top offerings of three design houses mentioned above.
(b) - memory capacity. Only 32 GB per 48-core node. For general-purpose
server application it's desirable to have at least 5 times more,
preferably 10 times more.