[Paper review] Exploring the limits of cconcurrency in ML training on Google TPUs
- 1 minProblems
- High memory cost of DL training
- Low GPU resource utilization because of insufficient-memory-induced model depolyment
Contributions
- xxxx
- xxxx
Main ideas
- xxxx
- xxxx
Key results
- xxxx
- xxxx
My thoughts and potential follow-ups
- xxxx
- xxxx