What are the future chips that Apple, Qualcomm, and Google are betting on?

      During the time I was using the Pixel 6 Pro, apart from taking photos, I rarely felt that this phone, which Google calls the "smartest Pixel phone", had any flaws, until one morning when the phone alarm woke me up.

　　Unlike ordinary mobile phones, where you slide to turn off the alarm, the Pixel 6 Pro prompts me to say 'Snooze' or 'Stop' to control the alarm. When I carefully said 'Stop', the phone that was beeping immediately quieted down.

　　This is a trivial little feature, but it keeps me in a good mood every morning when I am woken up by the alarm clock.

　　I finally no longer have to force myself to stay asleep and randomly look for my phone. All I need is one sentence to make the phone shut up. This is the first time I feel that my phone can "understand" me.

　　The secret to "understanding what people say" lies in the humble TPU.

　　Ubiquitous AI computing

　　On the SoC of a mobile phone, the presence of NPU is always much weaker than that of CPU and GPU.

　　This processor dedicated to neural network computing does not even have a unified name: it is called NPU on Kirin chips, and Neural Engine on A-series bionic chips; Google named it TPU, and MediaTek believes that it should be called APU for AI computing...

　　Although these chips have a variety of names and different architectures and principles, their purposes are roughly similar - to accelerate machine learning and improve the artificial intelligence computing capabilities of mobile phones.

　　If you pay attention to the performance of mobile phone processors, you will find that whether it is the iPhone's A series chips or the Android flagship Snapdragon chips, the CPU computing power improvement has been very limited in the past two years, and the performance "toothpaste squeezing" phenomenon has become more and more serious.

　　In comparison, AI computing power has become a parameter indicator that more manufacturers are willing to mention. Taking the A series chips as an example, Apple's A14 Bionic chip has nearly doubled the peak computing power of the previous generation and can perform 11 trillion operations per second.

　　A year later, the A15 Bionic chip can still bring a significant improvement of more than 40% on this basis, and can perform up to 15.8 trillion operations per second.

　　The Android camp has also made significant progress in AI computing power. On the AI ​​performance test list released by the Swiss Federal Institute of Technology, the Kirin 970, which introduced NPU for the first time, scored 23,600 points in AI performance. Four years later, the Google Tensor chip topped the list with a high score of 214,700, while the scores of Kirin 9000 and Snapdragon 888 also reached around 160,000 points.

　　Since AI computing power is growing almost exponentially, why is it so difficult for us to feel any changes? Is the word AI function, which sounds a bit advanced, too far away from us?

　　In fact, every time you unlock your phone, wake up your voice assistant, or even press the shutter button, you are having an intimate contact with AI computing.

　　The NPU is like a black box, which makes the AI ​​calculation process so fast that it is almost non-existent, so that you can't perceive the technology, but you are surrounded by more natural human-computer interaction. The evolution of Google Voice Assistant is a good example.

　　Since Siri added the 'Hey, Siri' voice wake-up function in 2014, the wake-up word has been almost bound to the voice assistant. Every time we talk to the voice assistant, we have to call their names tirelessly: Siri, Xiao Ai, Xiao Bu, Xiao Yi... If the voice environment is very noisy, this embarrassing process may have to be repeated many times.

　　This is because, for power consumption reasons, the mobile phone processor cannot waste computing power to parse every sentence of the user in the background for a long time. At this time, a low-power voice receiver that only recognizes wake-up words is needed to work permanently.