Apple's Tim Millet discusses A14 architecture, future chip designs
In an interview published on the eve of Apple's "iPhone 12" launch event, Apple VP of platform architecture Tim Millet has explained some of the work that went into the A14 Bionic system-on-chip, and what it means for the future of Apple's chip designs.
Revealed in the iPad Air launch during Apple's first special event, the A14 is widely anticipated to make an appearance during Tuesday's "Hi, Speed" event, where Apple is expected to unveil its 2020 iPhone lineup. Millet offered more details about the A14's design and creation.
Made using a 5-nanometer process, the A14 packs in 11.8 billion transistors onto the chip, up from the 8.5 billion of the A13, with the changes enabling Apple to be more precise in how it uses the chip to shape the user's experience.
"One of the ways chip architects think about features is not necessarily directly mapping [transistors] to a user feature in the product so much as enabling the underlying technology, like software in the graphics stack to be able to leverage a new capability in the GPU, said Millet to Engadget. "That will inevitably come as a visual feature in a game, or in a snappy transition in the user interface."
The expansion of the Neural Engine to 16 cores instead of 8 cores brought up the question of why Apple elected to devote transistors there and not budgeted for more GPU or CPU performance, which Millet suggests is down to how Apple views the feature.
"We saw the opportunity to do things that would have been impossible to do with a conventional CPU set," he explained. "You could in theory do many of the things the Neural Engine does on a GPU, but you can't do it inside of a tight, thermally-constrained enclosure."
Apple also had to balance power and energy efficiency, on which Millet says Apple erred towards efficiency, since it "applies to every product that we build." By working on making chips power efficient as a focus, it doesn't need to work on making the same chips efficient in different ways for an iPhone over an iPad.
Turning to its future of chip designs, Apple is keen to build and reuse its ideas. "Ultimately, we want to make sure that when we build a CPU for one generation, we're not building it necessarily for one," advised Millet.
An example of this is in the X and Z versions of the same chips that are used in the iPad Pro lines, as they are fundamentally tweaked and added-to variants of existing chip designs. Apple also builds to make the chip architecture "parameterized," making it potentially scalable to other hardware, like an Apple Watch.
One potential upcoming example of this would be Macs with Apple Silicon, with the first public launch of a Mac using the chips expected before the end of 2020. Developer kits used a Mac mini equipped with an A-series chip, which suggests the same chip designs may be in use for that product line.
In another interview from October 5, Millet discussed the progression of machine learning, including the closeness of working with Apple's software team and the importance of Core ML.
Revealed in the iPad Air launch during Apple's first special event, the A14 is widely anticipated to make an appearance during Tuesday's "Hi, Speed" event, where Apple is expected to unveil its 2020 iPhone lineup. Millet offered more details about the A14's design and creation.
Made using a 5-nanometer process, the A14 packs in 11.8 billion transistors onto the chip, up from the 8.5 billion of the A13, with the changes enabling Apple to be more precise in how it uses the chip to shape the user's experience.
"One of the ways chip architects think about features is not necessarily directly mapping [transistors] to a user feature in the product so much as enabling the underlying technology, like software in the graphics stack to be able to leverage a new capability in the GPU, said Millet to Engadget. "That will inevitably come as a visual feature in a game, or in a snappy transition in the user interface."
The expansion of the Neural Engine to 16 cores instead of 8 cores brought up the question of why Apple elected to devote transistors there and not budgeted for more GPU or CPU performance, which Millet suggests is down to how Apple views the feature.
"We saw the opportunity to do things that would have been impossible to do with a conventional CPU set," he explained. "You could in theory do many of the things the Neural Engine does on a GPU, but you can't do it inside of a tight, thermally-constrained enclosure."
Apple also had to balance power and energy efficiency, on which Millet says Apple erred towards efficiency, since it "applies to every product that we build." By working on making chips power efficient as a focus, it doesn't need to work on making the same chips efficient in different ways for an iPhone over an iPad.
Turning to its future of chip designs, Apple is keen to build and reuse its ideas. "Ultimately, we want to make sure that when we build a CPU for one generation, we're not building it necessarily for one," advised Millet.
An example of this is in the X and Z versions of the same chips that are used in the iPad Pro lines, as they are fundamentally tweaked and added-to variants of existing chip designs. Apple also builds to make the chip architecture "parameterized," making it potentially scalable to other hardware, like an Apple Watch.
One potential upcoming example of this would be Macs with Apple Silicon, with the first public launch of a Mac using the chips expected before the end of 2020. Developer kits used a Mac mini equipped with an A-series chip, which suggests the same chip designs may be in use for that product line.
In another interview from October 5, Millet discussed the progression of machine learning, including the closeness of working with Apple's software team and the importance of Core ML.
Comments
Speaking of which... Who else thinks that with these upcoming ASi Macs we’re going to lose the ability to customize for performance? Today I can buy an iMac with an i5, i7, or i9, but only because Apple can just buy those from Intel. Will Macs move to a more iOS device type of customization, where performance differences are marked by generation, not variant (expect of course for the AnX/AnZ)?
There are currently 4 Mac categories to design SoCs for: consumer desktop, consumer mobile, pro desktop, pro mobile. So that’s already at least four variants on the same generation SoC. Is Apple going to be able to make multiple SoCs for each category so that there’s a method of choosing more or less performance in CPU/GPU?
Or is it going to be... Here’s an iMac... choose your display size, storage and memory, that’s it!
Two things.
1. There are a lot of AI features on mobile and this will increase in the future. You offload those AI features to their own chips for the same reason that floating point math was offloaded at one time and graphics and security functions are offloaded now. Separating the specialized functions allows for botth the specialized and generalized functions to run better.
2. The other companies involved in AI - Microsoft, Amazon, Google - are using the cloud for AI. Apple lacks their cloud prowess and infrastructure so they are trying to compete using hardware. Apple is going to either aim to provide more AI functionality locally than the competition can provide using the cloud - which should be possible in theory so long as Apple can come up with applications for it - or provide similar AI functionality but with more "speed, security and privacy" than shipping AI queries and results back and forth over a 5G connection.
While it isn't this way on PCs - yet - go ahead and accept that going forward there are going to be 3 main processing units in mobile: CPU, GPU and AI.
Right now, only Apple knows how the 5nm process is going. It’s also possible that the yields are so good that they don’t have to do this.
They have a long way to go, especially in infrastructure.
I'm not sure why you think Apple is ahead with on device AI. I haven't seen anything that supports that. The opposite is possibly true.
Huawei has been using on device AI since 2017 and each new generation has just added more uses for it (battery management, voice and depth sensing biometrics and enhancement, image recognition, capture and stabilisation, translation etc.) They now use their own 3D Cube tech with MindSpore and the Ascend line of processors which are designed to scale from earbuds through phones right up to multimillion dollar clusters.
I was going to comment, but what's the point?
You'll just get schooled tomorrow when Apple announces the iPhone 12 and interleaves all of the new features with the ML that is integrated with the hardware and software.
Still,
You seem to have been on the losing side of your Huawei Telecom arguments, for quite awhile now, as more countries back away from Huawei, likely because of all of the Authoritarian shit that China has and is pulling.
https://www.reuters.com/article/us-orange-nokia-security-5g/huawei-ousted-from-heart-of-eu-as-nokia-wins-belgian-5g-contracts-idUSKBN26U0YY
The only big one left is Germany, and they appear to be putting human rights and security policies ahead of Trade priorities with China. Huawei could end up empty handed in Germany as well.
Lets just hope politicians don’t screw the product model.
With Apple able to transfer years of AI device OS components to macOS and add AI to ASi Macs from day one, I guess they’ll be way ahead of the pack in the PC market too.
Do you realize what the difference in responsiveness is between packaging up a request and shipping it off to the cloud and processing it internally using internal SoC data paths? Consider how an iPhone can give you a portrait preview with multiple planes of blur vs. Android's foreground sharp/backgroup blurry only in post.
Consider the I/O advantages of using internal graphics cores in the SoC vs. packaging them up and sending them to an external card via PCIe - and PCIe is vastly faster than the internet.
Internal SoC processing is so much more efficient than even using high speed data busses - you're not having to do things like package compression so you can fit enough throughput through even the enormous capacity of PCIe and are mere millimeters away from those graphics cores using SoC internal data paths and internal memory caches (which probably provide faster access than there is to physical RAM).
Another way to look at it is that there are a bunch of performance tiers because Intel charges differing amounts depending on the performance of the chip.
Apple doesn't use all core-i9s because they're expensive and produce a lot of heat and consume lots of power, so they use a bunch of processors at different tiers to adjust the balance between power, expense, and performance.
Maybe with Apple silicon all models will be top tier models in each SoC catagory - A14 for lowest level, maybe an A14X for iPad/midrange, and others for higher performance (the first Mac SoC is rumored to have 8 Firestorm high performance and 4 Icestorm high efficiency cores). I would expect an incremental cost for higher performance SoCs, both due to silicon real estate and higher cost per wafer because of lower quantity produced, but not like what Intel's charging.
Apple could defray the cost by selling CPUs, but having their own silicon is such a competitive advantage there's no way they're going to do it.
Having your own high performance silicon line that no competitor can buy and compete with you? Priceless.
Remember that one of Intel's problems is getting enough working chips out of a wafer.
Intel's 10th gen chips are 10 nm - except for the 10900s which are 14 nm. The 10900K is Intel's overclockable 10 core 20 thread chip, and though it has a nominal TDP of 125w in the real world it can consume over 300w. They've reached the limits of physics, and actually had to shave the top of the chip to get enough surface area to cool the chip, and even then it requires heat transfer material, heat pipes, radiators, and cooling fans to maintain operation near advertised clock speeds.
Compare that to the A12Z whose normal operating environment is passively cooled in a glass sandwich. In a Mac it would produce more heat, but not to the extent of competitors in the x86 line where discussion of models and performance always devolves into how good the cooling system is.
We'll see, but I wouldn't be surprised to learn that the highest performing Macs won't be coming out until A15 which is planned to be produced on TSMC's 3nm node - provided they don't run into undue quantum effects. Even Apple silicon will eventually run into Moore's Wall.