Intel Candidly Discusses Troubles at Credit Suisse 22nd Annual TMT Conference

[crazyeyesreaper]

For years Intel was able to maintain their endless tick-tock cycle however with the switch from 14nm to 10nm Intel realized all too late that they had bitten off more than they could chew. According to Robert Swan, Intel's Interim Chief Executive Officer and Chief Financial Officer, "we set out in the transition to 10 nm to attempt to scale much faster than we ever had at a time when I think most would argue the technology and the science and the challenges are more challenging they've ever been. So, we took a fairly aggressive scaling factor, roughly 2x of what the competitors do. So, we went for the analogy that a grand slam, I think, when the competition was hitting really solid singles."

Essentially Intel had hedged their bets that they could take a revolutionary step instead of the more typical evolutionary one thereby leaving their competition behind. Instead, it's resulted in the current situation that we are all very much aware of, that Intel is far behind their original predicted schedule. While that timeline has since been revised and they are now on course to release 10nm products in 2019, and 2020 Intel has also made it known that they plan to regain their leadership position as that transition begins.


In regards to supply constraints, Swan elaborated that they will continue to prioritize their Xeon product line first with their traditional small core and other devices coming second. Meaning those products will see continued supply shortages for the foreseeable future since that ties into Intel's eventual transition to 10nm. To help mitigate the issue, Intel has redeployed some of their 14nm equipment, but it is not enough to completely solve the supply problem. Right now they are trying to work with customers to figure out what constitutes actual demand rather than just inventory hoarding. Thus far Intel considers overall inventory healthy going from Q3 to Q4. For some customers this may be true, however for enthusiasts, that is not the case. Intel's latest Core 9000 series processors are in minimal supply having sold out pretty much everywhere making them little more than imaginary offerings for the PC gaming and enthusiast market. So while the inventory of certain products may be in a healthy state, you could say Intel's looking at the situation with rose-colored glasses.

You can check the entire transcript for more in-depth details from Intel at the source below.

View at TechPowerUp Main Site

 

[R0H1T]

For years Intel was able to maintain their endless tick-tock cycle

That's a bit of an exaggeration, Tick Tock lasted less than a decade & in fact barely 6~7 years if you take Intel's roadmaps strictly over the years.

 

[hat]

"Essentially Intel had hedged their bets that they could take a revolutionary step instead of the more typical evolutionary one thereby leaving their competition behind."

Is that why they sat on their hands since... Nehalem?

I dunno, man... things are looking grim for Intel, at the moment. I'm not so sure upgrading from 14nm to 10nm is going to be this fantastic thing... not without a major architecture overhaul, the likes of which not seen from Intel since Nehalem, or arguably Sandy Bridge. Skylake v6 on 10nm is not going to cut it.

 

[dj-electric]

Jim Keller is not at Intel to drink coffee and tell jokes.

 

[Vayra86]

Jim Keller is not at Intel to drink coffee and tell jokes.

I don't think its very hard to imagine what he's doing over there: designing a chiplet based CPU that can scale well and mitigates yield risks, trying very hard not to look like a carbon copy of Zen.

 

[londiste]

"Essentially Intel had hedged their bets that they could take a revolutionary step instead of the more typical evolutionary one thereby leaving their competition behind."
Is that why they sat on their hands since... Nehalem?

The context of the quote is manufacturing process node. 14nm > 10nm. They took more risks that we now know penalized them heavily. Architecture is a different discussion.

I dunno, man... things are looking grim for Intel, at the moment. I'm not so sure upgrading from 14nm to 10nm is going to be this fantastic thing... not without a major architecture overhaul, the likes of which not seen from Intel since Nehalem, or arguably Sandy Bridge. Skylake v6 on 10nm is not going to cut it.

Why not? There are variables that we do not know, mainly power consumption and possible clocks but pretty much everyone expects similar results to what 14nm > 7nm does for TSMC customers - Vega 20, Zen2, mobile SOCs. So far that boost should be far from negligible. If they can get 8 Skylake cores like 9900K out in under 100nm and half the power consumption (that is theoretically doable barring frequency walls) that would be good for them.

I don't think its very hard to imagine what he's doing over there: designing a chiplet based CPU that can scale well and mitigates yield risks, trying very hard not to look like a carbon copy of Zen.

All Intel has to do for a Zen lookalike glueing is increase the amount of UPI links to at least 3 in all their dies, that'll do it. Hopefully they manage to do something more innovative though.

 

[Vayra86]

All Intel has to do for a Zen lookalike glueing is increase the amount of UPI links to at least 3 in all their dies, that'll do it. Hopefully they manage to do something more innovative though.

Intel's problem isn't the interconnect, though. Its problem is also not scalability, Core scales fine in terms of performance. But with that, it also scales horribly in terms of risk and cost. Intel CPUs are too expensive to make at this point. Its not difficult to make ever faster CPUs, but it is difficult doing so while keeping costs under control, and the only viable product is one you can sell. Zen's strength is precisely its low risk, low cost perspective.

As for other innovative things... Jim Keller didn't manage to surpass Intel's Core in terms of IPC, it has a marginal efficiency lead per clock over Core, and... that's it. I think its realistic to consider that we've got the basic architecture at a near-optimal state and the only gains come from optimizing it based on core counts and power targets. We're also not seeing a radically different CPU design on ARM either, its just copying over the best practices developed under x86 mostly. Can't keep reinventing the wheel and think it keeps getting more round.

That also explains very well why Intel wanted to push its node shrinks so hard: its the place where most of the gains are to be had - not architecture.

 

[Midland Dog]

Jim Keller is not at Intel to drink coffee and tell jokes.

hes there to fix coffee joke, i mean lake

 

[londiste]

Intel's problem isn't the interconnect, though. Its problem is also not scalability, Core scales fine in terms of performance. But with that, it also scales horribly in terms of risk and cost. Intel CPUs are too expensive to make at this point. Its not difficult to make ever faster CPUs, but it is difficult doing so while keeping costs under control, and the only viable product is one you can sell. Zen's strength is precisely its low risk, low cost perspective.

Oh, I seriously doubt it. Intel is just taking much larger margins and has been unwilling to let that go. That would be the case even at same prices, not the larger MSRPs as they are (not even mentioning the current overpriced situation). The die in 9900K is still 20% smaller than Zen/Zen+ die. Die in 8700K is 35% smaller than Zen/Zen+ and that is still a competitive CPU for a large part of the market.

That also explains very well why Intel wanted to push its node shrinks so hard: its the place where most of the gains are to be had - not architecture.

Yep. And that is not only Intel. The others just rely on TSMC and Samsung to push the node. AMD, Apple, Samsung and other ARM implementers as well as AMD/Nvidia for GPUs and so on. Until there are gains to be had in shrinking nodes that path will be taken.

Edit:
tl;dr from what Intel said and we directly care about:
- Their current 10nm plan: end of 2019 client CPUs and FCPGAs with data center CPUs following early 2020.
- They expected single digit % rise of demand but got in teen %s, especially in datacenters with larger dies and thus larger capacity usage.
- They expected 10nm to be there and fabs were prepared. Now they brought some manufacturing capacity ready for 10nm back to 14nm (that probably took most of the mentioned $1.5B investment).

 

[bajs11]

great! by 2050 we may will have a 10 picometer Skylake refreshx34 called Guanolake
and it will be a much more imaginary offering than anyone alive now can imagine

 

[yeeeeman]

I don't think its very hard to imagine what he's doing over there: designing a chiplet based CPU that can scale well and mitigates yield risks, trying very hard not to look like a carbon copy of Zen.

Yep and actually they can learn from the mistakes AMD did with its chiplet design and make a better product.

 

[m4dn355]

Yep and actually they can learn from the mistakes AMD did with its chiplet design and make a better product.

It is gonna be intel's chiplet v. 1.0 vs. AMD's chiplet v. 3.1.
Game of cat and mouse, in this case AMD being the cat.

 

[Vya Domus]

Jim Keller is not at Intel to drink coffee and tell jokes.

He is also not responsible for absolutely everything like most think he is.

 

[bug]

The thing everyone overlooks is that while Intel bit off more than they could chew, when they finally chew it, they may still end up ahead. It's not guaranteed by a long shot, but I'd be really surprised if they didn't use at least some of this extra time to build a better Ice Lake than we would have gotten initially.

 

[m4dn355]

The thing everyone overlooks is that while Intel bit off more than they could chew, when they finally chew it, they may still end up ahead. It's not guaranteed by a long shot, but I'd be really surprised if they didn't use at least some of this extra time to build a better Ice Lake than we would have gotten initially.

It is not at all unexpected for intel to leapfrog AMD like in the days of core2duo or sandy bridge. Nevertheless 10nm is currently intel's biggest foe.

 

[bug]

It is not at all unexpected for intel to leapfrog AMD like in the days of core2duo or sandy bridge. Nevertheless 10nm is currently intel's biggest foe.

I'm thinking more like the problem Intel faces now, the others will probably face in the (near?) future. But we'll see. All I'm saying is just don't count Intel out just because they're having trouble with 10nm.

 

[mak1skav]

I don't think that anyone is "counting Intel out" but at the same time we can think that it was easier for Intel to look good in the past when there was no competition from AMD.

 

[bug]

I don't think that anyone is "counting Intel out" but at the same time we can think that it was easier for Intel to look good in the past when there was no competition from AMD.

True, but we don't want that. We want them on their toes

 

[yeeeeman]

It is gonna be intel's chiplet v. 1.0 vs. AMD's chiplet v. 3.1.
Game of cat and mouse, in this case AMD being the cat.

Well, we'll see about that. My opinion is that it won't be AMD v 3.1, but more like v 2.0. Lets not forget that we are not talking about a new uArch, but just a new layout of internal resources and reconfiguration of the internal buses. Sure, removing the ring bus might require some re-thinking of the uArch, but lets not forget Intel has a lot of cash, a lot of people and many teams of design that work in parallel and only the best stuff gets the green light. They also execute much faster compared to AMD because of this.

 

[R0H1T]

Well, we'll see about that. My opinion is that it won't be AMD v 3.1, but more like v 2.0. Lets not forget that we are not talking about a new uArch, but just a new layout of internal resources and reconfiguration of the internal buses. Sure, removing the ring bus might require some re-thinking of the uArch, but lets not forget Intel has a lot of cash, a lot of people and many teams of design that work in parallel and only the best stuff gets the green light. They also execute much faster compared to AMD because of this.

They'll need something better than UPI, heck we're not even sure what UPI yields outside multi socket systems.

Well there was the P4, then Larrabee, Atom (for smartphones & tablets) & lately 10nm, at each step they've had mountains of cash & failed spectacularly in such instances.

 

[yeeeeman]

They will use Jim Keller to make a "more than infinity" fabric. I mean, you didn't expect for Jim to go to Intel and talk nothing about what he did at AMD?
Also, Atom was not a failure in its days. The Zenfone 3 (if I recall correctly) was just as fast as Snapdragon oferrings of that time while consuming slightly more power. Given we are talking about an uArch optimized for low power compared to one built for 240V plug, that was quite an achievement. The reason Intel left Atom behind was the small margin of profit.
P4 was bad, I agree, but with their resources they managed to keep pace with AMD until the last very iterations of Athlon 64, where AMD had, for a brief period the performance crown.
I think we are subestimating Intel here. They have great talent, great people. They just made the wrong bet on something that is very tricky to fix and that is chip fabrication. Until now, I think they handled it pretty well, but things start to get risky for them if they don't do something in the next 6 months (when Zen 2 is rumoured to launch).

 

[R0H1T]

They will use Jim Keller to make a "more than infinity" fabric. I mean, you didn't expect for Jim to go to Intel and talk nothing about what he did at AMD?

Sure but that takes time & needs a rejig of the uarch, sometimes a major overhaul. You don't put untested cores on the glue & hope it sticks, you put the glue after the fact.

 

[yeeeeman]

Sure, but you have to understand that they know what the other is doing much more early than we know. Intel probably knows about Zen 2 chiplet stuff from long ago and they already had time to develop new interfaces, protocols and stuff. Jim keller is the guy to fit some unknown pieces of the puzzle. They have EMIB, they did the Intel-Radeon chip to test stuff. They are rumoured to have a brand new uArch named Ocean Cove (redesigned from ground up, not a successor of Core). They probably have much more stuff than we know in their hands and they find out stuff much more early than we do. I think they will have MCM chip in 1-2 years time which is ok (at least from desktop market point of view, for HPC it will could be disaster if Cascade Lake AP is slower and consumes more power compared to Rome), since chiplets are great in many core applications more than everything, not in gaming.

 

[bug]

They will use Jim Keller to make a "more than infinity" fabric. I mean, you didn't expect for Jim to go to Intel and talk nothing about what he did at AMD?

I would expect any professional worth their paycheck to first evaluate where Intel stands and then come up with a plan to work from there.
Doing for company B what you did (and worked) for company A is a recipe for disaster.

 

[LemmingOverlord]

Jim Keller is not at Intel to drink coffee and tell jokes.

Is that a Rowdy Roddy Piper joke? 'Cos if it is... I feel ya

*finger guns*