[jamalaramala]

That's exactly what I wanted to ask:

Their github page claims that it is possible to "tune LLaMa3.1 on Google Cloud TPUs for 30% lower cost", but they don't mention performance.