> That DeepLearning12 build cost a lot. For example, we had to get a $350 torque driver that could hit the tolerances needed not to crack the NVIDIA P100 GPUs. I was in the data center starting to install these, and one of our kind readers who worked on the Baidu AI research cluster in that same data center stopped me, letting me know I might break the GPUs if the torque was not perfect. Indeed, the worst story I heard about this was that HPE broke several V100 generation GPUs due to torque/ thermal paste thickness.
This is scary to me!