Apple A12 & A12X *** Now A12Z as well *** Now in a Mac mini

[Roland00Address]

It is an improvement over the old iPad Pro, due to more DRAM, more flash, better camera and slightly faster graphics. They're just giving a mid life kicker to this generation iPad Pro to make it a better value - you not liking it compared to the regular iPad is irrelevant to Apple. You obviously aren't their target market for the "Pro" version. For buyers of the Pro these are welcome improvements, or they can wait another 9-12 months for a bigger bang.

Human folly and all that, but I am allowed to want more in an ethereal sense. I am disappointed for I feel this is like the old iphone generation thing where

Main Number change, you get new features
S Number change, you actually get better specs with cpu, gpu, and ram.

[Note this pattern no longer applies but it was the old pattern for example iPhone 4 vs iPhone 4s, iPhone 5 vs iPhone 5s and so on.]

-----

It was 18 months ago (500 days) when the last iPad Pro upgrade compared to now. I would feel a little better if this was the change 12 months in not 18. It would be great if we get new silicon in 6 months from now. I just wish the iPad Pros were on a more "consistent / quicker cadence" instead of being updated 2015 / 2017 / 2018 / 2020.

 

[Richie Rich]

Why would they release an A12 variant?

The answer could be because there is no A14X suitable for iPad Pro. This could have several reasons:

• - there is no A14X at all - Apple skipped last ARMv8.3 chip in favor to A15X with new ARMv9 and SVE vectors
• - there is no A14X at all - A12Z is still beast, now cheaper and A14 provides smaller performance jump to be worth it
• - A14X is delayed for next year
• - A14X has bigger configuration (6+6 cores) aiming for 10W TDP for laptops

A15 design should be finished soon so Apple has performance numbers already. If A15 is bigger performance step than A14 then it makes sense skipping A14X entirely (A12X/Z is still beast, it beats Ice Lake at half TDP). Also not bad to have new ARMv9 ISA for both SoCs at the same time.

 

[Doug S]

The reason there is no A14X is because TSMC isn't ready for mass production at 5nm yet. Skipping A14X because A15 will be better is stupid, the next one will always be better so they'd wait forever with that plan.

 

[Doug S]

So just like I said, the A12Z is an A12X with the 8th GPU cluster enabled...

Apple’s A12Z Processor Confirmed to Reuse A12X Silicon

www.anandtech.com

 

[teejee]

So just like I said, the A12Z is an A12X with the 8th GPU cluster enabled...

Apple’s A12Z Processor Confirmed to Reuse A12X Silicon

www.anandtech.com

yes, could be higher frequency as well though. No info in the article contradicts that at least.

 

[Richie Rich]

Yes, with higher yields and better silicon Apple could change freq/voltage curve to get lower consumption. So due to thermal limitation new iPads can run more often at higher clocks.

A12Z is a monster even today. It beats brand new 6-core Renoir 4600U in every way (ST and MT). Despite having only 4 big cores and consuming half of the power in tablet. Apple has about 4 years development advantage over AMD so they are not in hurry.

IMHO next A-X SoC will be based on new uarch with 2048-bit SVE2 support. They can use it in MacBooks as well as Intel replacement and save a lot of money. ARM Holding plans new SVE2 cores for 2022 products and Apple could be one year ahead, so likely 2021.Rumors of replacing Intel in 2021 are no coincidence, just another piece of the puzzle fits into a larger picture.

 

[DrMrLordX]

A12Z is a monster even today. It beats brand new 6-core Renoir 4600U in every way (ST and MT).

Got any Pov-Ray results for that? How about, I don't know, 7-zip?

Or hell, how about Blender? Cmon, it's open source. Someone can compile it for iOS. Can't they?

 

[amrnuke]

A12Z is a monster even today. It beats brand new 6-core Renoir 4600U in every way (ST and MT). Despite having only 4 big cores and consuming half of the power in tablet. Apple has about 4 years development advantage over AMD so they are not in hurry.

Correction:
The 8 core (4 fast, 4 efficient) A12Z beats the 6 core Renoir 4600U in Geekbench 5. GB5 LOVES cores, doesn't matter if it's fast or slow. See my reply to you in the Quo vadis Apple Macs thread. GB5 is not a real world benchmark, though it claims to be.

Can you verify the power consumption of both chips during the GB5 run, or in any way prove that it consumes half the power during those runs?

Apple has a 4 year development advantage. In mobile. Where AMD has no chips.

 

[Hitman928]

Got any Pov-Ray results for that? How about, I don't know, 7-zip?

Or hell, how about Blender? Cmon, it's open source. Someone can compile it for iOS. Can't they?

I don't have Pov-ray, but there is a simple ray tracing benchmark in GB5. It seems to be a very simple scene/benchmark more akin to C-ray than Pov-ray but here are some multi-threaded results:

A12X iOS: 6.43 Mpixels / s
4700u Windows: 10 Mpixels / s

LENOVO INVALID - Geekbench

Benchmark results for a LENOVO INVALID with an AMD Ryzen 7 4700U processor.

browser.geekbench.com

iPad Pro (12.9-inch 3rd Generation) - Geekbench

Benchmark results for an iPad Pro (12.9-inch 3rd Generation) with an Apple A12X Bionic processor.

browser.geekbench.com

 

[Richie Rich]

.. here are some multi-threaded results:
A12X iOS: 6.43 Mpixels / s
4700u Windows: 10 Mpixels / s

Not bad for two years old CPU with only 4 big cores vs. 8-core/16 thread Zen2.... using twice as much power and 60% more die size. Ray tracing benefits from SMT a lot.

However in ST an old A12X beats Ryzen pretty bad:
A12X: ..... 1.30 Mpx/s
4700: ..... 1.25 Mpx/s

We can assume that A12X with doubled cores would beat Renoir pretty easy in RT. However not in emulation. Blender exists in ARM version and it runs on RPi4.

 

[Hitman928]

Not bad for two years old CPU with only 4 big cores vs. 8-core/16 thread Zen2.... using twice as much power and 60% more die size. Ray tracing benefits from SMT a lot.

However in ST an old A12X beats Ryzen pretty bad:
A12X: ..... 1.30 Mpx/s
4700: ..... 1.25 Mpx/s

We can assume that A12X with doubled cores would beat Renoir pretty easy in RT. However not in emulation. Blender exists in ARM version and it runs on RPi4.

4700u doesn't have SMT. So it's 8 cores against 8 cores but 4 of the 8 for A12X are small cores obviously. The single core results are definitely impressive though running the 4700u on Linux would probably result in a decent bump in performance on both single and multi-threaded results and would be closer to running on iOS than Windows.

For die size, according to Wikichip Renoir is 28% larger than A12X but Renoir also has way more IO circuitry to support PCIe and such which the A12X doesn't have. I believe Renoir's GPU is larger as well though I can't confirm that. In terms of cores + cache, they're probably pretty comparable in size.

The rest is nice speculation, but that's all it is. I'm not aware of any power consumption numbers for A12X. Personally, I wouldn't be interested in a desktop Apple CPU from a user perspective as I can't stand the apple environment, but it would be interesting to see from a technical perspective.

 

[Glo.]

Not bad for two years old CPU with only 4 big cores vs. 8-core/16 thread Zen2.... using twice as much power and 60% more die size. Ray tracing benefits from SMT a lot.

However in ST an old A12X beats Ryzen pretty bad:
A12X: ..... 1.30 Mpx/s
4700: ..... 1.25 Mpx/s

We can assume that A12X with doubled cores would beat Renoir pretty easy in RT. However not in emulation. Blender exists in ARM version and it runs on RPi4.

No, we cannot assume this, because 4700U is 8C/8T design, that is relatively low clocked, and under load multi core boosts around the same as A12X boosts: 2.6-2.7 GHz. So you brag about a CPU beating other CPU in Single threaded by 0.05 Mpx/s, forgetting about the difference iOS is yielding for performance.

If anything we can assume is that in similar software platform, x86 would still beat Best in the world ARM CPU, that is Apple made.

 

[amrnuke]

Not bad for two years old CPU

1) The A12X has only been out since early November 2018, it is not even 1.5 years old.
2) Zen2 has been available to select customers for about a year.
They were officially released 8 months apart.

with only 4 big cores vs. 8-core/16 thread Zen2.... using twice as much power and 60% more die size. Ray tracing benefits from SMT a lot.

A12X has 8 cores, 4 fast and 4 efficient. That plays into 1) why it loses and 2) uses less power and 3) has a smaller die size.
Neat fact that ray tracing likes SMT. So does a lot of real work.
The 4700U doesn't have SMT though. And it still smokes the A12X in MT tasks.

However in ST an old A12X beats Ryzen pretty bad:
A12X: ..... 1.30 Mpx/s
4700: ..... 1.25 Mpx/s

<5% difference.

We can assume that A12X with doubled cores would beat Renoir pretty easy in RT. However not in emulation. Blender exists in ARM version and it runs on RPi4.

We can assume no such thing. As you can see with these two chips compared 8 thread vs 8 thread , and see how much A12X chokes on multi-thread-heavy tasks. Either Tempest is really bad, or A12X just doesn't multi-thread well. Probably a little of both.

 

[Glo.]

We can assume no such thing. As you can see with these two chips compared 8 thread vs 8 thread , and see how much A12X chokes on multi-thread-heavy tasks. Either Tempest is really bad, or A12X just doesn't multi-thread well. Probably a little of both.

Check out the rest of ARM CPU reviews. We see similar pattern in them, that they choke in Multithreaded software. Especially the newest ones.

Why? Because ARM in general is designed for Single Threaded, lightly threaded workloads, whereas x86 is to fit all of the hands: Single threaded and Multithreaded. Both lightly and heavily threaded.

 

[Hitman928]

We can assume no such thing. As you can see with these two chips compared 8 thread vs 8 thread , and see how much A12X chokes on multi-thread-heavy tasks. Either Tempest is really bad, or A12X just doesn't multi-thread well. Probably a little of both.

A12X doesn't scale well because once you get above 4 threads the extra threads are handled by smaller, less performant (but higher efficiency) cores. So you would only expect good scaling up to 4T and then poor scaling beyond. If Apple were to put 8 big cores on the chip, it would probably scale well but they'd also probably have to make some changes to the memory/cache hierarchy to support scaling with the additional big cores. Power consumption and die size would increase significantly.

 

[DrMrLordX]

there is a simple ray tracing benchmark in GB5.

Wish we knew more about it though. It's much nicer to have a standalone benchmark where you can look at the code or at least do some instruction profiling to understand what it's doing.

 

[Hitman928]

Wish we knew more about it though. It's much nicer to have a standalone benchmark where you can look at the code or at least do some instruction profiling to understand what it's doing.

Geekbench published a document describing each of their sub-tests but it's not very detailed.

The Ray Tracing workload uses a k-d tree, a space-partitioning data structure, to accelerate the ray intersection calculations. The Ray Tracing workload operates on a scene with 3,608 textured triangles.The rendered image is 768 pixels by 768 pixels.

https://www.geekbench.com/doc/geekbench5-cpu-workloads.pdf

 

[amrnuke]

A12X doesn't scale well because once you get above 4 threads the extra threads are handled by smaller, less performant (but higher efficiency) cores. So you would only expect good scaling up to 4T and then poor scaling beyond. If Apple were to put 8 big cores on the chip, it would probably scale well but they'd also probably have to make some changes to the memory/cache hierarchy to support scaling with the additional big cores. Power consumption and die size would increase significantly.

There is a product out there that has basically the same single-threaded scores on GB5 overall as well as the RT test. The chip scales perfectly in the ray tracing test, and uses more power and has a larger die size than the A12X.

Spoiler

It's called a Ryzen 4700U.

 

[amrnuke]

Also very interesting to see Apple plop a full A13 into the iPhone SE for $399. If the screen isn't garbage and the battery life is decent (perhaps they downclock the A13?) that's a solid deal.

An A12 would have worked very well here too, in my iPad Air, the A12 is stellar, and I'm guessing it would have been cheaper, but this may give some clue as to just how cheap it is for Apple to make these <100mm2 A13s, that they can dump them into a $399 phone and still ostensibly turn a profit.

 

[Thala]

Got any Pov-Ray results for that? How about, I don't know, 7-zip?

Or hell, how about Blender? Cmon, it's open source. Someone can compile it for iOS. Can't they?

You cannot just compile it for iOS. The app needs to target the iOS APIs, which neither Pov-Ray or 7-zip is doing. You can however compile both apps for ARMv8 - which would result in a Win32 ARM64 application. In fact i did compile both apps for ARM64 - but i cannot run them under iOS apparently but rather on the Windows on ARM machines like the Surface Pro X.
In general if the app is available for a certain API, say Win32, MacOS, Linux - it is trivial to port them to ARM - but it stays a Win32, Linux or whatever app - its just the same app compiled for a different CPU architecture.

 

[DrMrLordX]

You cannot just compile it for iOS. The app needs to target the iOS APIs, which neither Pov-Ray or 7-zip is doing.

Oh I know. I was sort-of making a point. And that is that a certain user around here keeps gushing about how much better <insertApplechiphere> is than something from Intel or AMD, and uses something like SPEC or Geekbench to "prove it".

Blender would require a recode for the UI. The underlying rendering engine wouldn't require that, per se. A command-line version of Blender Benchmark might work.

 

[Richie Rich]

Oh I know. I was sort-of making a point. And that is that a certain user around here keeps gushing about how much better <insertApplechiphere> is than something from Intel or AMD, and uses something like SPEC or Geekbench to "prove it".

Blender would require a recode for the UI. The underlying rendering engine wouldn't require that, per se. A command-line version of Blender Benchmark might work.

You don't need GUI/API to run Blender RT engine. You can start render from console because it supports multi-cluster render. There are a lot of Raspberry Pi "super computer" clusters running Blender on YouTube. I've seen also list of Android smartphones with Blender times. However I'm not sure if iOS can run Linux without root same way as Android can.

 

[DrMrLordX]

You don't need GUI/API to run Blender RT engine. You can start render from console because it supports multi-cluster render. There are a lot of Raspberry Pi "super computer" clusters running Blender on YouTube. I've seen also list of Android smartphones with Blender times. However I'm not sure if iOS can run Linux without root same way as Android can.

Port it to iOS. Go on, do it. If it's headless rander from command line, you shouldn't have any API conflicts. You can't put it on the app store, but you can build an "executable" that'll last for 7 days.

And no, you can't run Linux on Apple phones or tablets. Haven't been able to do that in a looooooooooooooooong time.

 

[Eug]

A12Z is used for Apple's new ARM macOS dev platform.

Logic Pro, Final Cut, Photoshop, and Office are already running on it with fat "Universal 2" binaries.

Rosetta does dynamic translation to ARM.

 

[tamz_msc]

Shadow of the Tomb Raider running on that thing was mighty impressive.