Raw CPU performance, perhaps not. But people definitely do care about a specific set of user-facing, ML-driven functionality - think speech recognition, speech synthesis, realtime video filtering, and so on.

Many of these are only barely possible on "pre-neural" mobile ARM CPUs, and at a significant cost to power consumption. Developing for newer devices is like night and day.

Not sure that's true to be honest. Speech recognition on my old Pixel 2 is miles ahead of anything I've seen on any iPhone, which are 2-4x faster.

Google's speech recognition is damn impressive, but I'm talking performance/power consumption, not "quality". Sticking a 2080 into an iPhone won't give you better speech recognition results, but it will give you bad results faster.

Because speech recognition quality is a product of data harvesting, the one thing Google does well?

> > Many of these are only barely possible on "pre-neural" mobile ARM CPUs

> Speech recognition on my old Pixel 2

I don't think the Pixel 2 can be called "pre-neural". "[...] The PVC is a fully programmable image, vision and AI multi-core domain-specific architecture (DSA) for mobile devices and in future for IoT.[2] It first appeared in the Google Pixel 2 and 2 XL [...]" https://en.wikipedia.org/wiki/Pixel_Visual_Core

When speech recognition starts understanding European Portuguese without me playing stupid accent exercises, and mixed language sentences as well, then I will care about it.

I suspect access to a vast trove of user data is more important for ML than raw CPU power on the client.

I'm just talking real-time performance and power consumption, not accuracy.