alt Hacker NewsSome of this stuff you need to probe with very, very expensive equipment and fine-tune parameters, and to make it worse, RF performance also depends on the composition of the PCB layers, the amount of vias, hell even just rotating an unrelated component on a PCB can affect the RF performance of nearby components or traces. If you're into that, Hans Rosenberg has some darn good yt videos [3].
RF is a world of black magic, especially at the frequencies, symbol rates and encodings used for stuff like RAM. And the higher in frequency you go... the less "conventional wisdoms" apply.
There's in fact a very old article from 2007 [1] describing the issue from the other end. Some researcher tried to have a primitive form of what we'd call "machine learning" a few years later write FPGA bitstream to get a tone discriminator. Turns out the algorithm and test harness got a working bitstream... but it made no sense at all, it was very finely tuned to individual physical characteristics of that chip.
Link training blobs on modern chips do something very, very similar, at each link initialization they evaluate a lot of different parameters per pin to account for the current state the device is in to get the best (i.e. highest stable bit rate) possible link. And the parameters, value ranges and timings all vary between different chips, so if you write a blob for one combination of SoC and memory, it might very well be possible that you need an entirely different blob for another combination. And that is what the BSP vendor does and what is inside the blob... a tuned version of parameters that this specific board's firmware can use to achieve a working good link in the shortest possible time.
It's one hell of a ride that Flipper is on, and I seriously wish them all the best. There's a darn good reason this stuff has been proprietary, at best you'll get a high-level summary like [2].
[1] https://www.damninteresting.com/on-the-origin-of-circuits/
[2] https://www.ti.com/lit/an/snla415/snla415.pdf?ts=17793056953...