GTC Silicon Valley-2019: ImageNet in 18 Minutes Using Public Cloud
GTC Silicon Valley-2019 ID:S9573:ImageNet in 18 Minutes Using Public Cloud
Yaroslav Bulatov(South Park Commons)
We'll discuss how ImageNet can be trained from scratch to standard benchmark accuracy in 19 minutes on 64 NVIDIA Tesla V100 GPUs by relying solely on open-source tools and publicly available cloud infrastructure. Although previous efforts in this area used dedicated cluster hardware and specialized Internet adapters, our implementation performs well over Ethernet in a shared cloud environment. We'll walk through algorithmic improvements needed to achieve this result and lessons we learned in the process.