Why MS is going SnapDragon ( for mobile at least ? )

[gilesgoat]

Hi,

soo .. I just seen that MS seems it's going SnapDragon with its "new laptop" ...

I know SnapDragon only from Quest .. ( AND THE MEGA SHIT OF ANDROID ) so I can't say I may have seen "its best".

Is it really that worth it to make "a performant laptop" or "it's a way to save as much power as they can so they can fit some AI chip as well" ?

From what I've seen I dare to say "SnapDragon must be reasonably performant given that it runs Quest/VR reasonably well" ..

But yeah ... can it really replace Intel/AMD ?

[venusian]

It's purpose is to try and match what Apple has done with the M series processors, which is say, get within spitting distance of x64 performance for a fraction of the power consumption. Also to get local inference performance to a place where they can offload servers costs and forget about network latency.

Reports so far seem to be suggesting it'll hit performance targets, but not at Apple's low power levels.

The use cases for local inference are niche as far as I'm concerned.

[GoatFoam]

The atom chipset that microsoft were using in hololens was woefully underpowered. Maybe they finally got the message.

[Dio]

It's purpose is to try and match what Apple has done with the M series processors, which is say, get within spitting distance of x64 performance for a fraction of the power consumption. Also to get local inference performance to a place where they can offload servers costs and forget about network latency.

Reports so far seem to be suggesting it'll hit performance targets, but not at Apple's low power levels.

The use cases for local inference are niche as far as I'm concerned.

Apple's processor isn't more efficient than x86 in and of itself. The advantage they have is a systemic one - it's not just about the processor, it's that by controlling the entire software stack they can make the system more efficient, because they can specifically stop driver X waking up the CPU more often than it needs to, while in the PC environment the owner of driver X might not really care of even know that they're doing it.

As for local inference, in the medium term it's a much more interesting thing that cloud inference (I suspect largely because for local inference costs end after training, while for cloud inference costs continue forever - the widespread language models are already consuming far more power in inference than they did in training).