Valentin SawadskiThe biggest news for me at #GoogleIO today was TPU 8 with their specialised version for Training and Inference. Making custom silicon is a serious effort, so Google must see a future whey they are constantly training models on a massive scale. I wonder if all this is to develop their next Generation of Gemini Models, of if we soon a much larger model zoo of (personalised) models? 1/n
Does anyone know why Google went and built dedicated silicon for #AI model training? (Improving their Gemini models or an extended zoo of (personalised?) models?)
A bit of googling shows the original announcement is almost a month old now, but it does not go into detail what exactly the TPU 8t models will be doing 🤔
https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive 2/2
