Why ARM though? The article touts how this is a homegrown chip, and Amazon obviously has the resources to build a truly homegrown, optimized CPU. Why use ARM instead and import all of its idiosyncrasies?
I guess I could ask the question more broadly. Why does every company that "designs its own chip" use ARM instead of designing its own ISA? How much work does it save? How much optimization does it forsake? I'm reminded of John Regehr's post on discovering the optimal instruction set: https://blog.regehr.org/archives/669
It's not just the ISA that you get when you license from ARM: it's the entire HDL source. If you go even further and get an architecture license, you also get a whole bunch of fundamental technical documents and a test suite. This saves a ton of time.
There's the issue of software compatibility. Writing a port of Linux—kernel and userland—is a ton of work. And if you have other chips on board, such as a GPU, the only drivers available may be ARM-only binary blobs. There might be an entire ecosystem of ARM software you have to keep supporting (which is one reason why Apple doesn't [yet] design their own ISA).
RISC-V has a small chance of commoditizing some of this stuff someday, but not yet.
Designing your own incompatible non-mainstream ISA basically commits you to maintaining your own compiler toolchain(s) and port of Linux distribution(s). That's dozens to hundreds of software engineers, continuously.
> How much optimization does it forsake?
Not a great deal, unless your workload is very unusual or you spend all your time running one algorithm.
It seems that what Amazon are doing is adding extensions or changes to facilitate hypervisors - not changes to the instruction set at all, but changes in the way system calls, hypervisor calls, and interrupts are handled.
Because it is a lot of work to build from scratch, and build the whole software ecosystem. Risc-V is the other option but it is not yet as mature for server chips, and the software stack is only just starting to ship. Arm servers have only just reached maturity after some years.
Really a lot of work. Software compatibility and availability will make or break a processor. Unless you own the entire stack and are willing to deal with the struggles of a custom ISA, nobody wants to make a new one - the benefits probably aren't that great.
This can't be overstated for a cloud provider, especially one like AWS that wants to run everything for everyone. ARM has good OS and compiler support, and if you want people to move some workloads off x86 to another architecture, you're gonna want that migration to be as painless as possible.
Aarch64 (does this chip even bother supporting 32-bit?) is much less idiosyncratic than the older 32-bit arm. Also, arm server platform has UEFI (for all its faults), providing a x86 like plug and play experience. And, like others already said, they are getting an entire ecosystem as part of the deal.
I guess I could ask the question more broadly. Why does every company that "designs its own chip" use ARM instead of designing its own ISA? How much work does it save? How much optimization does it forsake? I'm reminded of John Regehr's post on discovering the optimal instruction set: https://blog.regehr.org/archives/669