Thread: iPhone 15 / Apple Watch 9 Event

[dada_dave]

Another thing to keep in mind is that the M3 will likely have the same cores as the A17 while the M2 had the same cores as the A15, so the M3 ST perf could effectively get a near 30% jump in performance over the M2, which I think is what @Cmaier was predicting? Though perhaps weighted more towards clock speed than IPC than he was thinking if I remember correctly.

GPU could be a very nice bump too depending on how the upgrades translate. I think a user -was it at @theorist9 ? - did a back of the envelope calculation for the M3’s blender improvement but I think he did it from the M2 using the A17 to A16 ratio whereas it would actually be the A17 to A15 ratio.

 

[Altaic]

Another thing to keep in mind is that the M3 will likely have the same cores as the A17 while the M2 had the same cores as the A15, so the M3 ST perf could effectively get a near 30% jump in performance over the M2, which I think is what @Cmaier was predicting? Though perhaps weighted more towards clock speed than IPC than he was thinking if I remember correctly.

That’s a good point. Also, if the A17 and M3 were designed in tandem, it could be that we’re also not seeing the expected uplift because A17 may have been designed for both the N3B and N3E nodes, where the later M3 lineup would be optimized for N3E. IIRC, N3B basically uses FinFlex 2-2, but has slightly higher density design rules due to double patterning and other process enhancements. Perhaps with the M3 (or M3 Pro/Max/etc.), there’ll be higher perf due to better utilization of other FinFlex options offered by N3E. I also came across a rumor recently that claimed the iPad Pro would get the M3 Pro, though I’m still ruminating on the implications of that.

GPU could be a very nice bump too depending on how the upgrades translate. I think a user -was it at @theorist9 ? - did a back of the envelope calculation for the M3’s blender improvement but I think he did it from the M2 using the A17 to A16 ratio whereas it would actually be the A17 to A15 ratio.

That was me, and it’s an admittedly naive extrapolation of M2 (A15 design) -> M3 (A17 design).

 

[dada_dave]

That’s a good point. Also, if the A17 and M3 were designed in tandem, it could be that we’re also not seeing the expected uplift because A17 may have been designed for both the N3B and N3E nodes, where the later M3 lineup would be optimized for N3E. IIRC, N3B basically uses FinFlex 2-2, but has slightly higher density design rules due to double patterning and other process enhancements. Perhaps with the M3 (or M3 Pro/Max/etc.), there’ll be higher perf due to better utilization of other FinFlex options offered by N3E. I also came across a rumor recently that claimed the iPad Pro would get the M3 Pro, though I’m still ruminating on the implications of that.



That was me, and it’s an admittedly naive extrapolation of M2 (A15 design) -> M3 (A17 design).

Aye though I think you gave it the 20% from A16->A17 when A15->A17 GB looks like 36%. However of course that’s with an extra core and the percentage change in core count for the M-series GPU is likely to be different. So 20% may be the better number to use for the rough estimate anyway.

Just some napkin math with blender open data (v 3.6.0): M2 Ultra scores 3447.15; naively extrapolating 20% GPU compute improvement with 4x raytracing speed results in a score of ~16546.32. RTX 4090 scores 13100.33. Even without any exciting M3 surprises, that's looking pretty good.

 

[Altaic]

Aye though I think you gave it the 20% from A16->A17 when A15->A17 GB looks like 36%.

Ah yes, you’re correct of course— I forgot to mention that

Either way, I feel pretty good about what the M3 lineup may bring (assuming it’s based on the A17).

 

[Jimmyjames]

Another thing to keep in mind is that the M3 will likely have the same cores as the A17 while the M2 had the same cores as the A15, so the M3 ST perf could effectively get a near 30% jump in performance over the M2, which I think is what @Cmaier was predicting? Though perhaps weighted more towards clock speed than IPC than he was thinking if I remember correctly.

GPU could be a very nice bump too depending on how the upgrades translate. I think a user -was it at @theorist9 ? - did a back of the envelope calculation for the M3’s blender improvement but I think he did it from the M2 using the A17 to A16 ratio whereas it would actually be the A17 to A15 ratio.

Yes, that’s very interesting. I had completely forgotten we’ve skipped the M chip based on the A16. That should indeed give a significant jump.

A14->M1 Single Core = ~2045 -> 2375 ——> Delta of 329 (GB 6)
A15->M2 Single Core = ~2267 -> 2734 ——> Delta of 467 (GB 6)
A17->M3 Single Core = ~2900 -> 3400 ——> Delta of 500. (GB 6)

Obviously a guess! If the M3 is more scalable, then perhaps more.

Edit: all scores are for a 16” MacBook Pro Max (M1/M2).

 

[Jimmyjames]

If we extend the difference between A15 -> A17 to estimate multi core and gpu scores

Multi Core
A15 -> 5462, A17 -> 7199 = 31%
M2 -> 14500, M3 -> 19000

GPU
A15 -> 19941, A17 -> 27158 = 36%
M2 Max -> 137396, M3 -> ~187000. 163000

Not bad!

Edit: Revising the GPU score downwards. Foolishly forgot that scaling is easier to 6 cores than 38!

 

[dada_dave]

Yes, that’s very interesting. I had completely forgotten we’ve skipped the M chip based on the A16. That should indeed give a significant jump.

A14->M1 Single Core = ~2045 -> 2375 ——> Delta of 329 (GB 6)
A15->M2 Single Core = ~2267 -> 2734 ——> Delta of 467 (GB 6)
A17->M3 Single Core = ~2900 -> 3400 ——> Delta of 500. (GB 6)

Obviously a guess! If the M3 is more scalable, then perhaps more.

If we extend the difference between A15 -> A17 to estimate multi core and gpu scores

Multi Core
A15 -> 5462, A17 -> 7199 = 31%
M2 -> 14500, M3 -> 19000

GPU
A15 -> 19941, A17 -> 27158 = 36%
M2 Max -> 137396, M3 -> ~187000.

Not bad!

Caveats: single core for the M2 Max/Ultra was higher than for the base M2. Same may be true for the M3. The above looks more like higher rank M2/M3 rather than base M2/M3s. Also between A15 and A17 there was a change in GPU core count which may or may not be reflected in M2 base/pro/max/ultra to M3 base/pro/max/ultra.

If my memory serves me right, the current leaks - which have slightly more believability as they come from developer strings but just as with Zen 5 or Intel come with additional caveats that we don’t really know the structure - there will be a change in CPU core counts for Pro/Max/Ultra chips but also possibly a smaller percentage change in GPU core count than reflected here. So we’ll see, but even with all these caveats, M3 should be a very nice overall bump compared to M2.

 

[Jimmyjames]

Caveats: single core for the M2 Max/Ultra was higher than for the base M2. Same may be true for the M3. The above looks more like higher rank M2/M3 rather than base M2/M3s. Also between A15 and A17 there was a change in GPU core count which may or may not be reflected in M2 base/pro/max/ultra to M3 base/pro/max/ultra.

If my memory serves me right, the current leaks - which have slightly more believability as they come from developer strings but just as with Zen 5 or Intel come with additional caveats that we don’t really know the structure - there will be a change in CPU core counts for Pro/Max/Ultra chips but also possibly a smaller percentage change in GPU core count than reflected here. So we’ll see, but even with all these caveats, M3 should be a very nice overall bump compared to M2.

Indeed. I should have specified the scores quoted are for a 16” MacBook Pro Max.

Edit. I’m far more confident about the single core score prediction than the multi core or gpu scores.

 

[dada_dave]

Indeed. I should have specified the scores quoted are for a 16” MacBook Pro Max.

Right

Here’s @Cmaier ’s original prediction:

I’m hearing single-core performance in the next generation is around 130% of M2 (high performance cores), and that the low power cores are much more efficient. 130% would put apple back on their historical 20% year-over-year performance improvement, more or less.

Core count-wise it looks like:

M3: 8 CPU, 10 GPU
M3 Pro: 12 CPU, 18 GPU (vs. 10/16 M2 Pro)
M3 Max: 16 CPU (12 perf/4 effic), 40 GPU
M3 Ultra: (up to) 32 CPU (24 perf/8 effic), 80 GPU

Sources:

Apple testing new MacBook Pro with 40-core GPU M3 Max chip

A Bloomberg report on Sunday revealed that Apple has been testing the base version of its new M3 chip with...

9to5mac.com

M3 Ultra specs reportedly include significant boost in CPU cores, up to 80-core GPU - 9to5Mac

More details about what to expect from the M3 lineup of chips continue to emerge. We’re expecting the first M3...

9to5mac.com

Should be very nice performance boosts across the board.

Actually looks right on the money for single core performance improvements.

 

[Jimmyjames]

Right

Here’s @Cmaier’s original prediction:



Actually looks right on the money for single core performance improvements.

Absolutely. Now the question is, when do they arrive? I do hope it’s going start this year. May well be delayed though.

 

[Jimmyjames]

Back to the iPhone 15 Pro. I hadn’t realised how poor the results for the 14 pro were on gfxbench. It’s about the same as the regular 14 or even a little behind. It’s also well behind some Android devices.

Anyone know what’s going on here? Is it gfxbench, thermal throttling on the iPhone, just very good gpus in the 8 gen 2 Snapdragons?

Also I just checked out the Geekbench ML (NPU) scores. There doesn’t seem much improvement there. Certainly not a 2x increase.

14 Pro Max -> 3212
15 Pro Max -> 3600.

 

[dada_dave]

Back to the iPhone 15 Pro. I hadn’t realised how poor the results for the 14 pro were on gfxbench. It’s about the same as the regular 14 or even a little behind. It’s also well behind some Android devices.

Anyone know what’s going on here? Is it gfxbench, thermal throttling on the iPhone, just very good gpus in the 8 gen 2 Snapdragons?

Also I just checked out the Geekbench ML (NPU) scores. There doesn’t seem much improvement there. Certainly not a 2x increase.

14 Pro Max -> 3212
15 Pro Max -> 3600.

It may be model dependent. Does it have sub scores?

 

[Jimmyjames]

It may be model dependent. Does it have sub scores?

I found a breakdown for the 15 pro, but weirdly, nothing for the 14 pro. Scores, but no detailed sub scores.

 

[dada_dave]

I found a breakdown for the 15 pro, but weirdly, nothing for the 14 pro. Scores, but no detailed sub scores.

Unfortunate

 

[Jimmyjames]

So close...

iPhone16,1 - Geekbench

Benchmark results for an iPhone16,1 with an ARM processor.

browser.geekbench.com

EDIT: just realised it’s also faster (in terms of this benchmark, sustained performance is yet to be seen) than the 8 core iMac Pro from 2017 and the 2019 8 core Mac Pro. Crazy.

Also if 2999 is indicative of A17 single core scores, that’s a 480 point increase on the A16 which is the largest ever (previous best 421 from A12 -> A13).

 

[Jimmyjames]

What is going on with Gfxbench? The 15 pro max is slower than the 14, not the 14 pro, the 14.

 

[Andropov]

Geekbench compute result looks promising: 27433 points. That's a +23% improvement in GPU, slightly above what Apple promised.

 

[exoticspice1]

What is going on with Gfxbench? The 15 pro max is slower than the 14, not the 14 pro, the 14.

I would wait for 3D mark benchmark suite and also actual gameplay from games.

I never liked Gfxbench. gfx bench is also piss poor on RTX cards. Maybe the new shader arch is not suited to GFX bench?

 

[dada_dave]

What is going on with Gfxbench? The 15 pro max is slower than the 14, not the 14 pro, the 14.

Same exact score as the A16 in the 14 Pro too. Those high A15 results in the 14 and 14 max are a bit suspicious. Previous A15 results were lower. Otherwise not sure.

 

[Jimmyjames]

More Geekbench weirdness. There are now 66 GB 6.2 cpu scores. That seems high for review units. Are people testing multiple times? The spread is very wide also. Single core from 2543->2999 and multi core from 6677->7779. Obviously it’s better to wait until the phone has done its indexing etc at the start, but that really seems like a wide range. I wonder if there are some fakes in there.

I’m quite suspicious of the result I posted to earlier here

Thread: iPhone 15 / Apple Watch 9 Event

Another thing to keep in mind is that the M3 will likely have the same cores as the A17 while the M2 had the same cores as the A15, so the M3 ST perf could effectively get a near 30% jump in performance over the M2, which I think is what @Cmaier was predicting? Though perhaps weighted more...

techboards.net

It has the highest single and multi core scores.