[Anandtech] Intel's Architecture Day 2018

[PeterScott]

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

I know some bits of this were getting some mention in other threads, but it seems like it was big enough to warrant it's own discussion.

Highlights for me were a new architecture (finally) Sunny cove, that looks to have potential (4-5 wide allocation) to deliver Intels first real IPC improvement in MANY years, and also what should be a nice improvement in the the IGP, and a look at Intels multichip solutions(Foveros) and Intel Big-little.

The Q&A was also somewhat interesting, like where they admit being node-locked on their designs really messed them up, and they will be more node agnostic in the future.

Overall, more actual news than we have seen from Intel in some time.

 

[Arachnotronic]

At the moment it's not clear what Intel is planning for the next year or two!

Hopefully a roadmap leaks or something.

 

[jpiniero]

Hopefully a roadmap leaks or something.

There was that leaked NUC roadmap but all that confirmed is that you won't see that much of 10 nm even through 2020.

 

[mikk]

NUC roadmap is uninteresting, we need a proper 5Q mainstream platform roadmap, preferably for both desktop and notebook. Maybe even more for desktop because I believe Intel is trying to launch ICL-U in H2 2019 even if it's a low volume launch, everything else is unclear.

 

[jpiniero]

NUC roadmap is uninteresting, we need a proper 5Q mainstream platform roadmap, preferably for both desktop and notebook. Maybe even more for desktop because I believe Intel is trying to launch ICL-U in H2 2019 even if it's a low volume launch, everything else is unclear.

I guess my point is that if Intel thought the Icelake supply was going to be great, they would have put together an Icelake NUC and it would have been on the roadmap.

 

[krumme]

No.

Ok. Same shit as old days then with a minor twist.
How are they going to fund 5 and 3nm? At that time AMD isn't tied to gf. Its new times. The situation is radically different.

 

[DrMrLordX]

I don't see Apple going for servers, or being able to cook a 60+ core dies either.
They make decent SoCs and so be it.

Decent SoCs? Uh. Okay. Obviously we do not see the A12x in the same light. Based on what I've seen, their tech is absolutely phenomenal. It's wasted on the phone/tablet business. AMD and Intel should be very worried . . . Intel most of all. Of course Apple will not go for servers, but they may well license their custom designs to a company that wants that market (Cavium, or Qualcomm before they shuttered their own ARM server effort). I'm sure they CAN cook 60+core dice since Cavium can basically do it already.

So to return to what I was saying, that Sapphire Rapids was the most interesting of their future products to me on the basis that it represented their one great chance to introduce a new base design on which they could then make fruitful future iterations (as opposed to the slow grinding slog of Core): in light of the fact that Sapphire Rapids may well be more Core, I must say that I am not really not interested in any of their future designs. Feel free to surprise me, Intel.

 

[Yotsugi]

Decent SoCs? Uh. Okay.

Their fat cores are nice, everything besides that is >okay.

I'm sure they CAN cook 60+core dice since Cavium can basically do it already

Cavium, was, like, a networking company that knew how to scale stuff up.

So to return to what I was saying, that Sapphire Rapids was the most interesting of their future products to me on the basis that it represented their one great chance to introduce a new base design on which they could then make fruitful future iterations (as opposed to the slow grinding slog of Core): in light of the fact that Sapphire Rapids may well be more Core, I must say that I am not really not interested in any of their future designs. Feel free to surprise me, Intel.

It's Core because Intel calls their wide and deep OoO designs Core.

 

[Arachnotronic]

Ok. Same shit as old days then with a minor twist.
How are they going to fund 5 and 3nm? At that time AMD isn't tied to gf. Its new times. The situation is radically different.

Funding 5nm and 3nm won't be a problem for Intel -- they literally generate more in revenue from logic IC sales than Samsung or TSMC combined.

Intel's process problems have been a result of bad technology execution, not due to lack of revenue/wafer scale.

 

[DrMrLordX]

Their fat cores are nice, everything besides that is >okay.

Understatement of the month, but okay.

Cavium, was, like, a networking company that knew how to scale stuff up.

And when Apple licenses their A-series core designs to Cavium, what do you suppose happens then?

It's Core because Intel calls their wide and deep OoO designs Core.

It's Core because the design is an iterative expansion on Sandy Bridge (and arguably Nehalem, though a lot changed from Nehalem-> Sandy). If the only thing that makes Core is . . . Core, then we may as well call Zen "Core" as well.

 

[Yotsugi]

Understatement of the month, but okay.

They're nice but that's it.

And when Apple licenses their A-series core designs to Cavium, what do you suppose happens then?

Nothing.
x86 will still rule the world.

It's Core because the design is an iterative expansion on Sandy Bridge (and arguably Nehalem, though a lot changed from Nehalem-> Sandy).

You can trace it all down to the good old P6.

If the only thing that makes Core is . . . Core, then we may as well call Zen "Core" as well.

I mean, pretty much.
Cyclone and S.LSI M3 are also of the same breed.

 

[DrMrLordX]

Nothing.
x86 will still rule the world.

Mmhmm.

You can trace it all down to the good old P6.

I wasn't going to go back that far out of respect to some of the changes from the Core2 generation to Nehalem, but yeah, technically . . . Intel has been flogging the same design since the late 90s.

 

[Yotsugi]

Mmhmm.

I mean, yeah.
x86 servers are red-hot now.

I wasn't going to go back that far out of respect to some of the changes from the Core2 generation to Nehalem, but yeah, technically . . . Intel has been flogging the same design since the late 90s.

Which isn't bad and how it should generally work.
Wide and reasonably deep OoO machines, they're great!

 

[Arachnotronic]

Mmhmm.



I wasn't going to go back that far out of respect to some of the changes from the Core2 generation to Nehalem, but yeah, technically . . . Intel has been flogging the same design since the late 90s.

Microarchitecture like any human discipline is iterative. This is like saying the invention of calculus wasn’t a big deal because it built upon the ideas that came before it.

The reason people don’t throw everything away and start from scratch is that it’s a good idea to reuse and extend ideas that worked.

 

[raghu78]

Funding 5nm and 3nm won't be a problem for Intel -- they literally generate more in revenue from logic IC sales than Samsung or TSMC combined.

Intel's process problems have been a result of bad technology execution, not due to lack of revenue/wafer scale.

Intel has left themselves terribly vulnerable with 10nm delays. TSMC 7nm based Rome server chips from AMD and other 7nm ARM server chips will arrive before 10nm Icelake-SP in late 2020. Whats more important is TSMC N7+ and N5 are on track for 2019 and 2020. Intel 10nm server products in 2021 will be competing with TSMC 5nm server products from AMD and ARM vendors. TSMC is on a leading edge node war with Intel. TSMC has said that HPC is their driver of growth from 2020. Guess at whose expense TSMC will gain silicon share. Yeah thats Intel. TSMC 5nm is going very well according to industry observers like Scotten Jones. TSMC is going to take a lot of silicon share from Intel over the next 3-5 years. btw thats not in just CPUs but also FPGAs and other segments where Intel is still doing quite well. TSMC 7nm products will ramp hard in 2019 for servers and at good yields as their process is in HVM from Q2 2018. Intel is right in the crosshairs of TSMC/AMD and they are going to ram the door down on the x86 monopoly of Intel especially in high margin servers and desktops.

btw Samsung crushes Intel and TSMC with their DRAM revenues. It does not matter if they are not big in logic revenue. The DRAM cash will allow Samsung to pursue leading edge for as long as TSMC or Intel.

http://www.icinsights.com/news/bull...uppliers-Forecast-To-Post-DoubleDigit-Gains-/

 

[TempAcc99]

You know well enough how that trainwreck ended.

Moving from complex OoO machine to exactly WHAT?

Why would a different ISA make, eww, any difference?
As long as it's still complex OoO machine.

The age of bold is far and away gone.

I'm gonna call AMD being first to chiplets and an IO die pretty bold.

I agree that Intel should come up with something new. Get rid of some legacy stuff. Even if that doesn't have a huge impact on die size it for sure makes validation faster as less things need to be validated. But just making minor changes and adding yet another AVX-version on top doesn't solve much. I mean most software doesn't even make use of AVX even if it could. think anything java, python or .net based. While python it would be the easiest, still standard packages usually aren't compiled against AVX.
Plus most consumer workloads hardly need it an then there is intels market segmentation with it probably preventing game devs investing heavily into it.

Anyway intel for sure needs somethign better than more AVX.

I don't see Apple going for servers, or being able to cook a 60+ core dies either.
They make decent SoCs and so be it.

Yeah Apples business model is clear. Lock consumers to their closed platform. They have no interest in servers. And even if they licensed the IP, it's not at all clear the A12 big-core could be made into a many core-chip. Then you will also need a mesh and all that pesky many core chips stuffs. Plus we should not forget the core is pretty big, bigger than skylake. No issue for apple because they sell it in products all costing >$700. Point being that such a chip would not be cheap at all. And then there is the software which often is the much bigger hurdle. You would need a team just to opitimze common Open-souce stuff for ARM.

 

[PotatoWithEarsOnSide]

If Core is Intel's 'wheel' then Intel will simply fade out of our memories. We've already established that the chiplet design in Zen 2 is going to be better than it, and offer significantly better scaling possibilities. If something exists that is better than your 'wheel' could ever be, then you really need to redesign from the ground up.

 

[DrMrLordX]

I mean, yeah.
x86 servers are red-hot now.

Fine then, you just watch.

Which isn't bad and how it should generally work.
Wide and reasonably deep OoO machines, they're great!

So you think Haswell -> Broadwell -> Skylake->Kabylake->Coffee Lake went swimmingly well? I don't.

Microarchitecture like any human discipline is iterative.

Intel has either lost their talent or pushed the boundaries of their design to its limits. Perhaps both. Iteration is not working for them. Innovate or get left behind.

 

[Spartak]

It pretty much depends on the state of 10nm.

...or not. Since they decoupled sunny cove from process.

 

[Yotsugi]

I'm gonna call AMD being first to chiplets and an IO die pretty bold.

Wow mommy Northbridges are sure awesome!

I agree that Intel should come up with something new

Should I point you at their advanced packaging or you're going to do it yourself?

I agree that Intel should come up with something new. Get rid of some legacy stuff. Even if that doesn't have a huge impact on die size it for sure makes validation faster as less things need to be validated

Legacy stuff barely costs anything.

So you think Haswell -> Broadwell -> Skylake->Kabylake->Coffee Lake went swimmingly well? I don't.

The latter three are the SAME core.
Node delays are bad mkaaay.

Intel has either lost their talent or pushed the boundaries of their design to its limits. Perhaps both. Iteration is not working for them. Innovate or get left behind.

It sure as hell works fine for them, as long as they're actually iterating.

...or not. Since they decoupled sunny cove from process.

Oh that it does.
They're changing their design approach.
BUT IT'S NOT YET TIME.

 

[Spartak]

The announcement was that they ARE ready. But we will see a year from now, probably earlier from leaks.

 

[PotatoWithEarsOnSide]

I know a good shrink...

 

[moinmoin]

Intel's issue is that they are iterative with the Core design and daring with the node processes. Both backfired, the latter failed and held up the former for way too long. Meanwhile AMD is daring with the system design (as they need to be in pursuit of catching up) and TSMC is getting ahead of Intel by being strictly iterative with the process node development.

Which isn't bad and how it should generally work.

It isn't good either if it comes with old baggage that isn't questioned ("working as designed") and turns out to be a security liability specific to that age old design (Meltdown). Of course wide and reasonably deep OoO designs are a keeper, but the fact that (depending on the ) all Core/P6 designs are affected point to a reluctance of questioning the base and instead always building and expanding upon it (not unlike the mess Windows OS turned into since the clean NT code base).

TSMC is on a leading edge node war with Intel.

Intel is not even participating in it in any way so far, which is of course fine with TSMC. I think Intel wasn't prepared for a pure play foundry outside of GloFo to get into the position of directly competing with them, never mind overtaking them. AMD's timely rejuvenation is a godsend for TSMC (and later likely also Samsung) considering the newer nodes are getting more expensive at a rate that the mobile mass market is no longer as easily targeted as before (and quantity is a necessity to keep the iterative process development going).

 

[Yotsugi]

It isn't good either if it comes with old baggage that isn't questioned ("working as designed") and turns out to be a security liability specific to that age old design (Meltdown)

That's mostly due to Intel's (corporate) attitude towards security (which is redacted SECURITY).
I mean, some Apple and ARM cores are vulnerable, too.

We've been here many times already,
yet you still continue to use profanity
in the tech forums.

AT Mod Usandthem

 

[PotatoWithEarsOnSide]

Their attitude isn't as you say.
Their attitude is profit first, profit second, and profit forever.
As you can see on the server side, their customers are buying new instead of taking the hits from fixes. That's a profitable model when the competition seriously lags, but it's a disastrous one when the competition gets ahead...which Epyc 2 will certainly do.
Now you begin to see the problem...profit forever is unattainable if you constantly screw your customers in the name of profit.

 

[Yotsugi]

Their attitude isn't as you say.

No, that's quite literally their internal attitude towards security, as it has always been.
Let's start counting IME/AMT holes!