Tensor SoC for Pixel smartphones
Pixel smartphones of the last two generations are based on SoC Tensor and are very far from the top solutions of Qualcomm, MediaTek Apple. But Google believes that this is not a problem at all.
In an interview, Monika Gupta, Senior Director of Product Management at Google Silicon Team, spoke about how Google approaches the creation of its new platforms and what guides it when choosing one or another development vector. It turned out that Google is looking to the future.
I don’t make decisions based on where machine learning is today, and I can say that because I work at Google. The same goes for the software that our development team makes. I know what the software development team wants to achieve from users in five years. This is the advantage of not being a commercial silicon supplier, but an internal silicon supplier. So these trade-offs are very complex, but I think they get a little easier when you’re vertically integrated.
Gupta added that the silicon team is talking to Google AI researchers to see exactly which machine learning models will be trending five years from now. That is, when developing Tensor, the search giant is trying, firstly, to focus on AI-related technologies and functions, and secondly, to strive to make the AI-related capabilities of SoC Tensor still relevant in five years.
As for benchmarks, Gupta says that’s definitely not Google’s goal.
I think classic benchmarks served a purpose at one point, but the industry has evolved since then. Classic tests were created at a time when AI and phones didn’t even exist.
They may tell a story, but we don’t feel like they’re telling the full story. So, for us, what we evaluate is the actual software workloads that we run on our chip, and then we strive with each generation of the Tensor chip to make them better, be it better quality, better performance, lower power consumption.