Apple A14 and A14X SoCs

I feel like a reasonable guess being there are at least two lines of SoCs above the “X” line designed for fanless iPad Pro. My expectation is one mobile all-in-one SoC for MacBook Pro 16” and below; and a CPU-only SoC for Mac Pro and higher-end desktop iMac (paired with AMD GPUs initially).

Maybe we will see Apple processors shipping with varying level of disabled cores for the first time. Then we might see perhaps advanced packaging and chiplets to scale from the desktop to workstation Macs, which is likely from a cost-benefit & volume perspective.

 

The only problem with that is that absolutely no sessions from this years WWDC covered anything else other than Apple's own GPU. Not even a mention of discrete GPUs from any vendor.

Granted the Mac Pro could be the last to switch meaning we will get another WWDC before it releases.

 

I'm guessing they might bin the CPUs and offer say 8 core and 6 core variants, with slightly cut down GPUs (also for yield reasons). Tend to agree that upgrade options would be fewer though as RAM and even NAND may be soldered.

HBM2 is probably unlikely and LPDDR4x/LPDDR5 should offer enough bandwidth for lower cost. The fact they went for LP4X for the iphone 12 certainly indicates that they're bandwidth efficient. A 128 bit LP5 interface should be plenty for the Macs.

I'm not entirely sure we'll see discrete GPUs with the Apple SoCs initially, partially due to power reasons and partially because of driver maturity. Would AMD invest significantly into drivers for ARM Macs so early?

 

If Apple intends to continue its AMD partnership, I don’t see why not. It could also very well be a semi-custom arrangement which AMD has NRE incentives to port the drivers over, given the existence of custom parts like the Apple exclusive Navi 12.

Though of course, if we speculate that Apple really wants to tightly control and unify the entire GPU HW-SW feature roadmap across all its products, AMD GPUs in ARM Mac probably would never happen, and eGPUs would be effectively deprecated.

 

Both DRAM and SSD have been soldered in the last few generations. So the current offerings already have reflected this, and I see no reason to expect it to change.

They typically only stock baseline models and top-of-the-line, while the rest are BTO.

 

The first 3 Apple Silicon machines
A14 deep dive

Quick summary: 4+4, 8 or 16 GB RAM, MBA, MBP, MacMini, excellent CPU performance.

I ordered a MBP with 16 GB. That's the first ARM-based laptop that has enough performance and dev tools to get me interested

 

Are such admissions allowed without being run out of the forums covered in tar and feathers?
I don’t think Apples sales volumes will change drastically, but I do believe that people will look at these small, quiet and performant system, that adds Very Long battery life to their list of comforts, and look at their own PCs and find them lacking. The relatively high power draw of x86 solutions simply make them....less nice. They actually draw more now (a lot!) than they did 30 years ago, which is weirdly out of sync with the rest of our machines and devices, as if PC development missed a vital evolutionary branching somewhere along the way.

 

Sadly there are too many memory hogs in my day-to-day workflow... Many are Apple’s own dev tools, and Electron/PWA apps.

My 6-core 2018 Mac Mini had been choking on 16GB, so I had to abandon it for clamshell in the company issued top-spec MBP 16. (Didn’t bother to open the mini up)

Guess I need to wait patiently for the rumoured Mac Pro mini...

 

The interesting thing with this M1/A14X is that it is able to match fastest x86 CPUs on high performance nodes with a high density / low power node, that with a much lower clock. Went for the Mini, partly for nostalgic reasons, as a successor of my former Acorn ARM desktop.
BTW I was surprised how smooth buying from Apple went, everything in stock and they are able to tell upfront when delivering. You pay through the nose but they got their supply chain in order.

 

I ordered the MBP on Nov 12th and I got a 3-4 week delay (week starting Dec 5th). On Nov 10th it was a one week delay.

 

Curious how they will scale up the M1 to the more powerful Mac range.
I bet they will stick with UMA with CPU/GPU etc integrated in a single SoC.
Doubling the CPU and GPU cores, and doubling the memory bus to 256-bit would still fit in a die smaller as 200 mm2.
Switching to LPDDR5 with 16 GB per stack would allow up to 64 GB and 200 GB/s bandwidth.
All consuming ~30 Watt, pretty insane.

 

The current M1 is probably the same as the "A14x" referred to in the thread title, fundamentally designed for a less than 10W power envelope. The die area/power envelope/memory subsystem you outline should allow a tripling of big cores and GPU cores, particularly as not all aspects of the chip would need to scale up, and LPDDR5 draws less power/bit/s.
Just how Apple will choose to scale their SoCs to higher performance levels is very, very interesting.

 

There would be zero reason to scale past 4 high efficiency cores just yet. The M2 (or whatever) for the higher end MacBook Pro 16” could easily get away with 8 high performance cores and 16 core GPU.

I have a feeling that many things can be done with the 16-core Neural Engine that were hitting the GPU before, like photo and video work.

That same SoC would go in a space grey Mac mini.

Both will have more Thunderbolt ports.

 

Tripling of cores might be feasible indeed, doubling already would be more than enough for the power laptops and iMac non pro range.
The M1 will end up more like 15Watt I think, based on graphs that were shown, 10Watt could be baseline after throttling down for the Air as it has no active cooling.
Most heavy load likely games with heavy CPU/GPU load or 3D CAD, not sure if this is a big market for Apple actually.

 

Assuming apple is heavily invested in accelerators how much do they even need cpu to begin with? Video editing has special accelerator in mac pro world, all arm based macs have probably decent gpu and decent neural network accelerator. Unified memory approach is great for accelerator heavy SOC. Unified memory/soc makes it possible to share work flexibly between units while avoiding overhead as resources like cache, memory bus and memory are shared. Apple can move both hw and sw to use new SOC and accelerators while using cpu only for things cpu is best at.

Apple isn't in business of selling faster cpu/gpu. Rather apple wants to sell better user experience and better SOC. World is different if you can integrate everything you need into a single SOC+supporting sw instead of trying to sell yearly cpu/gpu upgrade.

 

Don't forget about games or software, that use raw CPU power.

Handbrake 1.4 beta Universal Binary, https://github.com/HandBrake/HandBrake/releases/tag/1.4.0-beta.1

 

I'm not really holding my breath for triple a games on arm+apple. photoshop/video editing/... might just get fancy accelerator or be optimized to use gpu+neural net engine.

If apple wants they have choices they can make. For example this could be part of the mac pro soc instead of it being separate card.

https://appleinsider.com/articles/1...t-the-afterburner-accelerator-for-the-mac-pro

 

Well this was just posted: World of Warcraft on Apple Silicon.

But yeah not sure others will follow.

 

It's not simple:

I don't think, that Handbrake users care about ProRes and Afterburner.
It looks like Handbrake supports Apple VideoToolbox, but a "true Blu-ray ripper" would use x264 or x265 for the quality control and possible best quality.

 

TensorFlow Blog has a post with some benchmark results from the new Apple optimized Accelerated TensorFlow 2.4 on M1:

https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performance-on-mac.html

Since it's several times faster than the 13" Intel Macbook Pro I guess it utilizes the "Neural Engines" on M1.