Google's Tensor is supposed be all about machine learning - it just got crushed by the Apple A15 Bionic in our testing with the new Geekbench ML app

[Redaktion]

The Google Pixel 6 range is about to start its rollout in select markets with the promise that Google has made its new Tensor chip to really shine in machine learning tasks. Unfortunately, we just pitted it against the Apple A15 Bionic in the new GeekBench ML test and it got obliterated.

https://www.notebookcheck.net/Google-s-Tensor-is-supposed-be-all-about-machine-learning-it-just-got-crushed-by-the-Apple-A15-Bionic-in-our-testing-with-the-new-Geekbench-ML-app.575439.0.html

[an0n]

Interesting, that was my first thought that came to my mind after looked at geekbench generic result of so called tensor soc. It certainly matched to older sd 845/855/860 hence everything could not be far from that.

tl;dr google sells scam soc, too many false and misleading advertising.

[phusho]

Or just test didn't know how to use all the power of the new chip. I am pretty sure that Goggle have locked this learning coprocessor just for them, as they made same for previous pixel exclusive chips. May be it needs time to show it is full potential and some software updates.

[Dr0xium]

I'd be really careful about drawing such definite conclusions at this point in the game.  "ML" generally has several different types of optimizations that are possible in hardware.

1) Is it specifically tensor configured?
2) What kinds of pipeline operations are optimized? 
3) Does the general app space have access to the dedicated hardware?

The article doesn't seem to fill in these gaps, and we don't yet have hardware in our hands or Google commentary helping us add context.

I suspect that Apple absolutely has a lead here.  They invested earlier.  They have an entire team who has been working on this for a while.  They have a lot of dedicated/custom IP.  They hand-tuned the engine of thier own car, so to speak.

Google has a Samsung generic car that is a year old, that they bolted a supercharger onto. 

I think once we understand more about what the hardware specifics are for Tensor, how they appear to applications, how the pipelines can be used, we will learn more not just about "App XYZ has a gap of 70%" but WHY that is.  The why is important.  It could be that Tensor is 70% less capable.  It could be that its 20% less capable but the specific model in the benchmark was not running on the ML hardware at all.  Or was running using functions that the hardware doesnt provide as many resources for.  Or.... well... thats the problem: we dont know the important context of "why" yet.

All this article can really say with authority is: Google runs widget-spinning synthetic program 70% less fast than Apple - apparently on par with CPU level benchmarking from prior SoCs.

[Mike McNamara]

Apple blows everyone away, Samsung, OnePlus, Huawei, every one of them, the bionic A15 chip blows away Snapdragon 888.

[Dr0xium]

Let's add some context.

I have an S21 and S20FE to hand. 

So S20FE on an S865 scores
426/1356/750 for CPU/GPU/Math Proc

AN S21 with a S888 scores
423/1666/1158

This reflects that the CPU in the generational upgrade was not particularly suited for enhancement of ML but rather on execution pipelines rather than math.  Still, the GPU upgrades and enhancements to math processing are clear across the generations.

Against that background, the Tensor scored (based on this article's results):
307/1428/1720

We see here that the "medium" workload centric heterogenous processing strategy that Google went with de-emphasizes math on the primary processor and using lower clock speeds for CPU based ML processing.  The GPU is right where it needs to be in terms of raw math.  It doesnt best the 888 but its still a generational improvement vs 865. 

The google strategy of emphasizing ML workloads is very clear here, vs these chipsets as the numeric processing is higher even than S888's current gen capability by 48% of the 888's score!

Adding in the comparison against the Exynos 2100 which most of the IP blocks in this processor (basically a customized 2100), the Exynos scores
416/1466/736 on these same tests.

This makes it clear the IP strategy:
Reduced clock on the cores to emphasize the heterogenous strategy for simultaneous and medium-tuned workloads results in less top end power on CPU math.

The GPU is fundamentally a wash.  Compared to a "1500" baseline of an Intel i7-10700's math performance, they are within 2% between the base 2100 and the customized implementation.  This could be explained by the lower top end CPU clock, actually.

The math is MARKEDLY different between the unmodified Exynos 2100 IP and the Google version with a gigantic 736 -> 1720 difference when the ML is processed using the customized capability.

Does the Tensor "beat" the apple?  No.  And im not sure anyone expected it to.  Does it have more capability than just about any android ecosystem chipset today?  The Geekbench browser says "Yes - and its not even close."

[Partyco]

I didn't read past the first 2 sentences.

If you can write professionally without doing any research whatsoever about the device you're testing and then put misleading headlines on top, I can also leave negative comments of your work without spending the time to actually check it.

Tensor doesn't work like that, and you're a lousy blog.

[AnAppleADay]

Geekbench is a tool created by a guy who used to run a website dedicated to Mac hardware and software. It also says that the chip in an iPad pro is more "powerful" than a 32 core enterprise level server CPU. Why don't you try installing and running enterprise software on your iPad pro and see what you get?

[Ligma]

Firstly, Geekbench is biased trash, but so is this author.

Also, Tensor is nothing but a rebranded Exynos chip. Screw Google and ugly a** pixels. Samsung is where it's at.

[dvrk]

Sure they make good silicon but now their software needs to get up to Android performance

[Brynn Siebold]

Im sorry but i have both iphone and android smartphones, and iphones are completely useless at completing tasks and recieving notifications.

Half the time touchscreen does not recognize input accurately on iphones, and yet andriods no matter what touch screen always work even when hands are wet and raining outside.

When i download identical apps on both devices nearly half tge features are not availble on iphone as they would be on android.

The reason why iphones perform better is because nearly %75 of the feautures on iphone are fixed, dumbing down the system so it can deticate processing power to the its very limited capabilities.

Iphones are for people that dont understand how technology works,  they are absolutely useless phones.

[Steve Grey]

No where did I read any reports from Google claiming this to be the fastest or best soc ever, so most of what is written here is written by copy cat writers who get there information from YouTube. It is worth noting it is the best phone ever made for Google and will with future software updates supersede many other "Flagship" phones on the market at this price. I'm not a "fanboy" of any system, I own both an Android and an iPhone, both have there strengths and weaknesses.

[Shane Jowitt]

No one cares about tiny performance differences. It's Apple locking you in a small box or Android giving you some customizing as well as not forcing on you what they think you should be allowed to do with your phone.

It's the same with PCs people choose a PC over an apple because they want to be able to customize they want to be able to play games they want to be able to do everything not be crammed into a tiny spectrum of what their laptop has to offer.

If you want predictability go Apple if you want customizing, gaming for everything else that is Apple doesn't think you want you choose Android or PC.

[John McKenny]

Google designed Tensor for their specific workloads ML/AI.  Benchmarks make for good marketing and nothing else.

[Orange poop]

That fact that the writer only reports Geekbench scores without adding any additional context or expert details just makes this article blogspam trash.


Geekbench might not be the best test to diffentiate these chips in real-world apps, or might not be properly configured. No reports on energy usage either.