2 ms·
What framework can take advantage of these new chips right now?
by lostmsu 7y ago
What framework can take advantage of these new chips right now?
- m0zg 7y agoThese 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.
- kbumsik 7y agoWell 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.
- m0zg 7y agoBest I can tell, CMSIS-NN is just a software library, and not even a state of the art one at that. No hardware work was involved in supporting that.
- kbumsik 7y ago> No hardware work was involved in supporting that. It is a hardware-specific library that relies Cortex-M DSP instruction. > not even a state of the art one at that What do you expect? Here we are talking about microcontrollers which is always far behind of the cutting-edge computing.
- Symmetry 7y agoHelium 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.
- unlinked_dll 7y agoM4/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.