Engineering sample Snapdragon 8 Gen 1+ benchmarked

[Senzelian]

2 hours ago, leadeater said:

Well what was yours? You claimed Google was doing something different right? Something others are not doing and that would lead to Google surpassing Apple? If so then it's no on every count and Apple is ahead of Google.

You're interpreting more into my comment than there actually is.

 

1. I said that an aging CPU and GPU design doesn't hold them back. Tensor is already surpassing a lot of much higher end CPUs and GPUs on the market when it comes to the user experience. This specifically aplies to the Android market. Has nothing to do with Apple.
    
2. The other thing I said is that they will eventually overtake Apple with their CPU design.  This has nothing to do with the previous statement. Whether Google and Apple do the same basic thing doesn't matter. I'm literally saying, Google will do what Apple already does, much better. 

 

2 hours ago, leadeater said:

How can they "eventually overtake Apple" when Apple was doing it first, better, faster, more efficiently?

Mercedes Benz was building cars first, better than everyone else, faster and more efficiently. Are they still the number one? VW, Toyota and Tesla all overtook them.

There's your car comparison.

 

And honestly I think Apple will go the way of the dodo if they can't get their foot into the car market.

[leadeater]

1 minute ago, Senzelian said:

I'm literally saying, Google will do what Apple already does, much better. 

And your reasoning is?

 

2 minutes ago, Senzelian said:

Tensor is already surpassing a lot of much higher end CPUs and GPUs

Tensor is slower and less efficient than Apple's NPU. Tensor is also slower than Nvidia so this isn't factually true at all.

 

Googles TPUv4:

• TDP: 175W
• TOPS: Not published, assume 180 (TPUv3 was 90, TPUv2 was 45)
• TOPS/w: ~1.0

 

Nvidia H100:

• TDP: 700W
• TOPS: 4000
• TOPS/w: ~5.7

Nvidia: A100:

• TDP: 400W
• TOPS: 624
• TOPS/w: ~1.54

Apple A15/M2:

• TDP (of NPU): Est 5W
• TOPS: 15.8
• TOPS/w: Est ~3.16

 

I don't know where you have been getting your information from but it can't be correct if you think Google is leading in this area or improving faster than others are.

 

1 minute ago, Senzelian said:

Has nothing to do with Apple.

And yet you explicitly mentioned Apple....

[Senzelian]

2 hours ago, leadeater said:

Tensor is slower and less efficient than Apple's NPU. Tensor is also slower than Nvidia so this isn't factually true at all.

I don't think you understood what I meant with Tensor, because you're involving Nvidia.

I'm talking about Tensor, the SoC, that is used in the Google Pixel 6 and 6 Pro, which directly competes with Exynos and Qualcomm chips.

As we know, Google used an aging Samsung Exynos SoC as its basis for Tensor and yet was able to outperform much more modern SoCs from Qualcomm and Samsung. That becomes obvious the moment you use the phone.

 

I specifically said in my first comment "Google has shown with Tensor that even an aging CPU and GPU design do not necessarily matter in an everyday use case."

 

 

2 hours ago, leadeater said:

And yet you explicitly mentioned Apple....

 

Yes of course, cause Apple's A15 and M1/M2 are the only rivals to Google's Tensor SoC. And that has nothing to do with the performance of any of those chips. It is simply because both companies control hardware and software of their own devices. Microsoft doesn't do that, Nvidia doesn't do that, Samsung doesn't either.

 

 

2 hours ago, leadeater said:

And your reasoning is?

Because Google simply has the additional ressources to train an ML model.

 

And we can already see that with Google Assistant, Maps and all the new features that were announced during Google I/O 2022. And on top of that Google has the horsepower to work on much larger projects. Their cloud infrastructure is massive compared to Apple's. Apple is a nobody in the cloud space and when it comes to collecting data, there is no competitor to Google.

 

In short, there is simply more to work with. I could even see Microsoft taking over the mobile business sector entirely, which is now mostly dominated by iPhones. Microsoft inTune has already proven to be the choice of most large companies to enroll and manage company devices.

[leadeater]

13 hours ago, Senzelian said:

Because Google simply has the additional ressources to train an ML model.

And Apple doesn't? Apple totally does do this. Nvidia also does this, more than anyone else.

 

13 hours ago, Senzelian said:

Yes of course, cause Apple's A15 and M1/M2 are the only rivals to Google's Tensor SoC. And that has nothing to do with the performance of any of those chips. It is simply because both companies control hardware and software of their own devices. Microsoft doesn't do that, Nvidia doesn't do that, Samsung doesn't either.

Nvidia does have SoCs with ARM cores and AI accelerators, they are in basically every electric car on the road.

 

Nvidia also has Tegra as well.

 

Tegra X2

Quote

• CPU: Nvidia Denver2 ARMv8 (64-bit) dual-core + ARMv8 ARM Cortex-A57 quad-core (64-bit)
• RAM: up to 8GB LPDDR4[115]
• GPU: Pascal-based, 256 CUDA cores; type: GP10B[116]
• TSMC 16 nm, FinFET process
• TDP: 7.5–15 W[117]

 

Control of the entire software stack and OS is unnecessary. Only bad developers blame the tools and hardware, without really good reasons.

 

13 hours ago, Senzelian said:

As we know, Google used an aging Samsung Exynos SoC as its basis for Tensor and yet was able to outperform much more modern SoCs from Qualcomm and Samsung. That becomes obvious the moment you use the phone.

Yes I know that is what you were saying, but then you threw in Apple who is doing the exact same thing, first, better, faster, more efficiently and developing the capability also faster.

 

Google literally has to pull something out of a very big hat to surpass Apple.

 

13 hours ago, Senzelian said:

Their cloud infrastructure is massive compared to Apple's. Apple is a nobody in the cloud space and when it comes to collecting data, there is no competitor to Google

Apple has huge global datacenter capabilities, they just don't sell it to anyone to use other than through their services, you know some of the most widely used services in the world?

 

Are you going to say Netflix is also small too?

[Senzelian]

5 minutes ago, leadeater said:

And Apple doesn't? Apple totally does do this.

Not to the extend of Google.

 

5 minutes ago, leadeater said:

Nvidia does have SoCs with ARM cores and Ai accelerators, they are in basically every electric car on the road.

A car isn't a phone. This topic is still about ARM chips in phones, and specifically their progression and the advantages they bring to the user.

 

6 minutes ago, leadeater said:

Control of the entire software stack and OS is unnecessary. Only bad developers blame the tools and hardware, without really good reasons.

Ah yes, they're all just bad. That must be it. I guess Apple just bought up all the good ones. All praise Apple!

Apple clearly found it to be necessary. Google too.

 

9 minutes ago, leadeater said:

developing the capability also faster.

Yes so fast in fact, that I decided to use Siri instead of Google Assistant - no one said ever.

[hishnash]

11 minutes ago, Senzelian said:

Not to the extend of Google.

 

Depends on the training approach, apple have something google do not have. A lot more phones with powerful ML hardware on board. And yes apple does use these to do real world on device training based on real world user data. There is a method were they pertain a selection of models and ship them out to people phones, then when your phone is charging at night it will evaluate these and do a small amount of additional training then using a method known as `Differential privacy` the phones can report back on the results without exposing personal info (and yes this has been mathimaticlyl proven to not expose any single persons info).  when apple deeply ML most of the time you do not notice it, that is what they consider a sucseffull deployment of ML you should not even notice that its being used. Things like better accidentally touch rejection, or smarter charging times, or even when drawing with the pencil predicting were you are about to draw when the next frame will be displayed so the next frame can include that stock even through it started rendering before you made the stoke. 

Apple do also of course have vast datacenter training capacity most of googles data centre capacity is not in ML workflows its in data storage and retrieval for search and add tracking.

[leadeater]

1 hour ago, Senzelian said:

A car isn't a phone. This topic is still about ARM chips in phones, and specifically their progression and the advantages they bring to the user.

Doesn't matter if it's in phones or not, the advantages are common to all hardware platforms. Yes you were speaking about mobile phones but lets just say Google and Apple won't be the only ones doing this.

 

1 hour ago, Senzelian said:

Ah yes, they're all just bad. That must be it. I guess Apple just bought up all the good ones. All praise Apple!

Apple clearly found it to be necessary. Google too.

That's a weird take heh. All I said is your presumption you put forward is simply flawed and fails on this basis as stated. Google doesn't need it's own hardware and AI engine to do a good job with an AI engine, they could just as easily license something from Nvidia or Xilinx or whoever else. It actually does not matter other than how good and capable the hardware is, past that it's developer capabilities to use the hardware and Google isn't outshining anyone else I'm afraid.

 

Google is a leader in AI and ML but that doesn't mean in every area, in all ways, for everything 

 

The long and short of it is your original statement just was never based in any form of reality and objective analysis, really that's all. I just found it very odd to see you say Google will eventually surpass Apple when there is literally nothing indicating this outcome at all.

 

1 hour ago, Senzelian said:

Not to the extend of Google.

Google can't use all of it for everything and for every purpose, lets not make flawed reasoning here simply on the basis of size.

 

Google could well be using 99.99% of it's datacenter capacity to deliver cat videos, it's not but well you get the point.

 

1 hour ago, Senzelian said:

Yes so fast in fact, that I decided to use Siri instead of Google Assistant - no one said ever.

Yea and Android phones always take superior photos with their superior camera hardware..

 

Quote

Historically, Apple has not had a public reputation for leading in this area. That's partially because people associate AI with digital assistants, and reviewers frequently call Siri less useful than Google Assistant or Amazon Alexa. And with ML, many tech enthusiasts say that more data means better models—but Apple is not known for data collection in the same way as, say, Google

https://arstechnica.com/gadgets/2020/08/apple-explains-how-it-uses-machine-learning-across-ios-and-soon-macos/

 

Sound familiar? Surface level assumptions lead to surface level understanding and conclusions.

 

Edit:

Added some extra comments for better clarification

Edited June 29, 2022 by leadeater

[Senzelian]

30 minutes ago, leadeater said:

Google doesn't need it's own hardware and AI engine to do a good job with an AI engine, they could just as easily license something from Nvidia or Xilinx or whoever else. It actually does not matter other than how good and capable the hardware is, past that it's developer capabilities to use the hardware and Google isn't outshining anyone else I'm afraid.

https://www.anandtech.com/show/17032/tensor-soc-performance-efficiency/5

 

If you really feel to add anything to this, message me. I won't be going on and on and on about this anymore in a topic that has frankly nothing to do with this.

[leadeater]

9 minutes ago, Senzelian said:

https://www.anandtech.com/show/17032/tensor-soc-performance-efficiency/5

 

If you really feel to add anything to this, message me. I won't be going on and on and on about this anymore in a topic that has frankly nothing to do with this.

I think you missed the point there, that was a comment about software. Anyway thanks for the link, for whatever reason when I looked for Google Tensor reviews and benchmarks that did not come up. Well at least it does show their hardware is less bad than it used to be.

 

With that be forewarned to check any critical bits of information about things you are reading

Quote

unfortunately only the Apple variant is lacking CoreML acceleration, thus we should expect lower scores on the A15.

 

Which means that cannot be used to compare against Apple's hardware at all I'm afraid and software.