Leak of 3nm A17 chip - single core 60% faster?!

[3Rock]

It was also mentioned that was overlooked, that pushing it that fast would cause the chip to overheat, and that there was no cooling that could help keep it down to acceptable levels so Apple would not include that in the system, unless there was a beefed up cooling system.

so why do even start a thread on this if it’s fake?

6:30 on vid…

 

[Sn0wLe0pard]

Perhaps this is an engineering sample. In order to be able to put M3 in colorless Air models, this engineering sample can be «strangle» for example on 30%.
Or the leaker made a mistake, and is it generally an engineering sample of the M3 Pro.

 

[Scarrus]

There's nothing here that's right.

IPC improvements from A15->A16 were close to zero. Almost all gains were from pushing clocks. That's because the A16 was a stopgap design forced on them when their N3 ("3nm") design had to be put on hold because N3 wasn't ready.

4nm to 3nm would be a reduction of 25%, not 20%, but that's irrelevant since nothing in the A16 or A17 is 3, 4, or 5 nm in size. Those "sizes" are marketing nonsense that has taken over the industry, and they are entirely divorced from reality.

Nonetheless the 60% number isn't entirely impossible. I think it's entirely plausible, in fact, but it's misunderstood. If it's true, it's likely the score Apple was able to get by pushing clocks as far as they could go on their A17 cores, *in a desktop form factor*. In other words, you're looking at a plausible single-core score for an M3 Mac Pro right there.

How do you get 60% improvement? Not so hard. You get perhaps 15% from going to N3. You get the rest by redesigning the core to run at faster clocks (which the M1 and M2 almost certainly can't), along with moderate IPC improvements. *IF* the number is real - which it may not be - then they might be getting anywhere from 5-20% from IPC gains, and then the rest would come from boosting clocks to around 4.5GHz, give or take, depending on the IPC gains.

The multicore score is weird - you wouldn't expect scaling to be so bad - but if you imagine an engineer getting an early A17 back from TSMC and wondering "how far can I push this if I give it some cooling?" then that might make a lot of sense. The cooling might be good enough to run one core at 4.5GHz, but not enough to let all the cores run at top speed. Also, the A17 uncore is presumably not designed for that speed and may be showing severe deficiencies when pushed that far. One would expect an M3Ultramegawhizzbang to have a far more suitable uncore, so hopefully this is not an indication of how multicore will scale on the M3. Again, if the numbers are even real.

You could also design the CPU as such to give a lot more power and thus clock speeds to a single core when it’s the only one being really taxed and then back off when the load is heavily multi-threadad. For example going from 4.5Ghz to 3.5Ghz on the performance cores. I don’t see why that can’t be a possibility.


Technically it’s totally feasible. If they get this kind of short-burst performance(this can’t be sustained on a phone) it will be a hell of a feat and anihilate Android.

 

[JPack]

You could also design the CPU as such to give a lot more power and thus clock speeds to a single core when it’s the only one being really taxed and then back off when the load is heavily multi-threadad. For example going from 4.5Ghz to 3.5Ghz on the performance cores. I don’t see why that can’t be a possibility.

It's not a possibility because of the fundamentals of silicon design.

You need a lot of voltage to push high clocks. If it was designed to run 4.5 GHz in the first place, then why is it running at 3.5 GHz? When you increase voltage, power consumption increases exponentially.

Even with perfect scaling, 4.5 GHz only gets you 30% more performance.

 

[Scarrus]

4.5 Ghz was just an example… what fundamentals of the chip? You have the schematics to the different parts in Apple chips?

Of course you can push a chip. How do you know how much voltage an M3/A17 can take on a single core?

 

[JPack]

4.5 Ghz was just an example… what fundamentals of the chip? You have the schematics to the different parts in Apple chips?

Of course you can push a chip. How do you know how much voltage an M3/A17 can take on a single core?

A16 runs at 3.46 GHz. To get 60% increase in performance and perfect scaling, it would need to run at 5.5 GHz.

If you still don't understand the fundamentals, then use common sense. We saw about 50% increase in GB scores when Apple transitioned from Intel to M1. Similarly, a 60% increase represents around 3 generations of AS improvements, e.g. A12 to A15.

What do you think is more likely? Apple suddenly unlocked some low hanging, super performance after a dozen generations of Apple Silicon? Or the benchmarks are highly suspicious?

 

[Scarrus]

A16 runs at 3.46 GHz. To get 60% increase in performance and perfect scaling, it would need to run at 5.5 GHz.

If you still don't understand the fundamentals, then use common sense. We saw about 50% increase in GB scores when Apple transitioned from Intel to M1. Similarly, a 60% increase represents around 3 generations of AS improvements, e.g. A12 to A15.

What do you think is more likely? Apple suddenly unlocked some low hanging, super performance after a dozen generations of Apple Silicon? Or the benchmarks are highly suspicious?

I think I understand the fundamentals better than you. You can’t simply compare the Ghz of one Cpu to another when you have a die shrink and a different CPU generation.

 

[oofio2461]

That means M3 will have a better upgrade, and with 10/12 core CPU, it could beat the 13900k.

 

[Sydde]

The multicore score is weird - you wouldn't expect scaling to be so bad

A17, destined for phones, would almost certainly be 2P/4E, which would make the MC score reasonable. SC is always a P score, so, if your SoC is 2/4, it is simply not going to show linear scaling.

 

[Sydde]

A16 runs at 3.46 GHz. To get 60% increase in performance and perfect scaling, it would need to run at 5.5 GHz.

But, unfortunately, it does not work that way. Your entire system has to be able to support the 5.5GHz processor. The way Apple cores are designed, they would rather quickly starve and stall at that kind of speed because memory just would not be able to keep it fed. If it has a plaid-size L3, maybe, but you are still probably better off running steady than in fits and stalls.

 

[Boil]

That means M3 will have a better upgrade, and with 10/12 core CPU, it could beat the 13900k.

I am hoping we see a 16-core CPU (12P/4E) in the M3 Max, thus giving us a 64-core CPU (48P/16E) in the M3 Extreme, and thereby maxing out the CPU core count that macOS can address...?

 

[ian87w]

Leaker Max Tech reportedly got the Geekbench 6 scores for Apple's upcoming A17 Bionic chip coming to the iPhone 15, and if they're to be believed, will give Android phone fans something to grind their teeth over. The single-core A17 Bionic running score hit an astounding 3986 points, and the multi-core score was 8841 points. Compared to the iPhone 14's A16 Bionic chip, the power increase for the A17 Bionic clocks in at about a 60% increase on single core and 43% on multi-core. The current MacBook with the M2 chip has a single-core score of 2604 and a multi-core score of 9751 on Geekbench 6. In other words, the M2 CPU has been beaten by the single-core running score, with a very close multi-core score … for an iPhone!

In contrast, the Snapdragon 8 Gen 2 is the industry’s top flagship chip right now for Android. This chip scored 1953 points in single-core performance and 5449 points in multi-core performance on Geekbench 6. Apple's A17’s multi-core score testing is 1.6 times higher than that of the Android flagship. Its single-core running score is nearly twice as high.

Leaked 3nm A17 Bionic Geekbench 6 Scores show a Major Power Leap Forward coming to the iPhone 15 Pro this Fall

During Apple's last Financial Conference Call, Apple's CEO Tim Cook stated: "I think people are willing to really stretch to get the best they can afford in that category." He further noted that the iPhone had become "integral" to life. Rumors surrounding this statement has given rise to a...

www.patentlyapple.com

The trend is not unexpected. The M2 uses the A15 cores, so single core wise obviously the A17 should be faster with newer core and smaller node (can ramp up the clock speed more).

And the multi core improvement should not be a surprise either as the rumor is A17 will have 4 performance cores instead of the traditional 2.

So regardless of the numbers, it is completely expected that the A17 will have higher single core than even the M2, and higher multicore than A16 due to more performance cores.

I really should learn stating obvious things as if it's news, and make money from it.

 

[Hastings101]

I hope this will drastically improve battery life (who needs more speed?).

No joke. When has an iPhone since at least the 7 series (if not earlier) felt sluggish? Very cool rumour if true, but not exactly a big deal except for it making its way into the Mac line in a couple of years as the M3/4

 

[JPack]

But, unfortunately, it does not work that way. Your entire system has to be able to support the 5.5GHz processor. The way Apple cores are designed, they would rather quickly starve and stall at that kind of speed because memory just would not be able to keep it fed. If it has a plaid-size L3, maybe, but you are still probably better off running steady than in fits and stalls.

Yes, which is why I said "perfect scaling." It would have to include the appropriate decrease in latency, increase in front/back-end resources, memory bandwidth, etc.

My example was to explain to the poster why his idea of turboing a single core to some astronomical GHz is so ridiculous.

 

[apparatchik]

No joke. When has an iPhone since at least the 7 series (if not earlier) felt sluggish? Very cool rumour if true, but not exactly a big deal except for it making its way into the Mac line in a couple of years as the M3/4

I agree in general with one exception: gaming; iPhone & iPads are the largest gaming platform out there with around 1 billion active devices (yes, not all are gamers) but in comparison the Nintendo Switch has shipped a tenth of that (around 130 million), I think it's fair to say more than a tenth of iPhone / iPad users game on their devices, making them larger than the Switch. As Apple Silicon keeps improving the better performance keeps them getting closer to current-gen console quality graphics.

 

[unchecked]

Interesting. What is your source? I have done some searching, of course it cannot be verified yet, but that does not mean it's fake. I don't see anyone saying that?

Because unless and until it's verified and/or official, it's always fake until it's not.

Our default position on anything we hear from hearsay shouldn't be to believe that it's true, but to always default to think that it's false.

 

[MayaUser]

No joke. When has an iPhone since at least the 7 series (if not earlier) felt sluggish? Very cool rumour if true, but not exactly a big deal except for it making its way into the Mac line in a couple of years as the M3/4

trust me, you are just used to it...put an iphone X next to an iphone 14 pro on the latest OS side by side and you will see the difference in everything
The battery the performance the display are improved and you can feel it very easily
Like on cars, when you step from a Renault into an S class you feel the difference but the bigger difference felt is when you come out from an S class that you used it for 1 year and step into the Renault...in that case you feel the real difference...stepping backwards

 

[leman]

https://twitter.com/i/web/status/1635506320479784961

Going to point out that this 'leak' from a more trusted source is much more realistic of a year-over-year improvement.

This is much more realistic and if true, would cement the insane lead Apple has over the competition in both mobile and desktop.

 

[Confused-User]

A17, destined for phones, would almost certainly be 2P/4E, which would make the MC score reasonable. SC is always a P score, so, if your SoC is 2/4, it is simply not going to show linear scaling.

Obviously. I took that into account. The numbers are still way off for a 2P4E chip. Consider the A16's numbers. If the leak isn't entirely bogus, something odd is happening - unusually poor scaling, or throttling - as I wrote in my first post.

 

[Confused-User]

4.5 Ghz was just an example… what fundamentals of the chip? You have the schematics to the different parts in Apple chips?

Of course you can push a chip. How do you know how much voltage an M3/A17 can take on a single core?

It's not "of course". There are a variety of limits to how far you can push a chip, some somewhat finessable and some an absolute limit. Pipeline length, clock distribution, process characteristics (leakage, etc.), more ... it's a long list. You can't just say "oh, let's turn up the voltage". It doesn't work that way - or rather, not beyond a certain point, which depends on the specific chip.

I don't know if anyone has tried to overclock an M1 or M2, for example, but I'd bet real money that you can't push them over 4GHz. One of the big questions about the A17/M3 core, for me, is "Is Apple increasing clock headroom?". We don't know yet.

 

[Confused-User]

But, unfortunately, it does not work that way. Your entire system has to be able to support the 5.5GHz processor. The way Apple cores are designed, they would rather quickly starve and stall at that kind of speed because memory just would not be able to keep it fed. If it has a plaid-size L3, maybe, but you are still probably better off running steady than in fits and stalls.

"Plaid-sized L3"?

It is *exceedingly* unlikely that Apple will push any chip this year to 5.5Ghz, but if they could, memory bandwidth starvation would be unlikely to be a major problem. It would definitely have an impact, but not a big one. Apple already has very large caches and very wide paths to memory. They'd make up some of the deficit by clocking RAM higher (LPDDR5X?), and the rest wouldn't have a major impact. If you look at the impact of slower RAM in typical PCs, you'd find that overall, performance doesn't suffer all that much. Sure, there will be particular codes that do poorly - anything that's extremely memory-access-intensive, that doesn't cache well - but that's a pretty small portion of all workloads.

BT, avoiding "fits and stalls" is not necessarily a good idea. Race-to-idle is all about the opposite. Which policy is better depends on the system, CPU, and load.

 

[Confused-User]

The trend is not unexpected. The M2 uses the A15 cores, so single core wise obviously the A17 should be faster with newer core and smaller node (can ramp up the clock speed more).

And the multi core improvement should not be a surprise either as the rumor is A17 will have 4 performance cores instead of the traditional 2.

So regardless of the numbers, it is completely expected that the A17 will have higher single core than even the M2, and higher multicore than A16 due to more performance cores.

I really should learn stating obvious things as if it's news, and make money from it.

Your math is wrong, or you didn't look hard enough. Even if the A17 is 2P4E, the rumored multicore number is too low. For a 4E4P system it would be *insanely* bad. But I haven't heard anything about the A17 being 4P4E, and I would be truly astonished if Apple did that. What would be the point?

 

[floral]

Leaker Max Tech reportedly got the Geekbench 6 scores for Apple's upcoming A17 Bionic chip coming to the iPhone 15, and if they're to be believed, will give Android phone fans something to grind their teeth over. The single-core A17 Bionic running score hit an astounding 3986 points, and the multi-core score was 8841 points. Compared to the iPhone 14's A16 Bionic chip, the power increase for the A17 Bionic clocks in at about a 60% increase on single core and 43% on multi-core. The current MacBook with the M2 chip has a single-core score of 2604 and a multi-core score of 9751 on Geekbench 6. In other words, the M2 CPU has been beaten by the single-core running score, with a very close multi-core score … for an iPhone!

In contrast, the Snapdragon 8 Gen 2 is the industry’s top flagship chip right now for Android. This chip scored 1953 points in single-core performance and 5449 points in multi-core performance on Geekbench 6. Apple's A17’s multi-core score testing is 1.6 times higher than that of the Android flagship. Its single-core running score is nearly twice as high.

Leaked 3nm A17 Bionic Geekbench 6 Scores show a Major Power Leap Forward coming to the iPhone 15 Pro this Fall

During Apple's last Financial Conference Call, Apple's CEO Tim Cook stated: "I think people are willing to really stretch to get the best they can afford in that category." He further noted that the iPhone had become "integral" to life. Rumors surrounding this statement has given rise to a...

www.patentlyapple.com

"Leaker Max Tech..."
Ah, sorry. Forgot to put my 'suspicious' cap on.

 

[Confused-User]

Somewhat embarrassingly, I screwed up my math, because I was... I dunno, half asleep? It was dumb. :-(

How do you get 60% improvement? Not so hard. You get perhaps 15% from going to N3. You get the rest by redesigning the core to run at faster clocks (which the M1 and M2 almost certainly can't), along with moderate IPC improvements. *IF* the number is real - which it may not be - then they might be getting anywhere from 5-20% from IPC gains, and then the rest would come from boosting clocks to around 4.5GHz, give or take, depending on the IPC gains.

So the problem is that the ~15% performance gain you expect from N3 is iso*power*, not iso*clock*. In effect I double-counted those gains.

To hit 60%, maybe they get 20% IPC bump due to two years work (not one) and 70% more area to build wider cores. That leaves 40% from clocks, so they'd need to be at 4.9GHz give or take - not 4.5GHz. That's feasible in a desktop, not even close for a phone (or an ipad, for that matter). In fact it would be more than feasible on N5, if they did a design built for that speed, which the M1/M2 are certainly not. It should be easy on N3B... if they actually want to do that. We'll see.

 

[leman]

The multicore score is weird - you wouldn't expect scaling to be so bad

Keep in mind this is GB6, not GB5, so multi-core results work a bit differently. Still, that MC score would only make sense if either a) there is a very high single-core turbo (as you write) or b) the E-cores are not any faster then today. I find neither very plausible.

The other claim (3000/7800) makes much more sense.