Intel - Client 2.0 “In a few years we’ll be doing what Apple is doing today”

[vigilant]

Intel Moving to Chiplets: ‘Client 2.0’ for 7nm

www.anandtech.com

Interesting article on Intel essentially moving to a model that at least to me looks like the way Apple appears to be handling their SOCs for the last few years.

I encourage a read of it. The difference is Apple is already largely beating Intel in areas such as GPU, AI, low power consumption compute, and Media. Once Apple shows off an Intel competitive purpose built notebook class SOC I expect all of the pieces to be in place.

Thought I’d share, considering I’ve enjoyed the conversation around this area here lately.

 

[Tekguy0]

I find it hard to imagine that Intel will move towards Apple's model of SOCs. Keep in mind, Apple makes usually 2 models per generation, and sometimes only one, thus they can make each one pretty much perfect. However, Intel has already had mix-and-match for a while now, as can be seen with the large amount of chips with different CPU core counts, GPU core counts, and graphics/media accelerators such as Quick Sync. I like the idea of modular Silicon IP.

 

[Kpjoslee]

No, that is not what the article is implying. That article is talking about manufacturing, not the direction of Intel's future SoC. Intel is moving to multiple blocks of CPU, GPU, and IO which is similar to what AMD is doing right now, which allows each blocks to be manufactured on different processes for flexibility. Then moving down to individual IPs for more flexibility in the future.

 

[vigilant]

I find it hard to imagine that Intel will move towards Apple's model of SOCs. Keep in mind, Apple makes usually 2 models per generation, and sometimes only one, thus they can make each one pretty much perfect. However, Intel has already had mix-and-match for a while now, as can be seen with the large amount of chips with different CPU core counts, GPU core counts, and graphics/media accelerators such as Quick Sync. I like the idea of modular Silicon IP.

I don't think Intel is going to do SOCs, but it does look like they are looking to use a more modular scheme where they connect various different cores together to fit a given use case.

 

[vigilant]

No, that is not what the article is implying. That article is talking about manufacturing, not the direction of Intel's future SoC. Intel is moving to multiple blocks of CPU, GPU, and IO which is similar to what AMD is doing right now, which allows each blocks to be manufactured on different processes for flexibility. Then moving down to individual IPs for more flexibility in the future.

I should have clarified. I don't expect Intel to start building SOCs the same way Apple is. What I expect is Intel to use a similar approach of packaging various cores to fit a given use case, which Apple has been doing in their SOCs for years.

 

[leman]

I don't think Intel is going to do SOCs, but it does look like they are looking to use a more modular scheme where they connect various different cores together to fit a given use case.

Intel has been doing SoCs for years... it’s not really a novel concept. Each Intel CPU sold in last couple of years is a SoC. It’s just that their GPUs are still not capable enough to be considered a full-fledged system, but that might change with Tiger Lake.

 

[jz0309]

Intel has been working on this for years, do a search for "Foveros" ... problem they were/are having that everything was on a single piece of silicon so every component of their entire CPU was on the same technology node and that's not necessarily needed. With this new technology they can optimize each component in the best process for them and then bring them together.
Also know the Intel has for many many years "fused" their chips to be either a 8 core or 6 or 4 core with different speeds and other functionality enabled/disabled - same piece of silicon can be turned into many different products

 

[JPack]

No, that is not what the article is implying. That article is talking about manufacturing, not the direction of Intel's future SoC. Intel is moving to multiple blocks of CPU, GPU, and IO which is similar to what AMD is doing right now, which allows each blocks to be manufactured on different processes for flexibility. Then moving down to individual IPs for more flexibility in the future.

Agreed, Intel is proposing something entirely different from what Apple is doing today. The big word is "interposer."

There is something from a couple years ago where Intel describes the technology in detail.

Intel's Architecture Day 2018: The Future of Core, Intel GPUs, 10nm, and Hybrid x86

www.anandtech.com

 

[cmaier]

Agreed, Intel is proposing something entirely different from what Apple is doing today. The big word is "interposer."

There is something from a couple years ago where Intel describes the technology in detail.

Intel's Architecture Day 2018: The Future of Core, Intel GPUs, 10nm, and Hybrid x86

www.anandtech.com

I’m going to go ahead and claim I invented the use of interposers for CPU packages. Because why not?

Cliff Maier's PhD Dissertation

 

[deconstruct60]

I don't think Intel is going to do SOCs,

Errrr.

Newsroom Home

Intel Newsroom home page

newsroom.intel.com

and

The Intel Lakefield Deep Dive: Everything To Know About the First x86 Hybrid CPU

www.anandtech.com

Pack the RAM , CPU cores , GPU cores , and the PCH/IO hub into a package the size of a dime. That isn't SoC ?

but it does look like they are looking to use a more modular scheme where they connect various different cores together to fit a given use case.

Intel is going to use this elsewhere with a slightly different focus and emphasis. But deeper disinterest in doing a SoC now ( or over last several years ). That is a reach.

 

[deconstruct60]

I should have clarified. I don't expect Intel to start building SOCs the same way Apple is. What I expect is Intel to use a similar approach of packaging various cores to fit a given use case, which Apple has been doing in their SOCs for years.

Again a bit disconnected from what Intel has been doing. Intel has Xeon D , E , W , and SP line up. Intel has a mind bending load of CPU products coming out of their eyeballs. In part , Intel has too much product segmentation. Not too little.

This whole patch together lego blocks for even more permutations could end up being a dual edge sword where they could drift off into taking trying to be everything for everybody up to an even higher level of even more byzantine collection of CPU products.


It isn't that Apple is doing a SoC per se that is one of the primary advantages. Apple is winning in part because they haven't been chasing after everything for everybody. They don't even build a specific SoC for even half of the iOS/iPad product line up. The majority of products are all powered by "hand me down" SoCs. Some of them generation ( or two or three or four ) old.

That minimal number dies can get away with trick will be harder to do in the Mac market and still leave Macs competitive over the long term. Likewise the "hand me down" reuse as the enclosures have a much wider port and end user workload diversity.


Apple is likely also going to run into similar issues as Intel when get to place where want to weave in their own cellular modem and RF instracture. Being on bleeding edge TSMC isn't necessarily going to get you better Analog/Digital conversion implementations.

TSMC Updates on Node Availability Beyond Logic: Analog, HV, Sensors, RF

www.anandtech.com

Apple is the only major smartphone maker that is a primary buyer of discrete cell modems.
The space Apple has been sitting in isn't that broad. Their SoC in the watch is wrapping plastic around a logic board and calling it a "chip". That's OK, but also indicative to the relatively limited breadth of their implementations. Apple throwing the headphone jack overboard. Limiting port availability and diversity ( e.g., one and only one custom USB port). etc. are all as much having to deal with implementation diversity issues as it is simplifying the overall product.

 

[JMacHack]

The big word is "interposer."

Ever since Zen 3 dropped I've been convinced that unless other chip makers adopt the interposer design they're gonna fall behind. It's just too good from what I've seen.

 

[deconstruct60]

Ever since Zen 3 dropped I've been convinced that unless other chip makers adopt the interposer design they're gonna fall behind. It's just too good from what I've seen.

Other chip vendors?

Zen 3 doesn’t use interposer really either . There were multiple chip modules/packages before interposer came along . That is more so about the complexity of the connections than in making connections between dies .

If go to the broader definition then all of the packages that have smaller die contacts than logicboard pad/pins are interposed .

 

[thekev]

I find it hard to imagine that Intel will move towards Apple's model of SOCs. Keep in mind, Apple makes usually 2 models per generation, and sometimes only one, thus they can make each one pretty much perfect. However, Intel has already had mix-and-match for a while now, as can be seen with the large amount of chips with different CPU core counts, GPU core counts, and graphics/media accelerators such as Quick Sync. I like the idea of modular Silicon IP.

Most or at least many of those variants are due to binning, not separate processes. I would not be surprised if things like the Apple TV use chips that didn't bin well.

 

[cmaier]

3D stacks using diamond interposers with fins into thermal paste is the way to go.

 

[Tekguy0]

Most or at least many of those variants are due to binning, not separate processes. I would not be surprised if things like the Apple TV use chips that didn't bin well.

Sure, no doubt the best chips to to iPhones, the slightly less good ones go to iPads, and the worst to go the Apple TV and homepod. But the A12 had two designs, the A12, the A12X and the A12Z which is a binned version of the A12X w/ 1 extra GPU core. The A13 is only made in one version.

 

[JMacHack]

Other chip vendors?

Zen 3 doesn’t use interposer really either . There were multiple chip modules/packages before interposer came along . That is more so about the complexity of the connections than in making connections between dies .

If go to the broader definition then all of the packages that have smaller die contacts than logicboard pad/pins are interposed .

The thing they do with the chiplets and the i/o die separate on an interposer, whatever that's called. It might not be new but from papers I've seen it likely has benefits that outweigh monolithic designs.

 

[vigilant]

Sure, no doubt the best chips to to iPhones, the slightly less good ones go to iPads, and the worst to go the Apple TV and homepod. But the A12 had two designs, the A12, the A12X and the A12Z which is a binned version of the A12X w/ 1 extra GPU core. The A13 is only made in one version.

I’m not sure thats exactly how it works, at least not at this moment.

When the redesign of the Apple TV happened sure, it was if memory serves me correctly a single core variant of the iPad Product line.

I’d bet that the A11X was originally suppose to have that extra GPU core but couldn’t due to early yield issues, so they stored the good ones up to fill the line for a year and a half (roughly the amount of time since the previous iPad Pro launch happened). I’m not saying it’s a good, or bad thing, just saying that it is indeed a thing.

The lower end iPad models I don’t think are using lower binned parts... I’d expect they are using excess parts that can now be sold on more product lines as a result of the iPhone SOC line hitting scale better than it had at launch. To me at least, that makes the most sense.

Apples generalized tendency is to use what they can get at lower yields to fill the need of the iPhone, then once things shake out a few months later (roughly 6 months) release a variant of something that uses that same chip. I’d wager that they run everything at full capacity at the fab, and try to match ordering to meet the needs for a run, with the idea that excess capacity can be used to start a new product line, and try to stabilize demand.

 

[vigilant]

The thing they do with the chiplets and the i/o die separate on an interposer, whatever that's called. It might not be new but from papers I've seen it likely has benefits that outweigh monolithic designs.

I haven’t had enough time to focus on it, but I think it leads to a new way to design chips in general. Instead of trying to build a single SOC, build out the individual parts, and combine the known good ones pragmatically as a way to design systems. I don’t think Apple will be doing that in the near future, but it would be wrong to think they aren’t thinking about it.

 

[thekev]

Sure, no doubt the best chips to to iPhones, the slightly less good ones go to iPads, and the worst to go the Apple TV and homepod. But the A12 had two designs, the A12, the A12X and the A12Z which is a binned version of the A12X w/ 1 extra GPU core. The A13 is only made in one version.

Was it ever confirmed that the extra gpu core was a different hardware design as opposed to one core permanently turned off? The two aren't that dissimilar in these practices. Intel has a much broader range of hardware to cover, but I don't think within a particular processor line, different skus differ in terms of underlying chip design in most cases.

 

[deconstruct60]

The thing they do with the chiplets and the i/o die separate on an interposer, whatever that's called. It might not be new but from papers I've seen it likely has benefits that outweigh monolithic designs.

Not on maximum performance/watt. Which is Apple's main focus.

If you look at AMD's mobile "4000 series" offerings they don't use the approach. Probably won't for the "5000" series either (on refined 7nm). If can move to 5nm on "6000" series.... probably still monolithic.

It is a a better way to pursue the "maximum core count war" . Also a better way to build a "do everything, for everybody " SoC. Apple's SoC is a bit of a "black hole" subsuming other chips into a central package, but Apple has been selective there and desktops (or anything with lots of I/O ports and slots ) have major limits on just how far they can 'collapse" everything.

Apple is probably not going to chase the "more cores than anybody else" war in the workstation space. macOS isn't prepared to go there at all ( and there is zero benefit for the shared kernel scheduler aspects on the iOS, iPad OS, etc. products. ... which are the main money makers. )

 

[Tekguy0]

Was it ever confirmed that the extra gpu core was a different hardware design as opposed to one core permanently turned off? The two aren't that dissimilar in these practices. Intel has a much broader range of hardware to cover, but I don't think within a particular processor line, different skus differ in terms of underlying chip design in most cases.

No, it is the exact same hardware design, just with that extra core turned on. Apple supposedly tweaked the "thermal architecture" but it may have just been an improved heat spreader.

 

[deconstruct60]

No, it is the exact same hardware design, just with that extra core turned on. Apple supposedly tweaked the "thermal architecture" but it may have just been an improved heat spreader.

the Battery is basically the same. More likely Apple was doing two things.

1. Playing a yield boost game. As long as there was a defect in just one GPU "core" then could still get a working A12X out of the die. TSMC's 7nm yield rates have improved a several percentage points over the last two years. In short, it is cheaper for Apple to produce a A12Z now rather than two years ago. [ Not a huge amount but Apple can be a bit Scrooge McDuck when it comes to eeking out margins. ]

Apple was pretty aggressive with the A12X of moving it to 7nm rather early in the product maturity cycle.

2. Playing a "kick the can" game. The iPad Pro was much faster than any other tablet in 2018. Making it faster in 2018 really wouldn't do much to put a substantive gap on their competition. In 2020, it allows them to kick the can until 5nm is deeper into better yield rates. ( if they were pumping significantly more heat/power into the system then probably would need a bigger battery too. ).


The last couple of A-X series Apple has only moved on a process shrink (e.g,, A12 on 7nm , A13 also on 7nm so no X variant). Apple knew they were going to have to wait for 5nm arrive for A14X. The A12Z also gave them a 'fall back' if there was any hiccup with 5nm (and/or the camera changes they are looking to make). [ Decent change there is some small coupling to the Mac Apple Silicon implementation also. Several factors. So that being able to "bump" the iPad Pro for an extra 6-10 months probably was prudent. Apple is juggling more SoC designs so it is likely they would need more time. the A12Z gives them more time without impeding the iPad Pro (and any AppleTV variant that might use it. ) ]