Does the AI HAT+2 work with Tensorflow (the most common low level software for Deep Learning models)?
Yep, with the caveat that for all Hailo processors you can either use the Hailo model Zoo or " Hailo Dataflow Compiler (DFC)" to convert the files into a Ai Hat 2+ friendly format
Thanks a bunch!
I really tried to RTFM but couldn’t find the FM, neither for the HAT+ 2 nor the HAT+ in any of your docs. I found AI HATs under “Documentation” but that doesn’t tell me anything about “Zoo” or “Hailo Dataflow Compiler”—not where to find or get it, let alone how to use it. I found a tiny little bit on your web site under Documentation → Camera software, but this is—of course—very camera-specific.
But your answer lead me to https://hailo,.ai/products which made the rest a lot clearer. Hailo, and hence the Pi AI HAT+ (2), is mainly for edge computing, i.e. the application of already trained nets, but not really for (re-)training them using TensorFlow, which is what I was trying to do. For that, one or more NVIDIA cards, or with the RBPi formfactor, a Jetson Orin at the very low end are needed. I could have guessed it, as the Hailo NPU uses integer only, the Cuda cores on the Jetson use floating point. It is very tricky (i.e. it doesn’t usually work very well) to train nets using integer arithmetic (I tried using FPGAs—over 30 years ago), but with a little bit of math, trained floating point nets can be converted to nets using only integer arithmetic (especially when they were trained with quantization-aware-training), which is what the Hailo Dataflow Compiler mainly does.
Luckily, for those of us who want to accelerate training on tiny single-board computers, SparkFun also offers an NVIDIA Jetson Orin 
Excellent choice, I have an Orin myself!
The Intel ARC videos cards in a used low-end PC work well too for a somewhat reasonable price