Cascade Lake-X on 14nm++?

[Machinus]

I have read contradictory interpretations of Intel's 2018 processor roadmap, and I think some of the authors on these tech sites are really confused and using incorrect terminology.

I am hoping the astute staff/readers at Anandtech can clarify this.

Intel has advertised that they will be releasing additional chips called "Cascade Lake-X" at the end of 2018, but has not confirmed what this means.

The previous set of "-X" chips were produced on Skylake-X, which was the physical process called "14nm" by Intel, and compatible with the X299 chipset. Skylake is also "14nm," and uses the 100-series chipset.

However, the 14nm node is possibly the most unusual process that Intel has used in decades. There are currently four different product lines being sold under this process (the standard was two for a long time), and the market segments and capabilities of the products are not clearly delineated. So, despite Skylake being the second set of chips sold around 14nm (remember Broadwell? that was the first "14nm."), there are two subsequent processes that further HEDT refinements could be produced on: Kaby Lake, "14nm+", and Coffee Lake, "14nm++".

What is confusing to me is why Intel has not named the 2018 X-series additions to either the first or second optimization of 14nm. Does this mean that it's potentially a further semi-optimization of 14nm, but only used for the X-series? Or, a possibility that I think is more likely, are they just going to acknowledge the massively prolonged cycle at 14nm, skip any more products at 14nm+, and produce these new X-series chips at 14nm++?

If Cascade Lake-X is going to be produced at 14nm++, this would make the chips directly comparable to the Coffee Lake products released last year, in the same way that Skylake-X was comparable to Skylake, except in this case the X-series chips would have a different product name for some reason. Is this true? Will Cascade Lake-X be the HEDT version of Coffee Lake? Why have they given it a different name?

 

[xblax]

My guess is that Cascade Lake-X will be an updated uArch on 14nm++(+). Intel stated that 10nm will have lower clocks and higher power consumption initially, that means 10nm will definitely not be ready for mass production of competitive chips at the end of 2018. Also the low yield would make the production of big HEDT dies uneconomic if not impossible. Maybe Cascade Lake is a backport of Cannon Lake or Ice Lake to 14nm - they can't release refreshes forever if their 10nm process is garbage.

I would also speculate that the uArch of Cascae Lake-X will later be used for a successor of Coffee Lake-S in early 2019 on 14nm++(+). Maybe Intel knows how two fix their 10nm(+) process for an Ice Lake-S release. But I'm pretty sure they are now preparing a plan B to compete against Zen 2 in 2019 if 10nm turns out to be a failure.

 

[jpiniero]

Will Cascade Lake-X be the HEDT version of Coffee Lake? Why have they given it a different name?

It's pretty much a 14++ version of Skylake Server. They are making a few other changes; mainly support for Optane DIMMs (and I imagine now performance fixes for Meltdown and Spectre) but not much else. The different dies will likely have the same core counts too given how big the dies are.

 

[Machinus]

It's pretty much a 14++ version of Skylake Server. They are making a few other changes; mainly support for Optane DIMMs (and I imagine now performance fixes for Meltdown and Spectre) but not much else. The different dies will likely have the same core counts too given how big the dies are.

Do you have a reference for that?

Also, "Skylake Server" consists of almost 100 chips with different specifications. A HEDT line has fewer than 10 different configurations. Maybe your source can clarify what direction they are going?

 

[Justinus]

Skylake-X is already made on Kaby Lake's 14+, so its got to be on 14++ if it's not on a smaller node.

https://en.wikichip.org/wiki/intel/cores/skylake_x

It was my understanding it would either be 14++ or a third optimization specifically for it.

 

[Machinus]

Skylake-X is already made on Kaby Lake's 14+, so its got to be on 14++ if it's not on a smaller node.

https://en.wikichip.org/wiki/intel/cores/skylake_x

It was my understanding it would either be 14++ or a third optimization specifically for it.

It looks like you're right, that Skylake-X was on 14nm+. Cascade Lake-X will be released about 1 year after, which suggests to me that they will be on 14nm++ since Intel is under a lot of pressure to release 10nm.

With as much success as Coffee Lake had as a brand, I am surprised Intel did not want to release these chips in that line. Maybe the chipset being sold at the end of this year will be new?

 

[dark zero]

Seems that Intel's 14 nm is their version of AMD and nVIDIA 28 nm.... a total nightmare for them

 

[DrMrLordX]

Intel can't/won't bring Coffeelake to x299 since they have committed to using the distinct mesh-based interconnects on that platform. They also want the modified cache configuration of Skylake-X on x299. So it makes sense for them to respin Skylake-X as Cascade Lake-X on 14nm++, instead of simply trying to port Coffeelake over somehow.

 

[csbin]

Cascade Lake has 28 cores = Skylake-sp Refresh

https://twitter.com/TMFChipFool/status/937245082561974274

 

[Machinus]

Cascade Lake has 28 cores = Skylake-sp Refresh

https://twitter.com/TMFChipFool/status/937245082561974274

That does seem to confirm that they are basing Cascade Lake-X on Skylake server chips. So maybe this will be on a new 300-series chipset?

One interesting thing to note is that the Intel slide does not say "Cascade Lake-X," but just "Cascade Lake." Does that mean they might release new S-series chips as well?

 

[DrMrLordX]

I do not think we will see Cascade Lake-S on consumer desktop (if that is what you meant). Cascade Lake will be for Xeons and LGA2066.

 

[Topweasel]

Skylake-X/SP already was a large departure from the arch of Skylake, Kaby and Coffee Lake contain almost no arch differences. With Cascade Lake and on I think the idea is the consumer and enterprise core architectures will branch off on their own. Certainly it wouldn't make sense to name it after Kaby lake or Coffee lake since those are unchanged arch's. Maybe they will merge again in the future as things like the ring bus might hold them back when competing with future versions of Ryzen with 12+ cores for consumers.

 

[jpiniero]

Skylake-X/SP already was a large departure from the arch of Skylake, Kaby and Coffee Lake contain almost no arch differences. With Cascade Lake and on I think the idea is the consumer and enterprise core architectures will branch off on their own. Certainly it wouldn't make sense to name it after Kaby lake or Coffee lake since those are unchanged arch's. Maybe they will merge again in the future as things like the ring bus might hold them back when competing with future versions of Ryzen with 12+ cores for consumers.

I do think Icelake mainstream has the mesh (but not the extra L2 or AVX-512 unit) but until we get more info can't say for sure.

 

[Machinus]

Skylake-X/SP already was a large departure from the arch of Skylake, Kaby and Coffee Lake contain almost no arch differences. With Cascade Lake and on I think the idea is the consumer and enterprise core architectures will branch off on their own. Certainly it wouldn't make sense to name it after Kaby lake or Coffee lake since those are unchanged arch's. Maybe they will merge again in the future as things like the ring bus might hold them back when competing with future versions of Ryzen with 12+ cores for consumers.

I hope that a revision of X299 will be released if Intel plans on selling additional chips for it. It will not compare well with the 300-series that will be available at the end of the year.

 

[NostaSeronx]

Cascade Lake is the Core-version of Xeon Phi. It will be on 14nm# not 14nm++. It will have MCDRAM as well.

Whiskey Lake(AVX256) and Cascade Lake(AVX512) will use the same converged core on 14nm#.

14nm+++ will be used only for IoT oriented Coffee Lake and Gemini Lake respins.

 

[DrMrLordX]

Cascade Lake is the Core-version of Xeon Phi(which has been integrated inot. It will be on 14nm# not 14nm++. It will have MCDRAM as well.

Whiskey Lake(AVX256) and Cascade Lake(AVX512) will use the same converged core on 14nm#.

14nm+++ will be used only for IoT oriented Coffee Lake and Gemini Lake respins.

What the hell.

 

[NostaSeronx]

What the hell.

Cascade Lake really did start out being Core-based Xeon Phi...
"Mapped out micro-architecture changes required to make Knights Landing memory controller compatible with Cascade Lake processor."

14nm# is essentially a push for "Smart/NeuralNet" computational lithography. It does track(8T/7.5T/6T/5T/Rotated-track), fin(4/3/2/1-Fin), BEOL(1x/1.1x/1.5x/2x/2.2x/etc), LVT/RVT/HVT/LLVT, etc selection much faster. (LLVT is a 22FFL pulldown, there is a lot more pull downs from 22FFL fyi.)

 

[DrMrLordX]

I think you have things backwards.

Phi - what's left of it - is moving towards Core and away from Atom.

 

[NostaSeronx]

I think you have things backwards.

Phi - what's left of it - is moving towards Core and away from Atom.

I have it straight.

Knights Landing/Mill are both Silvermont-derived, while Cascade Lake is Skylake-derived. Originally, Cascade Lake would have replaced KNC/KNM but that was dropped.

What you don't know is that Xeon Phi is hopping over Skylake-derived cores to Sunnymont-derived cores. In which, the Sunnymont cores would be much wider than their standard CPU counterparts. (Sunnymont-Phi would have more physical core clusters per module.)

 

[jpiniero]

I dunno, I kind of think it's more likely that Intel is just going to take an (Icelake) core and add an additional 2 AVX-512 units and use that as the Phi.

 

[DrMrLordX]

I dunno, I kind of think it's more likely that Intel is just going to take an (Icelake) core and add an additional 2 AVX-512 units and use that as the Phi.

That's what everyone who is not NostaSeronx is saying.

But hey, he's got a reputation to maintain . . .

 

[Bouowmx]

I hope that a revision of X299 will be released if Intel plans on selling additional chips

Additional chips like? All even-number core counts between 6 and 18 filled.

2 and 4: Hopefully, something dumb like Kaby Lake-X will never see release again.
20 and above: requires XCC silicon. Would such an expensive processor make sense in a platform with just 4-channel, non-ECC memory?

 

[Machinus]

Additional chips like? All even-number core counts between 6 and 18 filled.

Intel can sell more than one chip with a certain number of cores. See: every previous product generation.

 

[eek2121]

I am aware, however I think we can all agree that today’s Intel 7 is quite a bit more efficient than the original 10nm process Intel struggled to roll out.

Clearly improvements have been made.