Discussion Intel current and future Lakes & Rapids thread

[DrMrLordX]

https://en.wikipedia.org/wiki/POWER9

Power 9 can be SMT4 or SMT8, 4 or 8 threads per core. Either way, the chip runs 96 threads with either 12 SMT8 cores or 24 SMT4 cores.

Right, I'd love to see a "four slice" design handling desktop tasks via SMT8. It would be interesting at least, if perhaps pointless.

 

[IntelUser2000]

What if,

management at the time of Otellini is at the fault for Intel's current problems. It's easy to blame current management for their misshaps. However because of long development time, mistakes get carried over. The world isn't a test system where you can normalize everything and start over.

Some people have said the reason Intel screwed up on 14nm and 10nm is because they pulled resources to better 22nm. And they said 22nm was a record yielding process. Otellini's team went out on a good note because they were at the peak. But that peak was decided years in advance.

In the early 2000s, the hindsight was that they believed clock speed scaling will go to the moon. The problems they are having right now are way bigger than that because its a process related issue - more accurately, its the increased difficulty due to reaching physics limits. Do they have a "Banias" to this problem or not?

 

[DrMrLordX]

I dunno that that's really true. Anyone who worked on Intel's 22nm process knew that clockspeed scaling would not go to the moon. They put a lot of work into that node making it more mobile-friendly which is why we got clockspeed regression moving from Intel's 32nm to 22nm (5 GHz on Sandy to ~4.5 GHz on Haswell).

14nm was initially more of the same.

 

[Dayman1225]

http://electroiq.com/chipworks_real_chips_blog/2017/12/01/iedm-2017-next-week-part-2/

Just a reminder that Intel will be detailing/talking about their 10nm and 22nm FFL nodes on the 6th December @IEDM. Glofo will also be talking about their 7nm process.

 

[IntelUser2000]

I dunno that that's really true. Anyone who worked on Intel's 22nm process knew that clockspeed scaling would not go to the moon. They put a lot of work into that node making it more mobile-friendly which is why we got clockspeed regression moving from Intel's 32nm to 22nm (5 GHz on Sandy to ~4.5 GHz on Haswell).

14nm was initially more of the same.

The reason I mentioned about Netburst was because it parallels the problems they are having today. Of course the problem they are having today is far greater than with Netburst. History repeats, just with different actors and different words.

Intel in early 2000s, had zero idea clockspeed would stall at 5GHz. They were envisioning 20GHz at one point. Intel few years ago believed their process leadership was absolute. In both cases, its lack of a fallout plan and inability for the engineers/managers to admit they might make a mistake that likely proved to be a failure.

 

[Dayman1225]

The (Alder)lake afterwards has a Gen13 GPU

Alder Lake comes after Tiger Lake.

 

[jpiniero]

The (Alder)lake afterwards has a Gen13 GPU

Alder Lake comes after Tiger Lake.

Contining to milk Skylake is going to get old at some point, although if it is 10+++ I wonder what Intel could get out of that. I am a little surprised Intel isn't just forcing client on Sapphire Rapids regardless of it's suitability.

 

[Dayman1225]

Contining to milk Skylake is going to get old at some point, although if it is 10+++ I wonder what Intel could get out of that. I am a little surprised Intel isn't just forcing client on Sapphire Rapids regardless of it's suitability.

Milking Skylake? At that point it would be a shrunk Icelake. Not Skylake. Assuming PAO.

 

[jpiniero]

Milking Skylake? At that point it would be a shrunk Icelake. Not Skylake. Assuming PAO.

You know, the usual. Process update, minor changes to the core, update the GPU to support modern codecs, HDMI and DP updates. I guess it does sort of make sense to have something to fallback on in case Sapphire Rapids is a bust.

 

[Dayman1225]

You know, the usual. Process update, minor changes to the core, update the GPU to support modern codecs, HDMI and DP updates. I guess it does sort of make sense to have something to fallback on in case Sapphire Rapids is a bust.

Or just or, that Sapphire Rapids is server only.

 

[Arachnotronic]

Contining to milk Skylake is going to get old at some point, although if it is 10+++ I wonder what Intel could get out of that. I am a little surprised Intel isn't just forcing client on Sapphire Rapids regardless of it's suitability.

Ice Lake is not Skylake, nor is any CPU that comes after it.

 

[Ajay]

Or just or, that Sapphire Rapids is server only.

I thought SR was server only - which would explain the Alder Lake client. Isn't PAO now PAOO or something like that?

 

[IntelUser2000]

Or just or, that Sapphire Rapids is server only.

I agree.

Ice Lake is not Skylake, nor is any CPU that comes after it.

Let's see whether Ice Lake is a further expansion of Sandy Bridge or it brings new concepts to increase performance.

 

[Dayman1225]

I thought SR was server only - which would explain the Alder Lake client. Isn't PAO now PAOO or something like that?

Intel hasn't changed it officially as far as I know.

 

[IntelUser2000]

According to the Hungarian site HWSW, here's what they are saying.

https://www.hwsw.hu/hirek/57865/int...ade-lake-cannon-lake-ice-lake-tiger-lake.html

Server
Mid 2018 - Cascade Lake on 14nm++
Mid 2019 - Icelake on 10nm+
Mid 2020 - Sapphire Rapids on 10nm++

Client
Early 2018 - Cannonlake on 10nm
Early 2019 - Icelake on 10nm+
Early 2020 - Tigerlake on 10nm++
Early 2021 - Sapphire Rapids on 7nm

Since this article is dated few months ago in late September of this year, we can peg Ashraf's information as newer.

"Alder Lake" in light of this does not bode well for Intel's 7nm development. The naming of Lakes must have some reasoning behind it. Alder Lake might be yet another 10nm client CPU, possibly in Early 2021.

While the platform might change greatly in Icelake, the core might not be a fresh new one. It'll be an extension, possibly for 7-10% gain.

 

[Dayman1225]

According to the Hungarian site HWSW, here's what they are saying.

https://www.hwsw.hu/hirek/57865/int...ade-lake-cannon-lake-ice-lake-tiger-lake.html

Server
Mid 2018 - Cascade Lake on 14nm++
Mid 2019 - Icelake on 10nm+
Mid 2020 - Sapphire Rapids on 10nm++

Client
Early 2018 - Cannonlake on 10nm
Early 2019 - Icelake on 10nm+
Early 2020 - Tigerlake on 10nm++
Early 2021 - Sapphire Rapids on 7nm

Since this article is dated few months ago in late September of this year, we can peg Ashraf's information as newer.

"Alder Lake" in light of this does not bode well for Intel's 7nm development. The naming of Lakes must have some reasoning behind it. Alder Lake might be yet another 10nm client CPU, possibly in Early 2021.

While the platform might change greatly in Icelake, the core might not be a fresh new one. It'll be an extension, possibly for 7-10% gain.

Thats of course assuming their roadmap/timeline is right. I don't think Sapphire Rapids it self is coming to client and 7nm node wouldn't be a client first node as anything beyond 10nm+ is server first IIRC

 

[IntelUser2000]

Thats of course assuming their roadmap/timeline is right. I don't think Sapphire Rapids it self is coming to client and 7nm node wouldn't be a client first node as anything beyond 10nm+ is server first IIRC

The whole server-first may mean implementation is done on server, unlike in the past where it was done first on client. But due to validation time it may not mean it actually arrives earlier. I mean, look at their Iris Pro devices. It takes nearly a year for availability. Or Xeon D.

Before we had server cores coming substantially later, a year or more sometimes. By just bringing release dates to be on a roughly equal footing qualifies for server-first. That makes sense to me. Ice Lake is likely early 2019. Sapphire Rapids on server is mid 2020 at the earliest. Unless Tigerlake is still on 10nm+ or significantly delayed to early 2021 or something how can server arrive first?

 

[Arachnotronic]

The whole server-first may mean implementation is done on server, unlike in the past where it was done first on client. But due to validation time it may not mean it actually arrives earlier. I mean, look at their Iris Pro devices. It takes nearly a year for availability. Or Xeon D.

Before we had server cores coming substantially later, a year or more sometimes. By just bringing release dates to be on a roughly equal footing qualifies for server-first. That makes sense to me. Ice Lake is likely early 2019. Sapphire Rapids on server is mid 2020 at the earliest. Unless Tigerlake is still on 10nm+ or significantly delayed to early 2021 or something how can server arrive first?

Also keep in mind Intel's early ship which means processors get delivered to customers roughly a year before general availability. Formal launch of Sapphire Rapids in mid-2020 could mean early ship in 2H 2019.

 

[Dayman1225]

Intel, ARM deepen foundry ties

ARM announced at the event it is extending its partnership with Intel’s foundry business to build IP tailored for the x86 giant’s 22nm FinFET node. It’s a fascinating example of co-opetition from the processor rivals. ARM will deliver IP to enable a Cortex-A55 geared for midrange smartphones to run up to 2.35 GHz or down to 0.45V in its so-called 22FFL process. ARM is already helping make in Intel’s flagship 10nm process a test chip due out before the end of the year using a next-gen Cortex-A SoC running at 3.5 GHz or 0.5V and delivering 0.25mW/MHz. An Intel executive showed road maps first released at a Beijing event in September, giving more details than ever on its foundry and IP plans. Observers said the level of candor was new for Intel but key to establishing a foundry business it has been struggling—largely with its own culture--to stand up for years.

I know I posted these two images before but I never posted the rest so here.

 

[Dayman1225]

https://www.intc.com/investor-relat...017/Nasdaq-37th-Investor-Program/default.aspx

Intel has an investors conference on the 5th hosted by Murthy Renduchintala

 

[jpiniero]

Also keep in mind Intel's early ship which means processors get delivered to customers roughly a year before general availability. Formal launch of Sapphire Rapids in mid-2020 could mean early ship in 2H 2019.

Plausible but sounded more like general availability to me. Maybe Icelake Server would just have a short window before the release of SR 10++. Given that it sounds like Icelake Server is monolithic margins are going to take a big hit so getting SR out ASAP has got to be a priority. But yet they can't let AMD have Rome out alone for too long.

 

[Yotsugi]

getting SR out ASAP

Rushing out server platform is a bad idea, especially given Intel's recent spotty track record (dying Avoton Atoms anyone?).

But yet they can't let AMD have Rome out alone for too long.

Easier said, than done.

 

[Dayman1225]

.........................




Don't not post the same image in three threads.


esquared
Anandtech Forum Director

 

[frozentundra123456]

dont see any indication of the 8 core CL there.

 

[IntelUser2000]

Cascade Lake has 28 cores.

Thanks to Ashraf
https://twitter.com/TMFChipFool/status/937245082561974274

So we now know Cascade Lake is basically a bug-fixed Skylake as some have speculated(CD for example). Since they are not increasing core count at all, for their sake I hope the market positioning is far better than they have now. For example, allow 28 core part to be available for $4-5K, rather than $10-13K.

The process fallout continues.