On 2025-05-28 00:53, Petitpierre, Arthur wrote: >> On May 23, 2025, at 11:56 AM, Christian Kastner <[email protected]> wrote: > Graviton4 is a neoverse-v2 core, so armv9.0-a.
> Let me know if you need help to get access to Graviton4-based instances. Thanks for the offer! > I don’t know for the other cloud providers, but on AWS the only way to run > your own hypervisor will be to run on .metal instances (ex: c8g.metal-24xl), > as nested is disabled on virtual instances. Hm, I need to go as low as armv8.0-a as apparently, that is Debian's baseline for arm64 [1]. But there is a simple compromise: Graviton4 for armv9.0-a, and some older but well supported SBC for QEMU+kvm and armv8.N-a. I realize that it might not make too much sense to run llama.cpp on an RPi, but it's a use case upstream supports, so I think the package should also do that. > You can expect most of the server grade arm64 processors built or integrated > by cloud providers from now on to be armv9+ , so from a market trend > prospective I’d say it’s about the right time. Funnily enough, the cloud provider use case for the ARM build never occurred to me, but thinking about it, I can easily see it be the predominant one. Best, Christian [1]: https://wiki.debian.org/ArchitectureSpecificsMemo#arm64
