Does anyone have an idea about when we should expect this chips to be incorporated in devices on the generally available market after an announcement like this?
Referring the previous M33 architecture, it took around 2 years to get the actual chips as a public developer if you don't have a priority from the vender. And it took less than 3 years to see consumer products using it.
I am not sure but I think this one might be different. In case of M33 it was fast because you just adopt the chip for the existing solution, but you need to come up with a new business idea, or a technical stack (that uses machine learning on the edge devices) to fully benefit from the new chip.
I've been comfortable designing with M3/M4 and for most applications M33 is not such a big step as to suddenly drop whatever you are doing and redesign your app around new features.
Even with new products I don't necessarily want to spend time learning new functionality since the cores I already know do everything I need.
Given possibility to learn new M33 functionality for a bunch of time I preferred to learn to make my designs more energy efficient with existing cores, for example.
Only just recently I have decided to start using M33 and mainly for security.
I can imagine most people designing go through similar process. The advantage of M33/M55 is/will be when you actually have a need for the new features and when you design your application around those new features, otherwise the cores provide only incremental improvement.
Anecdotally - probably in about 2 years, based on the time between announcement of the core and general silicon availability for past Cortex-M cores (M33, M23, M7, and M4).
These are not "chips" this is just "intellectual property" they are releasing. It will be another couple of years before you start seeing it in silicon, and then another year or so for them to reach the market _as a part of some device_, and gain framework support. You're looking at 2023 for this to be relevant, if all is well.
Well you can predict a little when you switch the scope to its CPU capabilities, not chip vendor platforms. For example, Tensorflow-lite recently added a support for CMSIS-NN (ARM's cuDNN for microcontrollers) so you can run it on the new ARM chips too when they are released.
Helium looks like it might be pretty performant in embedded DSP roles. I'd really like to see how an M55 stacks up against a Qualcomm Hexagon, normalized for silicon area.
M4/M7 was already making convincing arguments against TI/ADI DSPs. If you went and replaced the DSP block inside the ADAU chips (integrated codec + DSP) it would be even more convincing.
The tooling for proprietary DSP ISAs sucks, and with the dominance of ARM and available toolchains for cross compiling various languages over LLVM on basically any OS, it's just so much nicer than dealing with whatever garbage, dated, and poorly documented toolchain/IDE that they charge you to use for their ancient designs.
