Notice they require 32gb RAM and encourage 64gb RAM. Presumably the architecture keeps the model in CPU RAM and shuffles it dynamically to the PCIe card networ

Notice they require 32gb RAM and encourage 64gb RAM. Presumably the architecture keeps the model in CPU RAM and shuffles it dynamically to the PCIe card network? I'm guessing here. Whatever quirks come out of their hardware, I want to keep an eye on it. I also think that comparing the current generation of models, which are built and trained to maximize GPU or TPU bandwidth, could be improved if someone architected models to maximize their model to greyskull advantages. Given PyTorch runs on it, I don't think it'd be too hard to do.