I've been experimenting text based GPT from OpenAI's ChatGPT to Google Bard to now Gemini AI.
Since late '22 when the world was caught in attention on the magic of ChatGPT, people have been flocking to buy the next best accelerator cards (Nvidia), hoping they can be the next one to come out with the killer app – that ambitious goal to reach AGI first.
It's only up until very recently that I was made known that Gemini AI is entirely trained on Google custom TPU. I read a little about it (I found that TPU discussion is very little over here) and found that in the earlier days since Alexnet deep learning approach back to 2012.
Previous generalized computing (thank you Intel Xeon) tries to cope with big data workload. Accelerators were found to be particularly efficient in those machine learning workload. Early exploration by Google on designing custom TPU for themselves so they can have more flexibility catering to specific workloads. This decision alone turns out to be very valuable in today's Gen AI era.
With recent news from Apple that they have been training their AI models with Google's TPU, also the performance of Gemini AI to be on par with industry leading ChatGPT (heard that Antropic's Sonnet is even better – but Gemini is really good enough for my daily use).
These observations make strong statements to the Gen AI community, i.e., you can bypass Nvidia and still perform adequately well (even supercede for some reports) in your AI product offering.
The fact that Google does not hoard all its computing power to its own Gemini training, and the ability to extend its compute to customers like AAPL speak volume on the native TPU effectiveness.
Cost and time wise, Google is the prime candidate in AI era as TPU helps smoothen out the training cost. With ads revenue fueling the AI training (heard that OpenAI now needs more money), and superior TPU working natively, I'm of the opinion that Google will create a lot of value many years from now.
Leave a Reply