It failed.
But I think that they learned something really valuable. As a company, they don’t have a good understanding the server market or how to sell to it. Their premium product style just didn’t have a place in the race to the bottom server market.
We had two in my grad school lab that we used as web servers. But that was the problem - it was a premium designed product that wasn’t any better of a web server than a plain (cheaper) Dell 1U box.
Maybe things would be different now that they have a really unique chip. ARM as a whole ecosystem has more buy-in from other companies than PowerPC did. Their performance per watt and GPU design might make more of a difference this time around.
But I can’t see Apple being able to release and sell a data center centric server. I think others might have a better chance.
MacOS is not a server OS either and actually maturing M* support in Linux will take years. And for inference Nvidia is years ahead of Apple.
So best Apple can hope for is some niche market.
Apple doesn't have the business to business support organization to make this kind of thing work either.
Libdispatch is a burden to use for a server because of its book keeping overhead, and their history with kqueue makes it hard to convince someone to use it.
There’s no io_uring, but in practice, almost all servers run epoll anyway, don’t they?
MacOS’s implementation is an order of magnitude or more slower than (say) FreeBSD though, so its not a valid comparison that the syscall simply exists.
Another example of a syscall performing differently would be gettimeofday() in OpenBSD vs linux, its an extremely cheap operation in Linux but quite costly in OpenBSD despite being identical at the kernel boundary.
A mega-AI workstation could.
But server? Nah.
The issue is that the current Apple way of designing hardware doesn't really work for servers. An individual 1U server can easily cost $50,000 or more. Soldering everything together so that if a single memory chip goes bad you have to replace the entire machine instead of that one stick is going to make customers unhappy. Likewise, they expect to be able to add storage or network ports, and expect the software to have enough openness that the new systems will work with their existing enterprise SAN or what have you.
Obviously Apple could make open server hardware, but then people might start wondering why they can't make open laptop hardware.
The primary advantage of soldering compared to DIMMs is that the traces can be slightly shorter without routing them through an edge connector, but if you need that then you can use CAMM.
They solder the memory because then you can't upgrade it. Strong evidence for this is that they do the same thing with the SSD, for which there is no even plausible other justification.
> Strong evidence for this is that they do the same thing with the SSD
Worse, they have a proprietary module that they use in the Mac Mini. Couldn't have just used M.2.
If they just did motherboards with cpu/ram/pcie and let other OEMs build complete systems the story would be very different
All the capacity at all the fabs is booked up for years. They can’t make more if they want to.
So any hypothetical sales of OEM style guts would only cut into the number of Macs they can sell to existing customers. And there’s no way they would ever do that.
Higher margins?
42 people with iPhone will buy airpods, some subscriptions and apps. They also be vendor locked into Apple ecosystem and almost guaranteed to bring more iPhone sale in 3 years.
Apple must be dumb to canibalize their own market.
Now they have Apple Silicon.
I think most people here never seen the inside of sliding pizza boxes on server racks, really seeing them.