Not sure why this post is getting flagged to oblivion, but it is technically pretty interesting. GPUs remain expensive, and their use is typically prioritized

Not sure why this post is getting flagged to oblivion, but it is technically pretty interesting. GPUs remain expensive, and their use is typically prioritized for training, and then inference is run on CPUs. This provides a cost effective way to attach GPU resources on demand to regular instances, rather than having to run dedicated GPU instances.