Discussion Google Tensor SoC thread

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

FlameTail

Diamond Member
Dec 15, 2021
4,376
2,750
106
Tensor G3 die shot:


GPethfTakAM7GWO.jpeg

Node
Samsung SF4 (4LPP)

Die Area
10.71 x 12.62 = 135 mm²

The SoC is remarkably large (despite not having an integrated modem). The 16 MB SLC is bigger than Snapdragon 8G3 or Dimensity 9300. The TPU (Google's take on the NPU) also takes up a large amount of area.
 

FlameTail

Diamond Member
Dec 15, 2021
4,376
2,750
106
As the OP of this thread, I have decided to rename this thread from "Google Tensor G3" to "Google Tensor SoC thread", so that the discussion may not be limited to only the the Tensor G3, but extended to all Google Tensor SoCs including future ones.
 

FlameTail

Diamond Member
Dec 15, 2021
4,376
2,750
106
Google Tensor G4 rumours:


Screenshot_TensorSoCG4_Chrome.jpg
The regression in the "width" of the CPU is notable. The CPU component of the Tensor G4 looks more comparable to a midrange SoC from Qualcomm/Mediatek, instead of a flagship one.
 

soresu

Diamond Member
Dec 19, 2014
3,372
2,631
136
The regression in the "width" of the CPU is notable
Regression compared to what exactly?

Tensor has never been a top end SoC compared to the best QC/MTK options.

As far as I can see there is no generational regression on that chart either.
 

FlameTail

Diamond Member
Dec 15, 2021
4,376
2,750
106
Regression compared to what exactly?
As far as I can see there is no generational regression on that chart either
Tensor G3 was 1+4+4. According to this rumour, Tensor G4 will be 1+3+4. Clearly, a regression in the number of A7xx mid cores.

Meanwhile, Qualcomm/Mediatek/Samsung have been beefing up their CPUs in recent years;

8G1 = 1+3+4
8G2 = 1+4+3
8G3 = 1+5+2

I expected the Tensor G4 to upgrade to 1+5+4 (a configuration similar to it's distant relative born the same year- the Exynos 2400).
 
Last edited:

soresu

Diamond Member
Dec 19, 2014
3,372
2,631
136
The Tensor G5 SoC will be the first fully in-house design by the Google Silicon Team, in contrast to the decade+ experience that Samsung LSI has in designing SoCs
This is ignoring the years of experience Google have designing silicon themselves.

They have their own general ARM server CPU chip Axion, several generations of their own AI/ML training and inference TPU chips (the latter of which integrated into the Tensor Gx SoCs), and their Argos VCU chips that handle Youtube encoding.
 
Last edited:
  • Like
Reactions: Tlh97 and FlameTail

soresu

Diamond Member
Dec 19, 2014
3,372
2,631
136
Tensor G3 was 1+4+4. According to this rumour, Tensor G4 will be 1+3+4. Clearly, a regression in the number of A7xx mid cores.
A7xx are not super big, but they are big enough to add extra cost increasing area and increase licensing costs to ARM per SoC.

They probably did a cost vs benefit analysis based on user use and determined it wasn't worth the extra.

Google are very price conscious with their Pixel phone line.
 

soresu

Diamond Member
Dec 19, 2014
3,372
2,631
136
A tragedy for a lineup of phones that aspires to be 'the iPhone of Android'.
Journalists mark it as such more than Google themselves.

A far closer market comparison is Samsung/Android vs Apple/iOS, and that has always been the case.

The main point of Pixel is cost conscious build and standard Android software that updates pretty much the week Google announces each new version has gone gold.

Having suffered Samsung's Message app crashing my phone more than once it's beginning to feel like standard Android might be worth a hardware downgrade.

Not that my pokey Galaxy S21 FE is going to be that impressive vs Tensor G4.
 

coercitiv

Diamond Member
Jan 24, 2014
6,795
14,851
136
Having suffered Samsung's Message app crashing my phone more than once it's beginning to feel like standard Android might be worth a hardware downgrade.

Not that my pokey Galaxy S21 FE is going to be that impressive vs Tensor G4.
I don't even remember how the default Phone, Contacts and SMS apps look on my S20 FE, changed to the Google ones very quickly after purchase. When it comes to the Pixel experience, not being able to remove the search bar from the homescreen felt like a petty thing to do from Google, and not being able to double-tap to turn on the screen is one of the more requested features that Google ignores because reasons. It's still easier to fix the Pixel software experience than the Samsung one, that's for sure.
 
  • Like
Reactions: Tlh97 and soresu

LightningZ71

Golden Member
Mar 10, 2017
1,959
2,350
136
We're an Samsung Galaxy A52 5g family, up until recently. My wife's phone got to where it wouldn't charge unless you held the cable at a precise angle, with the moon in a particular phase, etc... and we replaced it with a Pixel 8a. I won't say that the phone is lightning fast as compared to the old one, but, it is consistently responsive, and the apps just consistently work. The Samsungs sometimes require killing and restarting apps, they can get really sluggish, and often get weirdness with wifi/cell communications.

The Pixel 8a is just better in every possible way, and is worth the modest extra amount they ask for it.
 

LightningZ71

Golden Member
Mar 10, 2017
1,959
2,350
136
Not very shocking? G3 is fabbed on a Samsung node that's known for being underwhelming at best, using ARM cores that are at least a generation behind contemporaries and not aimed at absolute performance.
 

FlameTail

Diamond Member
Dec 15, 2021
4,376
2,750
106
Not very shocking? G3 is fabbed on a Samsung node that's known for being underwhelming at best, using ARM cores that are at least a generation behind contemporaries and not aimed at absolute performance.
That graph not only speaks about Tensor's mediocrity, but also how surprisingly good the Huawei chip is.
 
  • Like
Reactions: rosetta

LightningZ71

Golden Member
Mar 10, 2017
1,959
2,350
136
If it's single thread, while great, that's not overly impressive. You can sacrifice efficiency for performance in ST and not cause too many problems unless you hit thermal saturation. I'm interested to see their MT performance and actual die size compared. I suspect that they aren't too terribly far apart given the known qualities of Samsung's nodes.