Solved! ARM Apple High-End CPU - Intel replacement

[Richie Rich]

There is a first rumor about Intel replacement in Apple products:

• ARM based high-end CPU
• 8 cores, no SMT
• IPC +30% over Cortex A77
• desktop performance (Core i7/Ryzen R7) with much lower power consumption
• introduction with new gen MacBook Air in mid 2020 (considering also MacBook PRO and iMac)
• massive AI accelerator

Source Coreteks:

 

[jpiniero]

What I meant was for the on premise data centers they are using a third party vendor for the hardware, like say HP. It may not even be running OSX, more like Linux.

 

[soresu]

Alternatively Apple could conclude that hard-core gamers are just not a desirable market (there's at least a segment that is barely socialized so always complaining, constantly trying to harass women, gays or any other victim they can find, always trying to find a way to avoid paying or to hack/ruin the experience for others), and as a business it's just much nicer serving the casual gamer market...

Don't make the media's mistake of conflating a noxious high pitched rabble with all hardcore gamers.

You're not really a hardcore gamer if you are more concerned with vitriol than actual gaming.

It's more like they are hardcore undesirables that also play games.

 

[name99]

What I meant was for the on premise data centers they are using a third party vendor for the hardware, like say HP. It may not even be running OSX, more like Linux.

That's probably true today. But everything takes time.
First you have to build the services (and remember Apple was what some would consider late to that party --- you can argue that this is because they had different standards wrt crypto, privacy, and the type of functionality offered, but they were late).
Then building data centers and migrating to running them takes time.
Then building up your own codebase (and exploring whether and when it makes sense to run on top of OSX rather than Linux) takes time.
Then moving to ARM takes time.
Then building your own HW takes time.

They're likely only halfway through this process.

But look at this more generically. EVERY substantial cloud vendor -- MS, Google, Amazon, Baidu -- has been exploring hardware options. This varies from using GPUs to Amazon's ARM cores to FPGAs to TPU and other AI accelerators. Why would Apple be the one large cloud company that doesn't investigate alternative, superior, hardware? And why wouldn't their exploration include not just these areas (accelerators, GPUs) but also better CPUs --- where better would mean anything from faster to cheaper to lower power to better integration with accelerators to better memory support to better security?

Apple has shown repeatedly that they're engaged in 5..10 yr projects that remain essentially secret (nothing but rumor and informed speculation) till the day of the announcement. The fact that we know nothing about Apple's data center operations today, or their plans for 2025, is par for the course. But there is a certain level of common sense.
I mean, come on, if you were ANY cloud provider and you looked at the endless stream of Spectres and Zombies and JCC-bugs and promises (next year, next year) about 10nm and EMIB, wouldn't you seriously investigate alternatives?

The other thing about Apple is that they are very disciplined about not doing too much too fast, and reusing what works. Meaning, I would guess, that
- TODAY they are experimenting with A11s, A12s, A13s on cloud workloads, to see what works and what doesn't. Are small cores useful or not? How about the onboard (ie small) GPU and inference NPU? Are there concerns with the caching system and the cost of inter-processor communication?
- TOMORROW they will ship desktop ARMs, which allows the cloud folks to start testing how well inter-chiplet communication works, and whether there are weaknesses in their IO system when pushed hard. All of which means that
- A GENERATION LATER they can rework this learning into better desktop cores, which can be reconfigured for cloud work, and serious mass scale deployment can begin.

 

[name99]

Don't make the media's mistake of conflating a noxious high pitched rabble with all hardcore gamers.

You're not really a hardcore gamer if you are more concerned with vitriol than actual gaming.

It's more like they are hardcore undesirables that also play games.

I expect this is probably true, that it's unfair to categorize this type by "gamer" as opposed to some other common bonding experience (maybe a particular TV show? maybe a love for guns?) I certainly feel your frustration around the simplistic claims of the media (I mean, try to even find statistics and you'll find a whole lot of generic claims about "games this" and "male gamers that" and "female gamers experience" but precious few actual numbers).

But companies have to operate in the world of media and stereotypes :-(
And until the stereotype settles onto something else, the situation is what it is, and companies can't be blamed for being shy. I do wonder if Google is going to, very soon, have to deal with headlines claiming "Racism, sexism, homophobia rife on Stadia chat sessions and Google is doing NOTHING"...
(Followed six months later by "Did you know that Google is spying on your Stadia chat sessions? How can they possibly justify this?"...)

 

[Gideon]

A new custom contender called Nuvia could be entering the ARM market, found this on Phoronix.

Link.

I'm surprised this got so little reception. Let me quote parts of the Anandtech's article on NUVIA.

“NUVIA was founded in early 2019 with the goal of reimagining silicon design to deliver industry-leading performance and energy efficiency for the data center. The company was founded by John Bruno, Manu Gulati and Gerard Williams III, who have collectively driven system engineering and silicon design for more than 20 chips, with more than 100 patents granted to date. NUVIA’s founders bring a rich silicon design heritage, having held a diverse array of engineering leadership roles at Google, Apple, ARM, Broadcom and AMD.”
...
The founding trio of Bruno, Gulati and Williams were key high-level architects at Apple whose expertise brought fruition to many generations of Apple’s SoCs and CPU microarchitectures. Williams was the chief architect on all of Apple’s CPU designs, including the recent Lightning core in the A13.

According do Andrei, they took quite a few extra people from Apple, and have quite a few people from SARC (Samsungs CPU guys)

On top of that they have the longstanding Red-hats ARM server guru as Vice President (Jon Masters), etc ...

Quite a lot of top talent, let's see what they can do!

 

[DrMrLordX]

@Gideon

Kinda makes you wonder if Apple is played out now. There's only so much talent to go around.

 

[Thunder 57]

I'm surprised this got so little reception. Let me quote parts of the Anandtech's article on NUVIA.


According do Andrei, they took quite a few extra people from Apple, and have quite a few people from SARC (Samsungs CPU guys)

On top of that they have the longstanding Red-hats ARM server guru as Vice President (Jon Masters), etc ...

Quite a lot of top talent, let's see what they can do!

I think the problem is they will need a heck of a lot more than the 53 million they have raised so far. But, if it looks promising they should be able to secure more funding.

 

[name99]

@Gideon

Kinda makes you wonder if Apple is played out now. There's only so much talent to go around.

It's unclear. Do people (or more precisely engineers) do spectacular work because they they are extra smart? Or because an ENVIRONMENT exists that allows them to operate at peak capacity?
Of course you need smart people, but there are actually quite a few of those. I think the real difference between companies is not their individuals, but the way they set up those individuals to interact with each other.

Assuming Apple maintains their existing chip management approach (which seems a reasonable assumption given that Srouji is still there), I suspect they'll do OK.

 

[Andrei.]

Keep in mind that we won't see post-Williams developments for another 2 years, and then they hired Mike Fillipo from Arm so I guess they're in a good state to continue their trajectory.

 

[DrMrLordX]

It's unclear. Do people (or more precisely engineers) do spectacular work because they they are extra smart? Or because an ENVIRONMENT exists that allows them to operate at peak capacity?

Hard to say. Arguably Intel is still doing great design work - none of us have seen Willow or Golden Cove. Yet the results, yech! Something changed there since 2006 that made it into a less-productive environment for implementing good designs. Was it people or was it the environment, or was it people changing the environment?

Assuming Apple maintains their existing chip management approach (which seems a reasonable assumption given that Srouji is still there), I suspect they'll do OK.

If the conditions are the same, you can't expect the remaining people at Apple to do a worse job. Depends on who they bring in to replace the Apple emmigrants and how they change workflow/environment.

 

[Richie Rich]

It's unclear. Do people (or more precisely engineers) do spectacular work because they they are extra smart? Or because an ENVIRONMENT exists that allows them to operate at peak capacity?
Of course you need smart people, but there are actually quite a few of those. I think the real difference between companies is not their individuals, but the way they set up those individuals to interact with each other.

Individuals is always the most important key factor. They create the environment because they start the companies up. Their visions inspire the others and they lead the whole company in the right direction. No matter if we are talking about Jobs, Musk, Gates, Keller etc. There are so many examples how bad management killed enviroment and forced best people to leave company resulting in company collapse. Remember Cyrix, DEC people went mostly to AMD and DEC was bought by Compaq and shut down later, Keller left AMD because they stopped his ambitious project and than with Bulldozer under Dirk Meyer practically collapsed whole company, Apple fired Jobs and almost collapsed under Sculley? Leaders are the most important.

I wouldn't be surprised they left due to seeing huge potential of actual Apple cores and frustration from wasting this potential in the same time. Apple is clearly approx 6 years ahead in CPU development over Cortex cores, and 4 years ahead of Intel and AMD. Thats huge motivation to take majority of server market for example. They can be the new DEC of 21st century. They cannot fail unless Apple will do the same move which is very unlikely. At least they can do decent exit by selling the company

 

[soresu]

Individuals is always the most important key factor. They create the environment because they start the companies up. Their visions inspire the others and they lead the whole company in the right direction. No matter if we are talking about Jobs, Musk, Gates, Keller etc. There are so many examples how bad management killed enviroment and forced best people to leave company resulting in company collapse. Remember Cyrix, DEC people went mostly to AMD and DEC was bought by Compaq and shut down later, Keller left AMD because they stopped his ambitious project and than with Bulldozer under Dirk Meyer practically collapsed whole company, Apple fired Jobs and almost collapsed under Sculley? Leaders are the most important.

I wouldn't be surprised they left due to seeing huge potential of actual Apple cores and frustration from wasting this potential in the same time. Apple is clearly approx 6 years ahead in CPU development over Cortex cores, and 4 years ahead of Intel and AMD. Thats huge motivation to take majority of server market for example. They can be the new DEC of 21st century. They cannot fail unless Apple will do the same move which is very unlikely. At least they can do decent exit by selling the company

First off, K10 came after K8, not Bulldozer - people are apt to forget K10 because of Bulldozer, but bugged Agena aside it was a solid improvement over K8, I was quite fond of my quad Deneb rig to be honest.

Keller and Jobs do not belong in the same context as you put them here - they literally existed within the same company for a time, in very different levels of the company, and Keller himself has typically been a transitory figure going from one company to another, while Jobs had his fingers in more than one pie (Pixar).

Lisa Su is a more appropriate figure to compare with Jobs - considering he was little more than a very savvy (and extremely cranky) business man when you subtract the myth and bluster.

Lastly, that 6 year Apple lead over Cortex figure is trash talk.

We don't know anything yet about even Hercules/A78 - and all we know of Matterhorn core that follows is that it brings MatMul instructions which accelerate machine learning work loads.

We know even less now about ARM's future plans than we did from that ambiguous leaked 2015 roadmap - given A77 had dramatic changes to the core on the heels of the huge change to 4 wide in A76, Matterhorn could well end up being a completely different beast altogether, especially if it is essentially another ground up design.

Something you might want to bare in mind is that ARM are not precisely aiming for Apple Axx parity, they are at the end of the day trying to get as many licensees as possible, and Axx cores are not small when considering the cost of area and power for a potential whole SoC design - so as much as it has frustrated us Android owners awaiting high end emulator parity with desktop Windows, it is always going to lag behind for those reasons.

Having said that, at this point there is already far more compute power in A76 than is being used by the vast majority of Android applications - there is still no Android PS2 emulator remotely matching PCSX2 in compatibility and speed years after Dolphin emu got to reasonable levels of usability on the platform.

There are also zero easily obtainable ARM SBC's that have anything stronger than A73 as far as I am aware - Rockchip's RK3588 couldn't come sooner for me.

 

[DrMrLordX]

There are also zero easily obtainable ARM SBC's that have anything stronger than A73 as far as I am aware - Rockchip's RK3588 couldn't come sooner for me.

I know, right? For anyone that wants to play with A76, it's basically . . . buy a Snapdragon 855 phone and try to run Linux on top of Android. Which is (sort of) what I'm doing soon. Wish me luck.

 

[soresu]

I know, right? For anyone that wants to play with A76, it's basically . . . buy a Snapdragon 855 phone and try to run Linux on top of Android. Which is (sort of) what I'm doing soon. Wish me luck.

Not sure if it's ever been attempted to take the PCB out and ditch the battery - many modern phones have HDMI/DP alt mode over USB C so that should take care of display/audio, assuming that you can shunt enough power with mains alone.

 

[Nothingness]

It's unclear. Do people (or more precisely engineers) do spectacular work because they they are extra smart? Or because an ENVIRONMENT exists that allows them to operate at peak capacity?
Of course you need smart people, but there are actually quite a few of those. I think the real difference between companies is not their individuals, but the way they set up those individuals to interact with each other.

I know people working at Apple and I'm not sure I'd work there (and I indeed refused three times proposals for interview from them): paranoiac environment (secrecy there is completely mad which reduces interactions between engineers to a minimum), not a lot of room for being creative (though I guess that depends on your level in the org chart), the obligation to work beyond hours (I do 50 hours a week but because I choose on what I work), etc. I might consider working for them if they open a design center close to where I live, but moving to the UK or the US and work for them is no option, even with their great recruiting package.

Note I like Apple and admire what they achieved in the chip industry. It just doesn't match what I (and many great engineers I know) want as a day job.

 

[Andrei.]

Lastly, that 6 year Apple lead over Cortex figure is trash talk.

We don't know anything yet about even Hercules/A78 - and all we know of Matterhorn core that follows is that it brings MatMul instructions which accelerate machine learning work loads.

Hercules is a smaller update. Matterhorn is a v9 high perf core with 256-bit SVE that finally gives up priority on small die area, and will be a super wide µarch. I would expect it to catch up with the A12/A13 - the issue is that yes that would indeed mean a 3 year gap between products, but that's also across process nodes, meaning the µarch gap would be 4-5 years. Not really trash talk, just reality of things. The only saving grace is if Apple significantly slows down their yearly improvements from here on, which based on what I heard, isn't happening.

 

[Nothingness]

Lastly, that 6 year Apple lead over Cortex figure is trash talk.

It currently is between 2 and 3 years. But that gap will reduce for sure.

There are also zero easily obtainable ARM SBC's that have anything stronger than A73 as far as I am aware - Rockchip's RK3588 couldn't come sooner for me.

I have a board with a Cortex-A75 (Snapdragon 845), but yeah I'd like something with more performance such as A76 (and preferably that does not reach crazy temperatures due to no fan).

 

[soresu]

It currently is between 2 and 3 years. But that gap will reduce for sure.


I have a board with a Cortex-A75 (Snapdragon 845), but yeah I'd like something with more performance such as A76 (and preferably that does not reach crazy temperatures due to no fan).

I think the whole Huawei shebang recently has stopped a Kirin 980 and 990 based HiKey board from coming out - a shame as I would definitely pony up for that.

 

[soresu]

Hercules is a smaller update.

Smaller than what?

A75, A76 or A77?

I'd say each of them were pretty significant, even A73 was a decent uptick on efficiency alone.

Are you sure they are throwing in the towel on area efficiency with Matterhorn?

Because Sophia cores tend to be the sleeker breed going by their past output.

A76 and A77 being big and bigger gels pretty closely with Austin's previous output win A15->A57->A72, so it would seem to be a serious change in direction for the Sophia team to go super size me for Matterhorn/A90?/A900? (going with 9s as a name direction for v9, also gels with A9 their first core!).

Either way even A76 is already blowing my mind, at the moment I'd be more than happy to have a half decent clocked SBC with 4C's of it running my silent media/browsing/NAS box instead of RPi 4 - which while being a great upgrade over RPi 3, the A72 is still a wee bit slow for full PC replacement of non gaming and CG work needs, and no where near fast enough for decoding AV1 beyond maybe basic 1080p.

 

[soresu]

but that's also across process nodes, meaning the µarch gap would be 4-5 years

Though A76 is significantly smaller than A12/13 right?

If so, then that gap is more a question of ARM's ambition on x node rather than an actual gap in capability - again I do think that ARM are constrained by the multitude of their licensees, as much as we would all like ARM to have parity with Apple, it may be in their interest to lag behind on core width in order to maintain maximum target audience.

If Nothingness is correct about A76 perf/watt, then Apple's core is still functionally about equal anyway in terms of battery life, if not raw performance - which is what the licensees care about most I would imagine.

Regarding Matterhorn width/core size though, I guess if anything could sway ARM to swing big on core width, it would be getting Samsung back as a high end core licensee for flagship Exynos after losing out on their high end royalties for 5 years since Mongoose launched - the potential loss of Huawei as a high end customer has to be a driving force in their decisions around designing to please those high end customers.

 

[Andrei.]

Smaller than what?

A75, A76 or A77?

I'd say each of them were pretty significant, even A73 was a decent uptick on efficiency alone.

Arm Unveils Client CPU Performance Roadmap Through 2020 - Taking Intel Head On

www.anandtech.com

Hercules is going to be the smallest update perf wise from the Austin family, ~10-15% upgrade with process.

Are you sure they are throwing in the towel on area efficiency with Matterhorn?

Because Sophia cores tend to be the sleeker breed going by their past output.

What Sophia made in the past doesn't constrict them on what they work on. Yes Matterhorn is a high-perf/HPC super-wide core. They couldn't go after this market in the past because they didn't have the design throughput and most customers wanted area efficient cores, but I think there's now more demand for Apple core competitors and they also want to use it for HPC and a base for future Neoverse CPUs, so essentially they're going for as high performance as they can even if it blows up area. At TechCon I think they said it (I don't remember exactly) it would be double the power or more over the power of current Cortex designs, more in line with what Apple cores look like. In essence I'm expecting double everything, which actually would be good in terms of PPA.

 

[soresu]

Arm Unveils Client CPU Performance Roadmap Through 2020 - Taking Intel Head On

www.anandtech.com

Hercules is going to be the smallest update perf wise from the Austin family, ~10-15% upgrade with process.
What Sophia made in the past doesn't constrict them on what they work on. Yes Matterhorn is a high-perf/HPC super-wide core. They couldn't go after this market in the past because they didn't have the design throughput and most customers wanted area efficient cores, but I think there's now more demand for Apple core competitors and they also want to use it for HPC and a base for future Neoverse CPUs, so essentially they're going for as high performance as they can even if it blows up area. At TechCon I think they said it (I don't remember exactly) it would be double the power or more over the power of current Cortex designs, more in line with what Apple cores look like. In essence I'm expecting double everything, which actually would be good in terms of PPA.

The media coverage on TechCon was gratingly sparse, even Charbax was strangely silent on it, despite ARM kind of being his thing - do you have any more detailed coverage links than the keynote posted on Anandtech?

 

[Andrei.]

The media coverage on TechCon was gratingly sparse, even Charbax was strangely silent on it, despite ARM kind of being his thing - do you have any more detailed coverage links than the keynote posted on Anandtech?

There was no direct coverage of the keynotes allowed. In general nothing concrete other than "we're working on bigger stuff and make everything better" and that Matterhorn will have 256-bit SIMD. It was just a teaser, expect more info next year.

 

[name99]

I know people working at Apple and I'm not sure I'd work there (and I indeed refused three times proposals for interview from them): paranoiac environment (secrecy there is completely mad which reduces interactions between engineers to a minimum), not a lot of room for being creative (though I guess that depends on your level in the org chart), the obligation to work beyond hours (I do 50 hours a week but because I choose on what I work), etc. I might consider working for them if they open a design center close to where I live, but moving to the UK or the US and work for them is no option, even with their great recruiting package.

Note I like Apple and admire what they achieved in the chip industry. It just doesn't match what I (and many great engineers I know) want as a day job.

Of course I worked there for ten years and loved every minute of it...
They treated me extraordinarily well, I had very good colleagues, they let me control my hours as long as I delivered.

Some people work for love of the product, some for the paycheck. If the paycheck is ALL that motivates you, it's probably not a good idea to work at Apple (or any other environment that expects more than that).

 

[name99]

Though A76 is significantly smaller than A12/13 right?

If so, then that gap is more a question of ARM's ambition on x node rather than an actual gap in capability - again I do think that ARM are constrained by the multitude of their licensees, as much as we would all like ARM to have parity with Apple, it may be in their interest to lag behind on core width in order to maintain maximum target audience.

If Nothingness is correct about A76 perf/watt, then Apple's core is still functionally about equal anyway in terms of battery life, if not raw performance - which is what the licensees care about most I would imagine.

Regarding Matterhorn width/core size though, I guess if anything could sway ARM to swing big on core width, it would be getting Samsung back as a high end core licensee for flagship Exynos after losing out on their high end royalties for 5 years since Mongoose launched - the potential loss of Huawei as a high end customer has to be a driving force in their decisions around designing to please those high end customers.

Don't get carried away by this sort of thinking. Even A12 or A13 are just not that large!
Look at an A13 die shot. The entire CU complex (2 large, 4 small cores, huge L2 and L3) is ~15% of a 98mm^2 die. What are you saving the area FOR?

Yes, if you want hundreds of cores on a die, or if you're creating a core that's a controller for some other purpose die size matters. But in the space of interest right now (eg phones, desktops) what's the difference between 15mm^2 for the CPU complex and 5mm^2? Are you happy to save 10mm^2 (so 10% of the SoC area) at the cost of losing 30 to 50% of your CPU performance? What's the point?