Discussion Google Tensor SoC thread

[poke01]

What?

Custom in house CPU or Cortex? One would assume so by the statement “fully self-designed”.

 

[FlameTail]

Custom in house CPU or Cortex? One would assume so by the statement “fully self-designed”.

The SoC is "fully self-designed", whereas the cores themselves are off-the-shelf Cortex CPUs and Mali GPUs.

The present situation is that Tensor SoCs are co-designed between Samsung LSI and Google Silicon team.

 

[hemedans]

By this point, the SoC + Modem is the biggest weak point for the Pixel.

Google Tensor G4 explained: Everything you need to know about the Pixel 9 processor

The Google Pixel 9 series sports a new Tensor G4 processor. Here's everything that you need to know about it.

www.androidauthority.com

The rest of the hardware is actually fine... great even. Excellent display, build quality and camera system.

Dont forget ufs 3.1, even some midrange now are using ufs 4.0.

 

[LightningZ71]

It made me realize how awful is Samsung EUV against SMIC DUV!

You're also looking at some of the most redined DUV machines out there vs. Early EUV machines with Samsung's unfortunately typical process implementation issues. SMIC can use heavy multi-patterning and likely be quite close or exceed Samsung's effective product results. SMIC also likely is less sensitive to cost and will not be as concerned about maximizing density or yields. However, in general, Samsung's track record speaks for itself.

 

[DZero]

You're also looking at some of the most redined DUV machines out there vs. Early EUV machines with Samsung's unfortunately typical process implementation issues. SMIC can use heavy multi-patterning and likely be quite close or exceed Samsung's effective product results. SMIC also likely is less sensitive to cost and will not be as concerned about maximizing density or yields. However, in general, Samsung's track record speaks for itself.

Indeed, but still Samsung should have the edge, not SMIC

I can see them eventually making 7nm chips in a very decent production cost. With DUV, showing up how strong could be the process.

 

[Doug S]

Indeed, but still Samsung should have the edge, not SMIC

I can see them eventually making 7nm chips in a very decent production cost. With DUV, showing up how strong could be the process.

Why? Samsung was never that strong when it came to multipatterning, that's why they went to EUV immediately for 7nm whereas TSMC did it with all DUV at first with N7 then added N7+ with a few EUV layers later as an optimization (and more probably as some mass production learning experience before plunging into EUV headfirst with N5)

SMIC reportedly had employees working "undercover" at TSMC so they probably got a lot of the benefit of their hard-won learnings in multipatterning, plus they get to use better DUV machines than what TSMC had for N7. So them being able to do 7nm is hardly surprising. In fact, with those better machines, I don't think doing a decent 5nm is out of the question for them. It is getting beyond that which will stymie them, as they have no access to EUV and multipatterning runs out of gas eventually even if you were willing to accept the cost of deca-patterning or whatever would be required to do an N3 equivalent with DUV only.

 

[Hitman928]

https://twitter.com/x/status/1823900261888467216

Why InFO_PoP? Are they stacking memory on top of the SoC?

 

[FlameTail]

Why InFO_PoP? Are they stacking memory on top of the SoC?

Don't all smartphone SoCs do that? The RAM is stacked on top of the logic die.

 

[Hitman928]

Don't all smartphone SoCs do that? The RAM is stacked on top of the logic die.

I don’t know, I don’t follow smartphone chips that much.

 

[poke01]

Don't all smartphone SoCs do that? The RAM is stacked on top of the logic die.

they do

 

[LightningZ71]

SMIC's machines should give them the ability to make chips that are in the quality triangle (density, performance and efficiency) of 5nm family processes, so long as they are willing to stomach heavy multipatterning. With further refinements to their techniques, they should be able to achieve some of the other advanced features of competing processes like varying fin height by region. They run into an absolute achievable density road block below the "5nm" feature size generation though. They can still be feature and performance competitive for a couple more years at increasing cost, giving them time to continue to work on their own lithography machines.

 

[hemedans]

Indeed, but still Samsung should have the edge, not SMIC

I can see them eventually making 7nm chips in a very decent production cost. With DUV, showing up how strong could be the process.

Thats google issue more than Samsung foundry. This is Tensor G3 vs Sd 888


Even though G3 has newer cores like Cortex X3 and A715 it fail to beat sd 888.

And this is G3, E2400, Kirin 9010 and 8 gen 3



E2400 has newer cores but you can see It's closer to 8 gen 3 than Kirin 9010.

 

[FlameTail]

View attachment 105517
Even though G3 has newer cores like Cortex X3 and A715 it fail to beat sd 888

That sounds really sobering when you put it that way.

Tensor G3
1 × Cortex X3
4 × Cortex A715
4 × Cortex A510
Samsung 4LPP

Snapdragon 888
1 × Cortex X1
3 × Cortex A78
4 × Cortex A55
Samsung 5LPE

The Tensor G3 uses;
• Newer ARM cores
• A 'beefier' 9-core configuration
• Newer process node

Yet it still fails to beat the 2 year older Snapdragon 888!

 

[FlameTail]

The Pixel Chief gave an interview recently:

An Interview with Google SVP Rick Osterloh About Pixel, Android, and Smartphone History

An interview with Google SVP Rick Osterloh about the Made with Google event, and understanding the company’s approach to Android and Pixel in the light of smartphone history.

stratechery.com

One of the things he said was that the goal of Pixel was to show what a 'premium Android' phone should be.

They can't do that if the silicon is subpar.

Android has been losing ground to the iPhone in the premium smartphone segment (>$500). It is so dire that Apple is taking home like 70% of the *profits* in the smartphone industry, leaving all the other Android OEMs to fight for the scraps. Research data shows >90% of Gen Z prefers iPhones. This is very, very bad and Google wants to change this.

 

[gdansk]

They can't do that if the silicon is subpar.

Why not? Phones are fast enough and have been for sometime. I run my phone in low performance mode all the time because my problem is battery life. I prefer getting 12-14 hours screen on time.

Most of the phone differentiators are in auxiliary units anyway. Camera processing, ML crap, GPU and so on. Now the Tensor might not be good at any of them - I haven't used one - but I don't think having a worse GB6 score prevents them from offering better user experiences with more accelerators.

But the price is too high for me.

 

[rosetta]

Why not? Phones are fast enough and have been for sometime. I run my phone in low performance mode all the time because my problem is battery life. I prefer getting 12-14 hours screen on time.

Most of the phone differentiators are in auxiliary units anyway. Camera processing, ML crap, GPU and so on. Now the Tensor might not be good at any of them - I haven't used one - but I don't think having a worse GB6 score prevents them from offering better user experiences with more accelerators.

But the price is too high for me.

Using a Pixel 6 at the moment made me realize how important the SoC is. Even light loads and just navigation apps make the phone overheat and battery drain simply connected to 5G is insane. I guess that's because of the old modem but still.

With Google saying they want to show what a premium Android experience looks like this doesn't help at all.

Besides will Tensor G5 not be outdated on release as well if it was meant for Pixel 9 originally? Did the use the additional time they now had to incorporate some newer features?

Otherwise would Tensor G6 be the noteworthy upgrade?

 

[sze5003]

Tensor is why I left the 6 pro for Samsung ultra series. I always had Google pixel phones before. I couldn't even go down the street without the 6 pro losing signal/data and this would require either toggling airplane mode on and off or rebooting.

Not to mention it would hear up quickly and battery life on mobile data drained very quickly. I hope they improve it if they plan on using it but at this point google is charging as much as other flagships but not providing the same equal level of hardware. The only thing I miss from the pixels are the cameras.

 

[rosetta]

Tensor is why I left the 6 pro for Samsung ultra series. I always had Google pixel phones before. I couldn't even go down the street without the 6 pro losing signal/data and this would require either toggling airplane mode on and off or rebooting.

Not to mention it would hear up quickly and battery life on mobile data drained very quickly. I hope they improve it if they plan on using it but at this point google is charging as much as other flagships but not providing the same equal level of hardware. The only thing I miss from the pixels are the cameras.

That's the Pixel 6 experience right there.
I'm interested in how the new modem in the Pixel 9 is going to improve connectivity and battery drain.

Hopefully Pixel 9 flops and they are forced to price Pixel 10 more aggressively

 

[DZero]

If Pixel 10 fails, I am afraid of Google's future. I mean, if even with TSMC they fails, is clearly that they are not doing the job well.
At least would be beyond UNISOC which is failing hard and even Huawei managed to comeback pretty fast. Their Kirin 8000 practically is on par with Dimensity 7300, which says how behind is UNISOC.

 

[hemedans]

If Pixel 10 fails, I am afraid of Google's future. I mean, if even with TSMC they fails, is clearly that they are not doing the job well.
At least would be beyond UNISOC which is failing hard and even Huawei managed to comeback pretty fast. Their Kirin 8000 practically is on par with Dimensity 7300, which says how behind is UNISOC.

I don't think Unisoc compete with Tensor or Kirin Line of soc. Unisoc is for ultra budget devices and they are doing preety good job.