Yes ! Thanks you, the Ara240 is at the heart of the matter. The reference to the 16168 module was regarding the Gateworks GW16168, which as far as I can tell isn't tremendously different than the Ara240 16GB M.2 module (apologies if that was an obscure reference). What I'm trying to build would more properly be called a pcie expansion card rather than an SBU. Switched PCIE riser cards were originally built to create M.2 m-key drive expansions, but I want to re-purpose that PCIE infrastructure to host 4 Ara240 m.2 modules, giving me a cluster to run LLMs locally. I'm not sure if the Tensor LLM structure would work as well as it does for my GPUs, but I'd love to give it a try. My highest hope is that I can achieve a heterogeneous compute environment on my pc. Anywho, the real kicker would seem to be that driver construction. My Cursor agent feels like it can do a lot, but there are real limits to what it can put together without some further software level info about the Ara240 chip.
