In the first two days, the start-up published two projects that aimed to squeeze the best performance out of each chip for cost-efficient model training and inference tasks.