There's really very little "innovation" in the ARM ISA. Or any ISA these days. It is an API. ARMv8 is nice but it's just warmed over RISC -- actually a lot of things that differentiated previous ARM ISAs were removed to make it more like a standard RISC! The value they add is basically nothing except holding the key to being compatible with ARM ecosystem. At this point extracting their tax on their proprietary ISA is close to rent seeking. That drag on the market is why people are looking to real open ISAs.
EDIT: That's not to take away from what architects (ISA designers) do. It's extremely difficult to develop precise and consistent ISAs especially when things like memory ordering, interrupts, pipelined execution, TLBs, etc come in to play. And they are gradually gradually adding features and improving. But this is not value to the tune of over half a billion dollars per year just for ARM licensing (then imagine x86 equivalent numbers for that very locked in ecosystem).
ARM is not in a rosy place with their ISA. Their biggest customers like Apple can and will lift and shift their entire ecosystem if ARM do what it's told. Android could start supporting other ISAs too if someone was motivated enough. And many of ARM's new customers outside the traditional low end embedded world (e.g., in servers) who have been the first ones to move to ARM have done so because their software stacks are very portable. This means they can relatively easily move off ARM as well.
Their logic designs for high performance CPUs don't seem to be top tier either. Not to say that in a disparaging way but for example the PA Semi team Apple bought blew past ARM's high performance mobile designs with their very first ARM core. https://www.anandtech.com/show/6330/the-iphone-5-review/
ARM is pretty decent but behind the likes of Intel, AMD, Apple for high performance cores. So if the ARM market gets bigger and more lucrative especially in the server space, they are vulnerable to other companies selling their own chips and cutting ARM's designs out of the high margin sales. Amazon could be designing their own core right now to use in their ARM servers in a few generations for example. Just as Apple did with its iphone chips. AMD could start making ARM server CPUs if that's where the money is.
> Amazon could be designing their own core right now
Would require an architectural license and lots of talent and time.
Getting stock Neoverse cores is an incredible deal for them, this is probably a big part of the reason they can offer cheaper services from their chips.
Arm cores aren't really all that behind, there are just tendencies for wildly incorrect comparisons (past Arm cores vs upcoming competition cores, scale-targeted configurations vs workstation-targeted ones, etc.)
Of course, the ARM tax. Is there any reason why that would be a problem for them to get?
> and lots of talent and time.
Well just more talent than ARM Ltd.
> Getting stock Neoverse cores is an incredible deal for them,
What's the incredible deal? IP licenses are the other part of ARM's business model.
> this is probably a big part of the reason they can offer cheaper services from their chips.
The problem is when other companies buy the exact same core or customers need single thread performance other designs offer, and then Amazon decides they can make a better design themselves. Like what Apple did.
> Arm cores aren't really all that behind, there are just tendencies for wildly incorrect comparisons (past Arm cores vs upcoming competition cores, scale-targeted configurations vs workstation-targeted ones, etc.)
Traditionally they have certainly been. It was a huge embarrassment that the small PA Semi start up could design their PA6T core in a couple of years from scratch and introduce it at the same time as ARM's A9 (which was a venerable and relatively efficient but pretty poorly performing core considering they had big mobile, PC, and server ambitions).
Things might finally be turning around very recently just with their latest core, but it's difficult to know yet.
EDIT: That's not to take away from what architects (ISA designers) do. It's extremely difficult to develop precise and consistent ISAs especially when things like memory ordering, interrupts, pipelined execution, TLBs, etc come in to play. And they are gradually gradually adding features and improving. But this is not value to the tune of over half a billion dollars per year just for ARM licensing (then imagine x86 equivalent numbers for that very locked in ecosystem).
ARM is not in a rosy place with their ISA. Their biggest customers like Apple can and will lift and shift their entire ecosystem if ARM do what it's told. Android could start supporting other ISAs too if someone was motivated enough. And many of ARM's new customers outside the traditional low end embedded world (e.g., in servers) who have been the first ones to move to ARM have done so because their software stacks are very portable. This means they can relatively easily move off ARM as well.
Their logic designs for high performance CPUs don't seem to be top tier either. Not to say that in a disparaging way but for example the PA Semi team Apple bought blew past ARM's high performance mobile designs with their very first ARM core. https://www.anandtech.com/show/6330/the-iphone-5-review/
ARM is pretty decent but behind the likes of Intel, AMD, Apple for high performance cores. So if the ARM market gets bigger and more lucrative especially in the server space, they are vulnerable to other companies selling their own chips and cutting ARM's designs out of the high margin sales. Amazon could be designing their own core right now to use in their ARM servers in a few generations for example. Just as Apple did with its iphone chips. AMD could start making ARM server CPUs if that's where the money is.