Provide ability to do real Machine Learning in Europe North
Currently on NCv1 with the ancient K80's that Google do not even give away for free on Google Colab is available in North Europe.
We have huge amounts of data in NE and thus it would be impossible or at least extremely cost prohibitive to move this data over to WE, where there are V100-NCv3 GPU's.
Even the NCv3 instances are ridiculously underpowered for SOTA ML work.
Just to give some perspective on how excruciatingly slow these are to work with:
My private gaming PC at home running a single 2080ti has 2.5x TensorFlops of a NCv1 running 4 K80 GPU's!
(K80: 5.6 TensorFlops, 2080ti: 56.9 TensorFlops)
Furthermore, horovod performance using Databricks (which we're using) is so slow that it doesn't even provide much benefit.
Considering the price per TFLOP in NCv3, something like NC96v3 (16x V100) seems quite reasonable to offer also since it's roughly equivalent of a DGX-2 workstation...
Dear Azure Customer,
Thanks for the feedback. Your feedback is very important to us. Our engineering team will investigate it and consider your feedback seriously. We will get back to you soon.
Thank you for understanding.
Azure CXP Community