M3 Chip Generation - Discussion Megathread

[iPadified]

Please take your tribal warrior nonsense somewhere else and leave us engineers to discuss reality.

I am a professor at an engineering university department since 20 years. Here are my observations:
*There are plenty of engineers that are tribal warriors. Being a tribal warrior is how you survive as an academic researcher
*Engineers have a very limited view on the reality that seldom encompass the whole picture

As far as I can see, none of us have developed a chip of this complexity for this node and is therefore clueless about the technical roadblocks, (human) resource restrictions, market focus and management decisions. The three last ones are often ignored by engineers in discussions as evident also here.

 

[JPack]

In fairness, Geekerwan also clearly said in the video that they made this video in 3 days, and that all the data in it are preliminary and may have inaccuracies. I will take that any day over reviewers who are both rushing out and not admitting that their reviews may have inaccuracies, which is sadly the norm.

No, they did not say that.

Geekerwan said the A17 Pro microarchitecture layout diagram might have mistakes, not their testing data. Literally no one else in the industry produced such a feature diagram for A17 Pro in such a short period. They consulted Andrei Frumusanu (Anandtech) and Maynard Handley (QuickTime developer). Both these guys are legends. Handley reverse engineered and documented M1 in a 350 page document.

 

[jeanlain]

I’m not sure why any other would be chosen.

Because when you look at individual results, you see plenty around 2630 and very few close to 2522. I suppose some unusually low like scores (obtained by phones working in the background) lowered the average a lot. That's why one should not use average, but medians.

 

[jeanlain]

Being a tribal warrior is how you survive as an academic researcher

I am an academic researcher and I disagree. Tribal warriors don't go very far and are immediately categorized as idiots.

 

[Xiao_Xi]

For me, there are too many conflicts with previously established baselines. The power consumption for A14/A15 reported by Geekerwan is dramatically different from what Anandtech reported for similar tests (by almost 80%)! E.g. Andrei measured over 4Watts for these chips running SPEC (and close to 5W in FP tests), Geekerwan only reports under 3 watts (7:18 in the video). These numbers do paint a daunting picture for Apple's 3N product, showing A17 pro being 20% less efficient than A16.

Were those baselines verified and replicated or just established by Anandtech?

 

[leman]

Were those baselines verified and replicated or just established by Anandtech?

Just Anandtech from what I know. Still, if you have two widely different sets of numbers, making an informed solution appears difficult.

At any rate, it seems that Anandtechs numbers are consistent with what Geekerwann is reporting as mainboard power draw. As I wrote in my previous post, if we go from these numbers, A17 Pro appears to be a performance-oriented architecture that aims to keep the power efficiency of previous generations while offering higher peak performance.

 

[altaic]

No, they did not say that.

Geekerwan said the A17 Pro microarchitecture layout diagram might have mistakes, not their testing data. Literally no one else in the industry produced such a feature diagram for A17 Pro in such a short period. They consulted Andrei Frumusanu (Anandtech) and Maynard Handley (QuickTime developer). Both these guys are legends. Handley reverse engineered and documented M1 in a 350 page document.


View attachment 2270260

@name99 Is that true? I’ve drawn quite different conclusions from this review than some other folks here, but either way it’d help know if they actually had consulted with someone who’s done proper research.

 

[jeanlain]

For me, there are too many conflicts with previously established baselines. The power consumption for A14/A15 reported by Geekerwan is dramatically different from what Anandtech reported for similar tests (by almost 80%)! E.g. Andrei measured over 4Watts for these chips running SPEC (and close to 5W in FP tests), Geekerwan only reports under 3 watts (7:18 in the video). These numbers do paint a daunting picture for Apple's 3N product, showing A17 pro being 20% less efficient than A16.

Their monitoring of clock speed and power consumption at the 7min mark appears quite convincing though. The A17 drops in clock speed while the A16 does not. This suggest throttling due to thermal problems, despite the supposedly improved chassis. (That's a very in-depth review BTW.)
That doesn't look good.
And what's with the Solar Bay result? Isn't the A17 supposed to be 3X faster than the A16 for ray tracing?

 

[jeanlain]

Now, at around 8:20 he reports a different set of numbers (I assume total device power draw?), which are much closer to Anandtech measurements. According to them, in SPECint A17 Pro is 10% less efficient than A16, 5% less efficient than A15, and about as efficient as A14. In SPECfp, A17 is as efficient or more efficient than there predecessors.

I suppose these correspond to multicore results, i.e., including the efficiency cores.
The P-core are the problematic ones.

 

[leman]

Their monitoring of clock speed and power consumption at the 7min mark appears quite convincing though. The A17 drops in clock speed while the A16 does not. This suggest throttling due to thermal problems, despite the supposedly improved chassis. (That's a very in-depth review BTW.)
This doesn't look good.

Not necessarily. If A17 is indeed a performance-oriented design, the observed throttling might be an energy-saving feature to better balance power consumption on sustained workloads. It's a phone after all. You'll have more three-room in a Mac.

And what's with the Solar Bay result? Isn't the A17 supposed to be 3X faster than the A16 for ray tracing?

Yeah, that was a bit underwhelming, right? This would translate to M3 Max scoring just between a 4060 and 4070 (desktop).

 

[altaic]

That doesn't look good.

Looks great to me. Seems like the new cores will scale nicely outside of a phone thermal envelope, no?

 

[leman]

I suppose these correspond to multicore results, i.e., including the efficiency cores.
The P-core are the problematic ones.

No, it's singe-core. I used a translator to read the annotations and they say "mainboard power draw"

 

[jeanlain]

Looks great to me. Seems like the new cores will scale nicely outside of a phone thermal envelope, no?

Scale nicely? I'm not sure how we can conclude that. We don't know how power consumption will evolve at higher clock speed.
The fact that the performance increase of the A17 P-core is achieved by a higher increase in power consumption isn't particularly good news.

 

[jeanlain]

Yeah, that was a bit underwhelming, right?

Indeed it is. Apple said up to 4X faster and showed a >3X difference in Solar Bay (8 fps vs 30 fps). It seems that the 8 fps figure was indeed achieved in "software", as they say (not even using the GPU to accelerate ray tracing?) because the A16 achieves 20 fps, not 8. How misleading.

 

[leman]

Scale nicely? I'm not sure how we can conclude that. We don't know how power consumption will evolve at higher clock speed.
The fact that the performance increase of the A17 P-core is achieved by a higher increase in power consumption isn't particularly good news.

It would of course be better if the improvements were in power efficiency. But I think what's rather interesting is that per/watt remained pretty much the same from the 5N generation. A At this point I expect that we will see linear (or slightly above linear) power consumption increases across the board, up to 8-10 watt single-core power draw (from todays 5-6 watts) for roughly 20-30% improvements in performance. I was speculating that Apple might go this route some time ago. It would make more sense for a Mac product IMO.

The bottomline is that we are kind of back to square one in terms of expectations. I hoped that the new u-arch will bring significant IPC improvements — that would make predictions easy. I think it's a safe conclusion that this didn't happen. They appear to be going for maintaining perf/watt while increasing the dynamic power range. It's as you say — we don't know how the power consumption will scale with higher clock. So I'm afraid that we will have to wait and see when the Mac product is released.

 

[jeanlain]

Not necessarily. If A17 is indeed a performance-oriented design, the observed throttling might be an energy-saving feature to better balance power consumption on sustained workloads. It's a phone after all. You'll have more three-room in a Mac.

Sure, but look at the power draw. It's much higher than the A16, and when the clock speed is reduced to the A16 clock speed, the power consumption is also reduced at the levels of the A16.
My point is that I have no reason to doubt their results. The power consumption of the A17 P-core is quite a bit higher than that of the A16 at nominal clock speed, and that's disappointing.

 

[leman]

Indeed it is. Apple said up to 4X faster and showed a >3X difference in Solar Bay (8 fps vs 30 fps). It seems that the 8 fps figure was indeed achieved in "software", as they say (not even using the GPU to accelerate ray tracing?) because the A16 achieves 20 fps, not 8. How misleading.

Did Apple use Solar Bay benchmark though? I was under impression that they were using some sort of internal test. It's very much possible that the hardware RT solution performance increases with scene complexity and amount of RT work required. Solar Bay is not particularly demanding. My M1 Max reaches 70 FPS without hardware raytracing.

 

[leman]

Sure, but look at the power draw. It's much higher than the A16, and when the clock speed is reduced to the A16 clock speed, the power consumption is also reduced at the levels of the A16.
My point is that I have no reason to doubt their results. The power consumption of the A17 P-core is quite a bit higher than that of the A16 at nominal clock speed, and that's disappointing.

Could be a different balancing approach to enable faster clocks. The power draw isn't too bad IMO and increases semi-linearly with performance (not by power law as often the case). Look at the SoC power draw (around 8:20 if I remember correctly), not the per-core one, that one is pretty meaningless for the big picture.

Edit: what I am trying to say is that Apple might have decided to keep the performance of A16 at the same power consumption while increasing the max attainable performance, instead of improving the performance at the same power and limiting the max attainable performance. I fully agree with you that it would be better if we had both. Maybe it was not attainable though or subject to some other constraints. I would be also very curious to know how much cache A17 Pro has. If I remember correctly, A15 and A16 use very large SLC to improve power efficiency.

 

[jeanlain]

Did Apple use Solar Bay benchmark though? I was under impression that they were using some sort of internal test. It's very much possible that the hardware RT solution performance increases with scene complexity and amount of RT work required. Solar Bay is not particularly demanding. My M1 Max reaches 70 FPS without hardware raytracing.

It looks like a different scene from Solar Bay. I don't know why their fps number is so low on the left. Maybe it's a different version of the tool which does more ray-tracing stuff. I'm under the impression that Solar Bay only uses RT for reflections.

 

[leman]

It looks like a different scene from Solar Bay. I don't know why their fps number is so low on the left. Maybe it's a different version of the tool which does more ray-tracing stuff. I'm under the impression that Solar Bay only uses RT for reflections.

The Solar Bay version I have here does not have that scene. And there is much more going on in Apple's video than reflections. It appears to have global illumination for example.

 

[Jamie I]

The M3 Max is supposed to get 4 extra performance cores and M3 Ultra is supposed to get 8 extra. What does this mean for performance and power consumption?

 

[leman]

The M3 Max is supposed to get 4 extra performance cores and M3 Ultra is supposed to get 8 extra. What does this mean for performance and power consumption?

Moar of everything?

 

[altaic]

So I'm afraid that we will have to wait and see when the Mac product is released.

I’m thinking the M3 iPad Pro will be the bellwether, considering thermals. Apple has always been pushing the iPad as a computer replacement, and maybe this one will be, perf-wise and perhaps gaming-wise.

 

[Andropov]

Sure, but look at the power draw. It's much higher than the A16, and when the clock speed is reduced to the A16 clock speed, the power consumption is also reduced at the levels of the A16.

It would be interesting to know how much the performance is reduced when clocked at the A16 frequency. It's hard to draw any conclusions otherwise.

 

[Jamie I]

A15: 2P+4E and 5 GPU cores
M2: 4P+4E and 10 GPU cores
M2 Pro: 8P+4E and 19 GPU cores
M2 Max: 8P+4E and 38 GPU cores
M2 Ultra: 16P+8E and 76 GPU cores

A17 Pro: 2P+4E and 6 GPU cores
M3: 4P+4E and 10 GPU cores
M3 Pro: 8P+6E and 20 GPU cores
M3 Max: 12P+4E and 40 GPU cores
M3 Ultra: 24P+8E and 80 GPU cores

I really do wonder what Apple is doing with the M3 lineup. M3 doesn’t have double the GPU cores as A17 Pro.

M3 Pro has 2 extra efficiency cores but M3 Max has 4 extra performance cores but doesn’t have the 2 extra efficiency cores. I’m so confused.