Ha! OK. That is relieving to hear. I think I readily believed it because of some of the enterprise features companies push for and get.

Then you can't take notes on your PC, reference documents (slide decks, Excel files, web pages, notes) for needed information. This might make sense for only a small subset of types of meetings.

Hello! I am a creative technologist with a background in journalism, and I maintain these feeds as a service to the Bluesky community. In curating the list of organisations included in the feed, I try to keep it to news orgs that produce original journalism and that I can verify the accounts of.

A lot of this depends on how much it will be running in the first place, and then how much power savings are gained (or latency reduced) on an NPU vs a CPU. Today it just doesn't matter so much but it eventually will.

It's good that Microsoft is doing this now. They can see at scale how different applications perform on different platforms with real-world usage. This is also similar to how genAI required a 30+ TOPS (Int8) NPU on a smartphone to run genAI applications, but Google did it with ~4 TOPS.

The direction Microsoft is headed with CoPilot+ will require NPUs, but they aren't really needed right now. Even applications running on newer CPU cores are more efficient because of things like Arm Kleidi and SME2, for example, but are not as efficient as NPUs can be.

As agentic AI applications are used more often or always running in background, a device w/o NPU will be at a disadvantage. GPU is more powerful & will be used as needed, but uses more power consumption - except for newer GPU cores designed with smaller AI compute cores within like embedded NPUs.

If battery life is important, then leveraging the NPU is key. For applications used occasionally, it will not matter. For someone doing many video calls and using eye tracking (to make it look like you are looking at the camera and not somewhere else on your screen (like other people speaking).

Many third party smartphone applications just run on the CPU because the developers do not have the bandwidth/resources to develop for many different NPUs. CPU is easiest to develop software for, GPU is harder, NPU is the hardest. But for many AI workloads NPU is the most efficient.

For clients: my.idc.com/getdoc.jsp?c...

And to make it more fun, NPU TOPS is less important than the size of memory, and the speed of memory - especially the bandwidth between the NPU and memory. That's why many newer SOC designs have doubled the interconnect bandwidth or more, and why some RISC-V companies have orders of magnitude more.

It is widely accepted that agentic AI agents that are always running will have to run on the NPU, and when they need more power they can use the CPU and GPU. And the NPU is required to be a CoPilot (<40 TOPS Int8) or CoPilot+ (>40 TOPS Int8) Windows PC now.

There might just be a little bit of tension at most. It's really up to developers, though SDKs will be steer them. Generally, from APU to GPU to NPU, the workloads are more general to more specific, higher power consumption to lower power consumption, easier to use to harder to use.

I am assuming this would only be for some portion. That might be the higher end (but not the highest end that gets used with dGPU) only, or it might be mixed within the higher end.

For clients of my research, I have a bit more that I say about Apple's chips including the N1:
my.idc.com/getdoc.jsp?c...