Cloud TPU Pods: AI Supercomputing for Large Machine Learning Problems (Google I/O'19)

[@bobsalita3417]

Would be nice to see TPU2/TPU3/GPU comparison chart supplemented with transistor counts, wattage.

[@martingorner6233]

Those numbers are not public but the only reason we are doing TPUs is that they offer a huge datacenter density and power advantage (in one word: cost) over GPUs, for ML workloads.

[@bobsalita3417]

Thanks. Great information. Tensorflow 2.0 isn't mentioned. Would like to know 2.0 implementation challenges, timelines, anticipated performance differences.

[@martingorner6233]

Keras/TPU support is planned in TF 2.1 and the TF 2.1 RC0 has just been published!

[@bobsalita3417]

Compute cost is an important cloud metric but what about BYO system (GPU) costs? A slide on BYO (GPU) costs vs GPU/TPU cloud would aid decision making.

[@MartinGorner]

The model I trained on stage on a 128 TPUv3 cores cost a little under $50. A TPUv3-128 is roughly the equivalent of 150 powerful GPUs on this workload. You can look up the market price of GPUs + interconnect. I hope this helps you make the buy on site / rent from cloud decision.

[@ShaneTheMathGuy]

Not sure why TPU does not support training multiple models on the same instance. This is not quite efficient for small models...

[@РодионЛукашов-п5у]

8:20 watch

[@visionscaper]

Usually 5 V100s will be in one node, I don’t believe that one TPUv2 is as fast as 5 V100s, not by a long shot.

[@bhuvaneshs.k638]

Nvidia Tensor Core vs Google's TPU

[@Diamond_Hanz]

NVIDIA!!!

[@xavierpan515]

We need TPUv4!

[@dinceraydn4992]

io 2020