What Does Google’s Tensor SoC Mean For Google’s Hardware And Software Efforts?

This previous week, Google introduced its upcoming Tensor SoC, supported by a weblog from Google {Hardware} lead Rick Osterloh detailing how the SoC will affect Google’s Pixel 6 and Pixel 6 Professional flagship telephones. Curiously, although Osterloh’s weblog mentions AI 4 occasions, it by no means as soon as mentions 5G—the first promoting level for telephones lately. That is probably as a result of Google’s greatest energy is in software program, and extra particularly, AI.

Google’s AI {hardware} expertise

If Google’s announcement of its new Tensor smartphone SoC sounded acquainted to you in any respect, it’s as a result of Google’s Tensor Processor Models, referred to as TPUs, are already in use within the cloud for AI inference functions. Google’s TPU v4 for the cloud put up some spectacular preliminary MLPerf coaching outcomes towards Nvidia’s A100, which is taken into account the gold customary for coaching. So, naturally, Google’s determination to call its new SoC (codenamed Whitechapel) Tensor, means Google is aiming for heavy ML duties, totally on the inference facet. One factor to think about is that inference and coaching are very several types of machine studying features and customarily require completely different particular AI {hardware}. When you think about that almost all shopper merchandise are constructed to do inference and never coaching, you wouldn’t essentially anticipate TPUs to operate the identical because the Tensor SoC. Not all {hardware} that may do coaching is nearly as good at inference and vice versa. For the Google Pixel, ML efficiency per watt is much extra essential than absolute efficiency since we’re speaking about smartphone with a restricted battery and an enormous show.

The vital factor to think about concerning the Tensor SoC is that it’ll goal AI efficiency firstly. Google is greatest in school on the subject of enriching the consumer expertise with AI-accelerated options. Google Assistant is on the core of that have, combining its complicated voice recognition and context algorithms with its adept search capabilities. When you think about that Google’s translation, images and dictation options additionally rely upon ML acceleration, it’s comprehensible why Google would need to beef up that efficiency. Suppose Google can lead on AI options and use circumstances and speed up them quicker than opponents. In that case, there may be little or no probability that its opponents will catch up, making its consumer expertise unattainable outdoors of its ecosystem.

Present and future silicon companions

Google has but to announce the SoC’s foundry companion, which is able to both be TSMC or Samsung, or its modem companion, both MediaTek, Qualcomm or Samsung. It’s unlikely that Google would drop Qualcomm’s SoC and follow its modem, however we’ll go away it within the checklist for now. It’s price remembering that if Google ditches Qualcomm, it dangers the Pixel not having the quickest modem on the town (particularly contemplating Apple at present makes use of Qualcomm in its complete 5G iPhone 12 lineup). Many rumors counsel that Samsung is each the foundry companion and the modem supplier, which might make sense as a result of Samsung is the one firm that might present Google each on the similar time.

In response to  Qualcomm, the Snapdragon 888 has an AI efficiency of 26 trillion operations per second (TOPS). I might anticipate that Google’s Tensor SoC would want to surpass that if the corporate actually needs to justify spending a whole lot of thousands and thousands of {dollars} additional yearly to construct a customized SoC for its Pixel line. That stated, if Google builds a robust sufficient SoC, it might foreseeably use these similar SoCs in its tablets and Chromebooks. That would supply an fascinating alternative for Google to unravel one of many greatest issues that plagues customized SoC designs: scale. With out scale, constructing a customized SoC is a dropping proposition, and if a design workforce isn’t aggressive with what’s already obtainable available in the market as we speak, it comes off as a loss-leading train. Samsung and Apple each do it as we speak as a result of their companies deliver them scale. Huawei did the identical till the US Authorities made it nearly unattainable. In brief, there’s a longtime monitor file of this being finished by smartphone producers, however usually ones with a lot greater volumes than Google.

Wrapping up

All that stated, I do imagine that Google has alternatives to enhance the general consumer expertise on the Google Pixel line by totally integrating the SoC, OS and gadget beneath one roof. Apple has proven that the whole-system strategy to energy, efficiency, and consumer expertise pays dividends—it outperforms the remainder of the trade in practically each benchmark. Nevertheless, Apple merely doesn’t have the quantity of knowledge or AI capabilities that Google does. I perceive Google’s want to take management of the entire gadget, however I fear that the corporate could hamper it’s competitiveness by selecting the improper modem companion or make another transfer (like not having sufficient RAM) that undercuts what it’s attempting to do with the Pixel 6.

Disclosure: My agency, Moor Insights & Technique, like all analysis and analyst corporations, offers or has supplied analysis, evaluation, advising, and/or consulting to many high-tech corporations within the trade. I don’t maintain any fairness positions with any corporations cited on this column.