Three chips in and Google Tensor is on life support

[Alexander_]

If someone does not like what the author described, it does not mean that it is not true.

The fact is that mobile chips can handle offline requests well. An example of this is HiSilicon chips from a few years ago.

So this goal - to do what others have already done several years ago, is not superfluous. The question is to approach its implementation as thoroughly as possible.

Because even Samsung, which does not make the best chips, still makes normal devices, because it controls many processes. Therefore, Google should approach the production of smartphones not in the "we will tell them - they will make it for us, and then we will collect everything together" type, but really involve its own experts in this, immerse themselves in it as much as possible, bring the optimization of the OS itself to order [and not hope, since it already exists, it seems that it should work somehow].

[Haahaa]

Why does this article read like a teenage boy saying everything bad he can think of about the girl who just turned him down for a date?

[Google fanboi]

So funny to see Google fanbois defending the pathetic Tensor under lacklustre Pichai's watch. Enjoy your Tensors and Google marketing lies.

Not all of them. Please do not lump us all in the same category. Many of us refuse to touch/upgrade to any Exynos-based Tensor chip Pixels with a 10 ft bargepole. Yes, we are well-aware it is complete trash. Also doubt these are actual true fans at all because if they were they'd be more critical on the company they're spending money on.

And yet, my 6a is running just fine. Stop with your Apple shilling bullshit.

Uhh, by the same token even a 7 year old phone is "just fine" if all you do is basic messaging, email, surf and youtube. Why not just buy a old used phones if not interested in actual progress? This has nothing to do with Apple shilling. Nobody should be defending this sorry excuse of a chip, AI/fanboism-shilling or not, it's inexcusable.

[Yonisanu]

Interesting points raised, but there are likely many factors not considered in the conclusions drawn.

More questions should be asked than statements made, as a rule of thumb for a modest appraisal of the situation.

Have you considered the price comparisons (including launch bundles), profit margins, long term goals, marketing, etc.?

[hwertz]

One possible problem may be capabilities rather than speed.  Some of the AI/ML solutions (like the one in my 11th gen Intel CPU -- which Intel seems to be preternding doesn't exist since they've commented on the 13th or 14th gen being the first to have ML acceleration in it)... the capabilities are very limited.  CUDA it ain't.  Very limited data set size, only supporting 1D input (i.e. audio, not even a 2D table of data).  My understanding is in Windows it's used for microphone background noise removal. In Linux there's a driver for it and I got it to show up in Tensorflow but it was too limited to run anything that basically wasn't specifically designed with it's limitations in mind so I got no "out of the box" models to train or predict with it.

I don't know what the Google Tensor's capabilities are but it could be it's missing FP32 or FP64 support, data size or shape limits, etc. make it unsuitable for running these things Google is running in cloud.

[Mishoo]

You murdered Andrei Frumusanu's name. 🤦🏻�♂️

[nokite]

What is the author talking about??

Suggesting that a mobile device should be able to run those large generative AI models?? That's ridiculous. Show me a Qualcomm chip that can do that!

Also saying that processing on-device will take less power... Have you ever precessed video? It's a much heavier task than networking.

With that said, yes, the Tensor chips are not great. Mostly due to heating up, in my opinion. I do not care about high-end performance, but about efficiency - YES, I DO.

[Neenyah]

What is the author talking about??

Looking at benchmark numbers where more is better and doing a whole opinion based on that, lol.

Suggesting that a mobile device should be able to run those large generative AI models??

Yes 😁

Also saying that processing on-device will take less power... Have you ever precessed video? It's a much heavier task than networking.

Very true (also a lot of thermal throttling in many phones).

[A]

Well there are native implementations of Stable Diffusion generators running on iPhones' neural engine and GPUs, but it's so useless - slow and hot.

[SteveT]

I must recharge my Pixel 7 twice to make it through a full day.

I don't know, maybe just use GrapheneOS?

Obviously, our usage patterns will differ, but the battery life on my Pixel 7 Pro with GrapheneOS is fantastic.

[A]

Obviously, our usage patterns will differ, but the battery life on my Pixel 7 Pro with GrapheneOS is fantastic.

Can't go against physics, SoC has a specific numbers of consumption and a specific battery capacity in Wh. If it makes Pixel fantastic, any other phone with 2x battery life will become, idk, marvellous. But as it's limited to google phones, you simply can't compare.

[Pas_]

I don't understand all the moaning in the comments.
It's quite obvious that Tensor is far behind. My Pixel 8 is barely faster than my old OnePlus 8 was, and that's with a 4 year old Snapdragon 865. What's more, it gets hot just messaging people and listening to music, because the modem is inefficient as all hell while also providing terrible cellular reception. Light tasks drain the battery, and it just stops charging if you're doing anything because the battery hits 40°C in no time. You have to recharge within the day if you plan on more than 5 hours of screen-on time.

This chip is supposed to be flagship-class, but it unintentionally continues the Pixel 5's expensive phone with mid-range performance tradition.