M3 Chip Generation - Discussion Megathread

[name99]

Like gaming?
View attachment 2278082

I doubt that having higher heat dissipation is detrimental. What would be the implication of the phone having better heat dissipation? Aren't heat dissipation and SoC consumption orthogonal?

Gamers (especially these sorts of games) are obsessed with the idea that what they lust after represents the entire world. It doesn't even represent the world of GAMING! Look at the real hardware that's most common on Steam.

Apple are willing to play along with this fantasy to the extent that it doesn't get in the way of real customers, but no more than that.
Which means they care a LOT about making sure the A17 works well when playing Candy Crush. But they are not going to increase weight, cost, or anything else, PURELY to make Genshin Impact behave differently. Hell, we all know that the people who truly care about these sorts of games will in fact invest in hardware to prevent this. Either they will mount a cooler on the back of the phone, or they will hook the phone up to a TV and blow a normal fan over it.

The iPhone is a phone FIRST. The fact that it can be used (with some rather minor hassle) as a console is a bonus, not the design point...

 

[name99]

Intel and AMD increases performance at a cost of efficiency every year and people criticise them. When Apple dramatically increases wattage to gain a small 10% increase in ST performance with A16 and A17, Apple people loves to say it’s not a problem. ‘Nothing to see here, folks.”

Be very very careful about this gloating. You might be unhappy when MTL is released...

Intel Core 7 165H "Meteor Lake" CPU debuts on Geekbench with over 2.5K score - VideoCardz.com

Intel Core Ultra 7 165H appears on Geekbench The 165H is the same as the Intel Core Ultra 7 1002H. A fresh wave of leaks has surfaced showcasing Intel’s new generation of Core Ultra series processors, powered by the Meteor Lake architecture. In recent weeks, there has been an increase in leaks...

videocardz.com

A17 performance of 2900, as a ~15% increase over A16, ain't looking so bad now, is it?

 

[name99]

To be honest, when a company controls both os and hardware. It is not that hard to manipulate Geekbench scores. It is like a player also the referee.

Proof of this claim? Exactly how do you envisage this "manipulation"occurring?

 

[name99]

The problem is they also opened up a use case mode where getting power delivery from an external display. That whole run it like a console gaming system set up is likely a path to some screwed up phones over time.

The fact the self regulation subsystem doesn't stop things before they get 'horrible high' is a dubious design choice ( if not a defect. ) . If developing this for 'bursty' activity then it should shut down when timeline for a burst should be over. 15-30 minutes is not a 'burst'.

What do you imagine to be the failure mode here?
Presumably Apple stops the phone before it hist danger temperatures AND presumably Apple stops the phone before it hits battery current limits.
So your belief is Apple should throttle the phone just because?

People have been demanding for years that Apple give them something like DEX, where the phone plugs into a monitor and can be used as a computer. Now Apple gives you that, and your first response is "I want it to run slower, like a phone not a computer"???

 

[PgR7]

If you can't even explain what you tested and what your "efficiency" score refers to, no-one is going to take you seriously...
Maybe spend some time first reading the room and reading the thread? This is a thread for people who care about engineering, not for d**k-measuring.

We only have a few measurements of power consumption and its from Geekerwan so you dont have to investigate much to know what benchmarks im refering to, my Snapdragon 870 with Android 13 debloated and undervolt on GPU makes 3000 points in Geekbench 5 with 4w power consumption, A17 Pro makes 6200 points geekbench 5 with 14W, it makes about 5600 points too with 9.5W consumption. In GFXBench 1440p aztek vulkan I get 25 fps with 2.8w the A17 Pro does 64 fps with 11.5W.

Its only when you activate low power consumption that it gets better efficiency numbers

 

[APCX]

We only have a few measurements of power consumption and its from Geekerwan so you dont have to investigate much to know what benchmarks im refering to, my Snapdragon 870 with Android 13 debloated and undervolt on GPU makes 3000 points in Geekbench 5 with 4w power consumption, A17 Pro makes 6200 points geekbench 5 with 14W, it makes about 5600 points too with 9.5W consumption. In GFXBench 1440p aztek vulkan I get 25 fps with 2.8w the A17 Pro does 64 fps with 11.5W.

Its only when you activate low power consumption that it gets better efficiency numbers

This sounds like a wild guess. How do you know your measurements are accurate? How do you know geekerwan’s measurements are accurate?

 

[aibloop]

I am finally going Apple Silicon in 2024 with a Mac Studio M3 Ultra OR... Depending on how good the Mac Studio M3 Max is, I might opt for saving some cash w it.


M3 RT raytracing cores + Neural engine cores working in tandem to upres/denoise... Killer features for M3... And there goes the longtime Nvidia edge! (I doubt it will be faster then a 4090 btw, just close enough is good enough for me)

M1/M2 is comparable to a AMD workstation. BUT since Nvidia offers RT raytracing logic and good denoise/upscaling machinelearning... Apple HAS been incomplete feature-wise against the PC, but w M3 they gain the features missing. OFC, it could be that the RT raytracing is lacking, however I doubt it. M3 will make a Mac workstation a great renderstation.

2024 is looking good for Apple.

(Animator, VFX raytracing/3d Rendering/unreal RT etc)

 

[Boil]

- M3 is likely 10 GPU cores, Pro+ is 20, Max+ is 40.
Given the A17, one might have expected these numbers to be more like 12/24/48? But there's always an area budget, and maybe this year the area is devoted to scaling logic more than just ramping up GPU core count?

The newly-added hardware ray-tracing is gonna need some space...?

- 16 core NPU. It's not widely remarked upon how (very unlike the GPU or CPU) the NPU core count remains the same from A to Max. Kinda strange that, and not sure what it means. Maybe that, for now anyway, Apple see this as mainly about facilitating language UI, and there's only one person talking whether you have an iPhone or a Mac Studio?

Didn't Apple say the Neural Engine in A17 Pro is now "twice as fast"...?

 

[NT1440]

The newly-added hardware ray-tracing is gonna need some space...?

My understanding is that’s baked into the new GPU itself rather than a separate chiplet. So from the perspective of space, RT is essentially “free” because that area was always going to be used for the GPU.

 

[PgR7]

This sounds like a wild guess. How do you know your measurements are accurate? How do you know geekerwan’s measurements are accurate?

You can doubt about every benchmark run, I had my doubts too but just by feeling how hot the back of the phone is it makes sense, I thought if its taking the consumption of the soc only, but if i increase the brightness of the screen it increases the consumption, Geekerwan supposedly tears down the phone and simulates with a digital power supply the consumption, but I think that in the 15 PM review he used software telemetry instead.

The thing is this SD870 makes 750 points in multicore GB5 per watt, but the A17 Pro in low power mode makes about 1470 points per watt, about 590 points per watt with 9,5w and about 440 points per watt using 14W

 

[APCX]

You can doubt about every benchmark run, I had my doubts too but just by feeling how hot the back of the phone is it makes sense, I thought if its taking the consumption of the soc only, but if i increase the brightness of the screen it increases the consumption, Geekerwan supposedly tears down the phone and simulates with a digital power supply the consumption, but I think that in the 15 PM review he used software telemetry instead.

The thing is this SD870 makes 750 points in multicore GB5 per watt, but the A17 Pro in low power mode makes about 1470 points per watt, about 590 points per watt with 9,5w and about 440 points per watt using 14W

Sorry if I missed it, can you clarify how you are making these measurements?

As for Geekerwan, it seems as though they cooled the phone, allowing the power to hit a level it wouldn’t be able to normally. If you look at other tests, the phone doesn’t hit anywhere near 14 watts.

 

[PgR7]

Sorry if I missed it, can you clarify how you are making these measurements?

As for Geekerwan, it seems as though they cooled the phone, allowing the power to hit a level it wouldn’t be able to normally. If you look at other tests, the phone doesn’t hit anywhere near 14 watts.

There is an app that gets you an overlay where you can get performance data, I can get the power consumption of the battery all time. Geekerwan did another measurement without cooling 5600 points and 9.5w, its still less efficient.

 

[APCX]

There is an app that gets you an overlay where you can get performance data, I can get the power consumption of the battery all time. Geekerwan did another measurement without cooling 5600 points and 9.5w, its still less efficient.

Right but we’re talking about the power consumption of the cores, not the battery. Battery consumption will take place as a result of power use by a variety of components within the phone.

 

[PgR7]

Right but we’re talking about the power consumption of the cores, not the battery. Battery consumption will take place as a result of power use by a variety of components within the phone.

And as I told you, its the battery consumption, not SoC, I can increase the power consumption just by increasing the brightness of the screen, or if i get it to the minimun it lowers

 

[leman]

My understanding is that’s baked into the new GPU itself rather than a separate chiplet. So from the perspective of space, RT is essentially “free” because that area was always going to be used for the GPU.

Of course it’s not free, it’s additional functionality that occupies die area. Preliminary analysis says that the GPU core die area has increased by 20% - and that with density savings from the new process. Although it’s unclear how much of that is taken by the new RT function and how much by increased caches/register files etc.

 

[APCX]

And as I told you, its the battery consumption, not SoC, I can increase the power consumption just by increasing the brightness of the screen, or if i get it to the minimun it lowers

I think perhaps there is some miscommunication. I realize you are referring to the battery consumption. That’s the issue. The topic is core power consumption: that is what you should be measuring, otherwise we have no basis for comparison.

 

[PgR7]

I think perhaps there is some miscommunication. I realize you are referring to the battery consumption. That’s the issue. The topic is core power consumption: that is what you should be measuring, otherwise we have no basis for comparison.

But Geekerwans test the battery power consumption, originally he used to remove the battery of the phone and puts a digital power supply and gets the consumption figures from there, but in the 15 PM I think he used software telemetry as I did, If he got core power consumption is even better for me because I get the core consumption + the rest of the phone and It still beats the A17 Pro in efficiency

 

[NT1440]

Of course it’s not free, it’s additional functionality that occupies die area. Preliminary analysis says that the GPU core die area has increased by 20% - and that with density savings from the new process. Although it’s unclear how much of that is taken by the new RT function and how much by increased caches/register files etc.

I meant “free” as in RT was always going to be part of this GPU so the space was factored into the design from its inception. Rather than “where can we bolt this on”, it was always going to be part of the floor plan, whatever size that ended up being. That was (in my opinion) why RT would only make its appearance on 3nm.

Does that make sense?

 

[APCX]

But Geekerwans test the battery power consumption, originally he used to remove the battery of the phone and puts a digital power supply and gets the consumption figures from there, but in the 15 PM I think he used software telemetry as I did, If he got core power consumption is even better for me because I get the core consumption + the rest of the phone and It still beats the A17 Pro in efficiency

You can’t speak meaningfully about power consumption if you can’t measure it. It seems like you can’t measure yours, and it remains unclear if and how Geekerwan measures it.

We need to speak about efficiency at a given performance level. You may think you are more efficient because at a lower score you use less power than an A17 at a higher performance level, but you don’t know how much energy the A17 expends to match your performance, and just as importantly, you don’t know how much energy your phone would use to match the A17 if it could!

It doesn’t scale linearly.

 

[PgR7]

You can’t speak meaningfully about power consumption if you can’t measure it. It seems like you can’t measure yours, and it remains unclear if and how Geekerwan measures it.

We need to speak about efficiency at a given performance level. You may think you are more efficient because at a lower score you use less power than an A17 at a higher performance level, but you don’t know how much energy the A17 expends to match your performance, and just as importantly, you don’t know how much energy your phone would use to match the A17 if it could!

It doesn’t scale linearly.

I did measure it, and yeah I know that in scenarios with limited performance it will be different, but I would say that that A17 looks like wants to boost a lot to feel very snappy and it will have more consumption in burst loads, yeah it will take my phone double the time to load a webpage or an app but it will consume less than half battery

 

[Confused-User]

I did measure it, and yeah I know that in scenarios with limited performance it will be different, but I would say that that A17 looks like wants to boost a lot to feel very snappy and it will have more consumption in burst loads, yeah it will take my phone double the time to load a webpage or an app but it will consume less than half battery

Here is the heart of the problem. You don't understand what nonlinear means.

The measurement you're describing above - even if you could do it reliably, which you can't - is not a valid comparison of efficiency.

If you want to compare efficiency, you have to compare energy usage while keeping time constant. That is, you'd have to run the A17 slower, so that it takes the same time to complete the job as your phone. Then you'd compare energy used, and that would tell you which is more efficient.

There are some other possibilities, too. You could measure total performance while consuming the same amount of energy per second. That's not quite right, I think, but it's probably close enough.

In fact, measuring both at peak performance might, averaged over a large number of use cases, still favor the iphone (note that I'm not saying A17 here because this is no longer just about the SoC) because, being much faster in the race to idle, it might reduce utilization of other energy consumers in the phone. But then that's not about efficiency any more, or at least not efficiency of the chip.

 

[PgR7]

Here is the heart of the problem. You don't understand what nonlinear means.

The measurement you're describing above - even if you could do it reliably, which you can't - is not a valid comparison of efficiency.

If you want to compare efficiency, you have to compare energy usage while keeping time constant. That is, you'd have to run the A17 slower, so that it takes the same time to complete the job as your phone. Then you'd compare energy used, and that would tell you which is more efficient.

There are some other possibilities, too. You could measure total performance while consuming the same amount of energy per second. That's not quite right, I think, but it's probably close enough.

In fact, measuring both at peak performance might, averaged over a large number of use cases, still favor the iphone (note that I'm not saying A17 here because this is no longer just about the SoC) because, being much faster in the race to idle, it might reduce utilization of other energy consumers in the phone. But then that's not about efficiency any more, or at least not efficiency of the chip.

I know what you saying and there is not much power consumption data of the A17 where i can investigate that, in thinking about something like 3DMark Wild Life extreme stress test, the test last 20 minutes, we can see the score and the battery percentage consumed. Im thinking too of power consumption in Genshin Impact, geekerwan measured performance and power consumption there. But he tests that in a city that you unlock later in the game

 

[AgentMcGeek]

As always, how do you know that we don't get a "perf-to-space improvement on the GPU side"? You don't see the performance boost of what you don't test...
As GPUs get more general (see what nVidia is doing) part of what is going to happen on Apple designs is following innVidia's footsteps, eg
- making the lanes of the GPU capable of executing independently (as in Volta),
- providing more powerful synchronization primitives, and
- allowing Scratchpad storage to be shared across cores (as in Hopper).
- Hell maybe even some FP64 support.

These don't excite the games crowd, but they are an essential part of the evolution of GPGPU.

I was just making a rough assessment based on Apple’s own claims + what we know so far: +20% more cores, for +20% more space, for +20% perf improvements.

I also mentioned that some new features or capabilities such as RT might be one of the reasons why we’re not seeing prima facie anything better than a 1:1 ratio. As to what these other features might be, you covered it pretty well.

 

[leman]

Regarding the GPU, I did a quick test and A17 Pro delivers 2TFLOPs compute throughput. That gives us the GPU clock of 1.3Ghz, same as A15.

 

[diamond.g]

Regarding the GPU, I did a quick test and A17 Pro delivers 2TFLOPs compute throughput. That gives us the GPU clock of 1.3Ghz, same as A15.

You think adding a GPU core was more power efficient than bumping the GPU clock up 10-20%?