Yeah, I would love to be able to buy an ARM server (or a RISC-V server), but so far that's just impossible.
Let's just limit the discussion to Ampere. There is supposed to be Gigabyte G242, with several SKUs, that supports an older version of their chips. I've had extensive email discussions with Gigabyte sales staff, and yet I've been unable to find a single reseller that will even backorder any version of the G242 for me. And there's even a G252 now. Lovely!
Yes, there is a supply chain crisis. Yes, ordering novel products is never easy, nor cheap. But, heck, do I understand why Intel has a lock on the server market...
Availability of their existing dev kits is at best spotty. The AGX Orin [0] dev kit is at least available for backorder at $2k USD - but I haven't seen any of the Xavier kits available for many many months...
Yeah - strong supply constraints right now - side effect of what happened with GPUs... in the first half of 2020 Xavier devkits were very easy to buy at MSRP.
Hey, Intel is trying to get into the foundry business, right? Wonder if Renee has any good contacts over there. Maybe they could dodge the supply chain issue. (Just kidding, I know switching fabs is a big project -- but I do sort of wonder if there was a missed opportunity here... Ampere needs some tricks up their sleeve if they want to compete with Amazon and their unlimited funds, right?).
It wouldn't be possible until around 2024, and regardless TSMC has a ridiculous ARM ecosystem. IFS is probably most competitive with RISC-V and licensed x86 cores rather than ARM. That said, IFS apparently does Graviton packaging, so maybe?
If your talking about the CCR2116-12G-4S+, its a nice router with the AL73400, which is a 16 core A72 platform. OTOH, for a general purpose machine roughly in the same ballpark your probably better off with the NXP LX2160A on something like the solidrun honeycomb. The latter is more suited for general purpose compute given the pcie slot, sata, m.2 slot, pluggable ram, etc. Its also 16 A72's, with the advantage its a systemready certified platform.
But, those are really old generation parts given there are a number of A76 based systems appearing on the market. Those chips are a giant step above the A72 based machines the same way the A72 is a big step above the A5x ones. (aka its a gravaton1->gravaton2 kind of step, or the diff between atom and the mainstream i5/7 parts)
That is going to be the uplift going from a A72->A76+ platform. The IPC is nearly 2x, and translates to going from a early 2000s era x86 to something fairly modern.
however, after using Oracle free tier, I was surprised at the amount of stuff like Docker images that don't support ARM, a lot of images hat to be built manually(GH actions did that for me) or had source images which had x86 binaries!(looking at you bitnami)
in the end I needed x86-64 server to have incompatible images/services on.
If Ampere ever became available, it would make sense for all businesses to host non-realtime work loads on-prem. You might not even notice a change in your electric bill.
Current Ampere servers (even dual socket ones) is available from distributors in Europe and NA with custom configurations in minimum order quantity of 1.
The internet speed and reliability would be poor and you would require in house expertise to maintain the infrastructure even if it is simple. I think it still makes sense to rent them from providers.
> The fact is that the x86 cores have focused so much on performance per core, largely due to per-core licensing costs that dwarf CPU hardware costs, that it has led to a window for a cloud-native processor segment designed to provide the correct level of performance per core tied to cloud vCPUs, but then offer massive numbers of cores.
Why is the x86 core count related to the per-core licensing costs?
If you're paying for software on a per-core basis, you want fast cores, which means big cores. This also works well on desktop since a lot of stuff depends on single-core speeds, so it's basically where x86 has settled in. If you're scaling a workload across many thousands of cores and don't need specialized features, this makes a lot less sense, and smaller/slower cores are very competitive. ARM CPUs like Ampere and dense variants of x86 are targeting this niche.
How ready is Linux (or a BSD; I'm willing to learn!) for home server use?
I understand there will be some growing pains for ARM chips as they venture into server land; the x86 ecosystem with a bazillion different motherboard / component options is much, MUCH further ahead.
What about Docker or kubernetes? How about kvm / libvirt? Are they ready to rock on ARM?
Having a few (mostly idle) x86_64 home servers consuming an unpleasantly high amount of power doing nothing, I really look forward to ARM. Not sure if the power savings would eclipse the savings I get from basically freeloading off of 3-year-old workstations destined for the recycler I grab from my employer, but as someone who frets over power utilization for a major cloud provider every day, I can't wait for ARM to get here!
I have an RPi 4 8GB running Fedora, and everything works fine, including running virtual machines with KVM & libvirt. Small VMs, given the limited memory. The whole thing cost me £238 (actually under £200 because I could reuse one of the SSDs).
Let's just limit the discussion to Ampere. There is supposed to be Gigabyte G242, with several SKUs, that supports an older version of their chips. I've had extensive email discussions with Gigabyte sales staff, and yet I've been unable to find a single reseller that will even backorder any version of the G242 for me. And there's even a G252 now. Lovely!
Yes, there is a supply chain crisis. Yes, ordering novel products is never easy, nor cheap. But, heck, do I understand why Intel has a lock on the server market...