Discussion Intel current and future Lakes & Rapids thread

[jpiniero]

EOY 2017 - Cannonlake Y
Q1 2018 - Cannonlake U 2C, with ramp up over the year, meeting their claims of "HVM 10nm 2018"

It's probably more like a paper launch Cannonlake Y in end Q2 2018, with volume in H2 2018. U likely will get canned. The EOY 2017 stuff is likely something that's not sellable like a 2+0 Y Celeron part but they are selling it because BK said they would ship something in 2017.

Q2 2018 - Coffeelake GT3e and 6C GT2

Launch of CFL 6+2 is January actually. I think 4+3e is Q2, but it'll be small volume for sure.

BTS 2018 - Coffeelake 8C Desktop

That's probably about right.

*GT3e takes extra time to make. They always came later. With Icelake generation we have the possibility of eDRAM being replaced with HBM.

I have to think Iris in general is going to get canned. Especially now that AMD is reasonably competitive on the CPU side there is no reason to think OEMs would bother with Iris when AMD would be so much faster in IGP gaming.

 

[Qwertilot]

Well they might always take it as a challenge and double down on iGPU instead.....

They do seem to have kind of given up though - they spent several years really pushing it but haven't really moved the bar at all for the past few years.

 

[IntelUser2000]

I'm telling you, there's no way Intel is shipping a 10 nm server part without EMIB. They obviously planned for it at one point but it's not happening now. The problems plaguing 10nm doesn't necessarily mean that 10++ would be delayed in kind too.

Your insistence on EMIB is strange.

Your point about 10nm++ applies to 10nm+. Remember Skylake? Broadwell couldn't clock much beyond 4GHz even if its life depended on it. That's why BDW-E and BDW-C suffered. Skylake comes and does 4.7GHz. Process failures on 14nm was basically why Desktops didn't exist for Broadwell.

Intel processors require operating at very high frequencies. It doesn't matter if it can make 700mm2 chips if it can't do 4+GHz frequencies. For Servers it can be ok with that.

And Icelake Server is 10nm+, not 10nm, not 10nm++.

They do seem to have kind of given up though - they spent several years really pushing it but haven't really moved the bar at all for the past few years.

They didn't really give up. They didn't have delayed process as part of their future. Remember they prided for so long on process leadership, and just 2-3 years ago it was true. Kaby and Coffee was a last minute decision.

Cannonlake is Gen 10, Icelake is Gen 11, and Tigerlake is Gen 12. Even nowadays process tech is one of the major contributors to performance. From 2009 to 2012 the iGPU portion went from 65nm to a 22nm process, one generation jump every year. Those were also the years where massive improvements happened in Intel iGPUs.

If they didn't have delays with process, we would have had Cannonlake GT3e/GT4e with 72/104EUs respectively on the high-end, with Icelake Gen 11 about to be released. The picture would have looked quite different.

 

[jpiniero]

Your insistence on EMIB is strange.

I'm convinced the problem with 10nm is functional yield; ie: most if not all of the wafer is busted. There's probably quality/clock speed problems too but they pale in comparison. They won't even get usable yield of a tiny 70 mm2 Cannonlake die until the middle of next year... so you can see why a big monolithic server chip isn't happening.

EMIB would mitigate the problem by using as small dies as possible. Obviously it doesn't solve the entire problem, and there's a cost to it, but they don't have any alternative.

Your point about 10nm++ applies to 10nm+. Remember Skylake? Broadwell couldn't clock much beyond 4GHz even if its life depended on it. That's why BDW-E and BDW-C suffered. Skylake comes and does 4.7GHz. Process failures on 14nm was basically why Desktops didn't exist for Broadwell.

Skylake is on the same process as Broadwell. Intel did release desktop chips for Broadwell, y'know.

 

[Dayman1225]

Launch of CFL 6+2 is January actually.

Not according to this.(Assuming you are talking mobile)

 

[PeterScott]

They didn't really give up. They didn't have delayed process as part of their future. Remember they prided for so long on process leadership, and just 2-3 years ago it was true. Kaby and Coffee was a last minute decision.

Cannonlake is Gen 10, Icelake is Gen 11, and Tigerlake is Gen 12. Even nowadays process tech is one of the major contributors to performance. From 2009 to 2012 the iGPU portion went from 65nm to a 22nm process, one generation jump every year. Those were also the years where massive improvements happened in Intel iGPUs.

If they didn't have delays with process, we would have had Cannonlake GT3e/GT4e with 72/104EUs respectively on the high-end, with Icelake Gen 11 about to be released. The picture would have looked quite different.

I think big factors in IGP stagnation was reaching "good enough" status for most users, and OEMs not being interested in buying models with bigger more expensive IGPs.

 

[Dayman1225]

My thoughts on release.

EOY 2017 - Cannonlake Y
Q1 2018 - Cannonlake U 2C, with ramp up over the year, meeting their claims of "HVM 10nm 2018"
Q2 2018 - Coffeelake GT3e and 6C GT2
BTS 2018 - Coffeelake 8C Desktop

Spring 2019 - Icelake Y
Q2 2019 - Coffeelake 6C, Kabylake-R, and Cannonlake U, all get replaced with Icelake derivatives
BTS 2019 - Icelake Desktop, and GT3e* mobile Icelake

*GT3e takes extra time to make. They always came later. With Icelake generation we have the possibility of eDRAM being replaced with HBM.

Here's my take on it

I agree with the Majority of what you have said but -

EOY 2017 - Cannonlake Y
Q1'18 - Cannonlake U 2C,
Q2'18 - Coffeelake GT3e and 6C GT2
H2'18 Icelake Y/U HVM - Due to BKs claims on "Higher complexity designs" in the July call.
BTS 2018 - Coffeelake 8C Desktop

H1 2019 for Icelake S/H

Regarding my Icelake Y/U optimism - it is mainly due to his(BKs) mention of "higher complexity designs" in July call - and this:

Source

Reminding you that this is a highly optimistic "prediction"

 

[IntelUser2000]

I think big factors in IGP stagnation was reaching "good enough" status for most users, and OEMs not being interested in buying models with bigger more expensive IGPs.

That's true, but good enough has been reached back in the Core 2 days. With Cannonlake they planned 40EUs for the base GT2, 72EU for the GT3e, and 104EU for the GT4e. Gen 11 and Icelake would have done even better. They want to move every generation from Cannonlake to Tigerlake. Clearly iGPU performance isn't something to be ignored.

Comparatively too. The Intel system that's on par/better than AMD chips in all metrics will command more sales and higher margins than if the GPU was substantially behind. With Bristol Ridge they didn't even have graphics lead.

 

[DrMrLordX]

Yea, with Icelake we'll see full Desktop lines, full Notebook lines, and -SP server parts all in 2019.

I still want to see some indication of the status of 10nm+. The earnings call specifically dealt with 10nm which will only ever be used for Cannonlake . . .

The EU nations are becoming bankrupt, even more so than the rest.

Exactly. They are hungry for outside capital sources.

I have to revise my assessment on Brian Kraznich. The Q3 results are quite good. Despite what looks like a very messy situation, they've effectively countered Ryzen with Coffeelake.

All it means is they're still making money. I don't think Ryzen shipped in sufficient volume to really hurt Intel. They're still struggling on the foundry side.

I find it amusing that in the same post you (1) bash Intel for shipping 50,000 CPUs as a paper launch and then (2) applaud AMD with EPYC coming in Q4 2017 when it was launched in June 2017:
https://www.anandtech.com/show/11562/amd-epyc-launch-event-live-blog-starts-4pm-et-

I don't recall anyone having problems buying 1950xs when they first started listing on sites.

 

[Jan Olšan]

I find it amusing that in the same post you (1) bash Intel for shipping 50,000 CPUs as a paper launch and then (2) applaud AMD with EPYC coming in Q4 2017 when it was launched in June 2017:
https://www.anandtech.com/show/11562/amd-epyc-launch-event-live-blog-starts-4pm-et-

Well I don't want to get into blue VS red debate here, but surely the retail+OEM desktop market has slightly higher volume demands than completely new server platform. Epyc sold little not because it was manufactured in insufficient volume, but because the server guys are still more evaluating than using it. Dell, HP only launch actual servers with it in Q4 etc. Basically AMD probably has them, they are just waiting for their takers.

BTW when we are on that topic, BK actually painted very similar picture about Skylake-SP/Purley ramp, too, in the conference call. Even Intel has to go through slow ramp with its new platform, despite having all the status quo on its side and actually shipping pre-launch hardware for many month, for evaluation and testing. The server market is just slow to adopt new thing.

Stacy Aaron Rasgon - Sanford C. Bernstein & Co. LLC

Hi, guys. Thanks for taking my questions. First, I wanted to ask a little bit about the revenue drivers. Revenue upside for the year and I guess marching up for three quarters, but it continues to look to be the same trend, almost entirely from PC and memory. In the meantime, this is, in your own words, your biggest server refresh in a decade, and yet we're not seeing any upside. We're doing 7% year over year pretty much for the full year, which I guess I would classify as barely high single digit. We're not really seeing meaningful ASP increases from mix as Purley ramps. I was wondering if you could just give us a view on I guess your expectations for data center, where it's coming in, why we're not seeing more upside from the Purley ramp versus what we've seen from prior platform introductions, and whether or not 7% actually would qualify as high single digits by your definition.

Brian M. Krzanich - Intel Corp.

I'll start and Bob can jump in, Stacy. First, because you've got a couple questions in there, what I would tell you is that the Purley ramp has just really started. It starts with the big cloud service providers. You saw Google talk about it a couple weeks ago. You saw Microsoft talk about it earlier this week. Those ramps are just beginning. I talked about the 200 design wins or design-ins at customers. Those are just starting to come out, so we're really early. These ramps, remember because we have to – it will go through the cloud service providers and the Tier 2 cloud service providers, then on into comms and eventually into enterprise as well. This is going to take a year or more. So I think it's early to make this judgment. The ramp is right within the range we predicted it to be from a ramp of that product

source: https://seekingalpha.com/article/4117111-intel-intc-q3-2017-results-earnings-call-transcript (view html to get around the subscription-forcing javascript)

 

[Eug]

Not according to this.(Assuming you are talking mobile)

I’m curious. How significant is the 4.5 W to 5.2 W bump for Y?

 

[jpiniero]

I’m curious. How significant is the 4.5 W to 5.2 W bump for Y?

The discussion long way back was that the increase was due to putting back in the FIVR.

 

[Yotsugi]

The discussion long way back was that the increase was due to putting back in the FIVR.

They are putting FIVR back?
Intel is weird.

 

[witeken]

They are putting FIVR back?
Intel is weird.

Not the same FIVR. It's called McIVR (with the Mc of Multi-chip) and is compatible with EMIB.

July 2015: http://wccftech.com/intel-ice-lake-return-fvir-wip/

 

[jpiniero]

Not according to this.(Assuming you are talking mobile)

That roadmap is old/obsolete I think. I think they moved up the mobile launch to CES because of Raven Ridge.

That's true, but good enough has been reached back in the Core 2 days. With Cannonlake they planned 40EUs for the base GT2, 72EU for the GT3e, and 104EU for the GT4e. Gen 11 and Icelake would have done even better. They want to move every generation from Cannonlake to Tigerlake. Clearly iGPU performance isn't something to be ignored.

With AMD now reasonably competitive on the CPU side, Intel wouldn't be competitive in IGP gaming even if it had 104 EUs. The problem is drivers more than anything - Intel can't/won't spend the resources to do all the optimizations/cheats that AMD and nVidia does for every game. At some point you need to prioritize resources, and now that Intel is a server-first company (not to mention one cutting back), something's got to give.

 

[Yotsugi]

Intel wouldn't be competitive in IGP gaming even if it had 104 EUs

Besides, there's more to GPUs than shader math.

server-first company

That almost sounds like a joke right now.

 

[raghu78]

I find it amusing that in the same post you (1) bash Intel for shipping 50,000 CPUs as a paper launch and then (2) applaud AMD with EPYC coming in Q4 2017 when it was launched in June 2017:
https://www.anandtech.com/show/11562/amd-epyc-launch-event-live-blog-starts-4pm-et-

Maybe you should learn about server qualification cycles. AMD started shipping initial volume in Q3 to server OEMs but the server platforms launch only in Q4 2017.

Sent from my SM-G935V using Tapatalk

 

[Dayman1225]

I'm sure the majority of you all have already seen this from the Coffee Lake thread - but I thought it was appropriate to post it here

 

[Yotsugi]

I'm sure the majority of you all have already seen this from the Coffee Lake thread - but I thought it was appropriate to post it here

Mid-H1 2018 risk production?
Really Intel?
So much for Custom Foundry.

 

[DrMrLordX]

We're finally seeing some dates on 10nm+. That means risk production in Q2 2019 and commercial products in Q3 2019 (probably). No specific meaning, "in definition", blah blah blah.

In other words desktop Icelake = Q3 2019, ouch? By the way, I'm interpreting the desktop products as being in the "high performance" category. So maybe Q2.5 2019? That's much better.

 

[Dayman1225]

We're finally seeing some dates on 10nm+. That means risk production in Q2 2019 and commercial products in Q3 2019 (probably). No specific meaning, "in definition", blah blah blah.

This is for their Custom Foundry clients (which there is little off), I would assume Intel Custom Foundry Partner timeline != Intel internal CPU timeline. We'll wait and see!

 

[Dayman1225]

Mid-H1 2018 risk production?
Really Intel?
So much for Custom Foundry.

I'll say it again and again - better late than never. I do wonder though. Do they also get HVM 10nm H2 2018 like BK claimed for themselves? If so it isn't horrible but not great

 

[Yotsugi]

I'll say it again and again - better late than never. I do wonder though. Do they also get HVM 10nm H2 2018 like BK claimed for themselves? If so it isn't horrible but not great

It's horrible because TSMC is going HVM with their 7nm SoC node like half a year earlier.

That's not how you get design wins.

 

[mikk]

By the way, I'm interpreting the desktop products as being in the "high performance" category. So maybe Q2.5 2019? That's much better.

Icelake-S is for Hero/Mainstream. No wonder Intel is launching 8C CFL in H2 2018.

 

[DrMrLordX]

This is for their Custom Foundry clients (which there is little off), I would assume Intel Custom Foundry Partner timeline != Intel internal CPU timeline. We'll wait and see!

True, but still . . . well we'll see. I haven't seen any dates on 10nm+ anywhere else outside general speculation that Icelake desktop is due sometime in 2019.

Icelake-S is for Hero/Mainstream. No wonder Intel is launching 8C CFL in H2 2018.

Well if that's the case then I guess that puts Icelake-S in Aug/Sept 2019?