Has anyone used the BeagleBone AI yet ? As you may know around 2016 time-frame Texas Inst purged their DSP roadmap and Signalogic was unable to convince them to support PCIe card multicore (c6678) server solutions for AI, deep learning and ML applications. Their marketing and web page attitude towards servers was "the box that must not be named". I posted an article both here and on LinkedIn about what happened.
Now it looks like maybe we'll support a TI solution again in the form of the BeagleBone AI (AM5729 SoC) as it purports to be a single-board server solution that doesn't need USB JTAG, non-network terminal connection, Windows IDE, and other ancient TI methods in order to be usable.
I'm looking for feedback, either from actual users, or in the form of reaction / expectations. Does this stand a chance ? Based on their groundhog day behavior, it's not expected that TI would ever market this as a server solution -- that would be up to Signalogic and other third-parties. But at least it would be a server that millennial (and younger) engineers can develop with as they are accustomed -- a first for a TI SoC solution.
Looks like a nice upgrade to the Beaglebone Black SoC, so that's kinda cool. The rest of it seems to be the library support for AI, which is always dubious these days. Whether it'll be successful depends on how well it compares to the processing power on the other typical small, cheap, processing boards, like an RPi or whatever.
It is nice that there are powerful little platforms like this around that are not expensive. I've done SDR projects for customers on a number of them, and seeing the capabilities of the Beaglebone platform evolving in the right direction is nice from that perspective, but it's certainly not the only thing out there.
First thanks for your feedback. "it's certainly not the only thing out there" - yes you are so right. Therein lies the root of TI's demise as a semiconductor influencer.
TI will not have major success with any of their c66x SoCs until they start marketing their products to engineers with ML and data science degrees as interchangeable with other servers, suitable for IoT and cloud native applications. Nvidia mastered that years ago and results are obvious.
"Whether it'll be successful depends on how well it compares to the
processing power on the other typical small, cheap, processing boards,
like an RPi or whatever.
I would mostly agree, but SWaP matters too. TI is still cutting edge in power and package size. That is their remaining strength, but as you imply it depends on which application. For apps that need H.264 or some other type of codec before CNN/DNN processing AM5729 looks promising. Searching around shows AM5728 max power around 6.5W, adding another 1 - 2 W for EVE cores on the AM5729 leaves it well under an Nvidia Tegra-K1 or Intel Atom + GPU.