3 ms·

Using ARM processors combined with their own custom silicon will give Apple the potential to add features that can't readily be replicated by Intel-based laptop

Using ARM processors combined with their own custom silicon will give Apple the potential to add features that can't readily be replicated by Intel-based laptops (as we already see in the phone market). Does the same logic then hold in the datacentre? With the ability to add their own IP mean that AWS, Google etc can start to add new features (e.g. specialised accelerators) that would not be possible with Intel CPUs?

detaro 6y ago
To a degree. But afaik Intel already will do custom extensions to Xeons for them and for many, especially larger-scale, accelerators it makes sense to not have them on the main CPU, so they have their own cooling and power delivery etc (although I don't know how open Intel is about giving them low-level access to system resources, e.g. the QPI bus, whereas a fully custom design could add the interfaces needed freely)
Traster 6y ago
I think the big difference between what Apple does and what happens in the data centre, is that Apple builds custom silicon along with a full technology stack on top of it. They aren't throwing some ML specific stuff on their chip and then just hoping someone uses it - they've got specific applications (voice recognition, face recognition etc) that they're targeting. It's much more difficult in the data centre where you have customers who are going to be writing their own code. You have to build something very generic, you have to build an API, you have to create a software eco-system of libraries and examples and a community of users, and even then lots of people will be worried about lock-in. It's absolutely possible, and Google are already doing custom silicon with the TPUs, but it's a very different challenge.
- klelatti 6y ago
  Fair point but there are examples where customers could be using AWS / Google API's that could hide the new hardware (voice recognition would be a good example). I guess too that for Apple there is a huge advantage to having the custom silicon on the same SoC as the CPU whereas there is much less penalty for having it on a separate chip (as per TPUs) in the datacentre.

detaro 6y ago

To a degree. But afaik Intel already will do custom extensions to Xeons for them and for many, especially larger-scale, accelerators it makes sense to not have them on the main CPU, so they have their own cooling and power delivery etc (although I don't know how open Intel is about giving them low-level access to system resources, e.g. the QPI bus, whereas a fully custom design could add the interfaces needed freely)

Traster 6y ago

I think the big difference between what Apple does and what happens in the data centre, is that Apple builds custom silicon along with a full technology stack on top of it. They aren't throwing some ML specific stuff on their chip and then just hoping someone uses it - they've got specific applications (voice recognition, face recognition etc) that they're targeting. It's much more difficult in the data centre where you have customers who are going to be writing their own code. You have to build something very generic, you have to build an API, you have to create a software eco-system of libraries and examples and a community of users, and even then lots of people will be worried about lock-in. It's absolutely possible, and Google are already doing custom silicon with the TPUs, but it's a very different challenge.

klelatti 6y ago
Fair point but there are examples where customers could be using AWS / Google API's that could hide the new hardware (voice recognition would be a good example). I guess too that for Apple there is a huge advantage to having the custom silicon on the same SoC as the CPU whereas there is much less penalty for having it on a separate chip (as per TPUs) in the datacentre.