Official ARM manycore/server thread

[ldesnogu]

Three interesting articles on TheNextPlatform:

Has Arm Discovered the Ecosystem Keys?: this gives an overview of where things are in Arm HPC
A64FX Arm Chip Gets A Big Push From Cray: Cray will build a system with Fujitsu Arm A64fx with SVE
Nvidia Arms Up Server OEMs And ODMs For Hybrid Compute: Nvidia has ported CUDA to Arm-based systems

[ldesnogu]

AWS are officially launching their new ARM-based chip: Coming Soon – Graviton2-Powered General Purpose, Compute-Optimized, & Memory-Optimized EC2 Instances.

The CPU used, Neoverse N1, is the server variant of Cortex-A76 (more info on AnandTech). So no SVE.

[ewmayer]

Originally Posted by ldesnogu

The CPU used, Neoverse N1, is the server variant of Cortex-A76 (more info on AnandTech). So no SVE.

Laurent, to your knowledge, what is the cheapest workstation/server-node hardware that is SVE-capable, and what is the bit-ness of the SVE? 256 and 512-bit would be really useful to play with - I don't know if anyone has a wider ARM-based SIMD implementation, Fujitsu's 512-bit seems to be as wide as they come to date.

[ldesnogu]

Originally Posted by ewmayer

Laurent, to your knowledge, what is the cheapest workstation/server-node hardware that is SVE-capable, and what is the bit-ness of the SVE? 256 and 512-bit would be really useful to play with - I don't know if anyone has a wider ARM-based SIMD implementation, Fujitsu's 512-bit seems to be as wide as they come to date.

As far as I know the only chip in production with SVE is Fujitsu A64fx. I don't think this chip will ever come in an affordable system so I'm afraid we'll have to wait more and rely on QEMU simulation

Or ask nicely Cray and Fujitsu to provide you with an access to such a system?

NOTE: on Firefox all replys get extra empty lines between paragraphs. I just edited to remove them. Is this a known bug (it happens on both Linux and Windows Firefox)?

[ewmayer]

Originally Posted by ldesnogu

As far as I know the only chip in production with SVE is Fujitsu A64fx. I don't think this chip will ever come in an affordable system so I'm afraid we'll have to wait more and rely on QEMU simulation

Or ask nicely Cray and Fujitsu to provide you with an access to such a system?

Well, if there's no prospect of anything other than special-case (code-development, verify run of next M-prime) access to SVE-capable hardware, there would seem little point in spending time porting all the asm-routines to it, aside from personal edification. The only use case I could legitimizing the expected months of coding effort under such circumstances would be a promise from vendor to use code to do whole bunch of GIMPS assignments by way of stress-testing a new supercomputer installation.

I was hopeful that something like 256-bit SVE might be making its way into various consumer hardware like smartphones and PCs at some point in the near future.

[Belteshazzar]

Anandtech's SPEC testing shows Graviton2 coming ahead of x64 EC2 instances for performance per dollar.

[henryzz]

It sounds like some MLucas benchmarks might be interesting on those cores.

[ewmayer]

Apple Aims to Sell Macs With Its Own Chips Starting in 2021 - Bloomberg

Rumors of such a move to ARM-based SOC have been rife for a couple years now. I wonder if the first such chips will use 128-bit-wide vector arithmetic, or 256-bit.

[ldesnogu]

Graviton2 is now publicly available: New – EC2 M6g Instances, powered by AWS Graviton2

[nomead]

Ampere Computing early access program:
https://amperecomputing.com/eap-program-request-form/
No idea about which version of their ARM hardware it is for, though.