14nm 6th Time Over: Intel Readies 10-core "Comet Lake" Die to Preempt "Zen 2" AM4

[Despoiler]

We (the consumers) don't need any more cores, at least for the time being...

I see this as an "if you built it they will come" situation. We don't "need" more cores right now because Intel has been in the driving seat with fast and narrow mainstream chips. Now that AMD took over we have slightly slower and much wider. Few devs have thought about what to do with all of the cores. Or perhaps it's more accurate that they haven't tied a specific benefit to having excess cores. Something like a top AI mode only available to a certain core count.

 

[DrMrLordX]

Damn this is goofy. Ice Lake-S may never happen at this rate. If AMD releases Matisse in a timely fashion, Intel is going to get wrecked. Comet Lake-S will not be suitable competition.

Never thought I'd say it, but Intel should have listened to Piednol!

 

[maddie]

Damn this is goofy. Ice Lake-S may never happen at this rate. If AMD releases Matisse in a timely fashion, Intel is going to get wrecked. Comet Lake-S will not be suitable competition.

Never thought I'd say it, but Intel should have listened to Piednol!

Remind me.

 

[jpiniero]

Damn this is goofy. Ice Lake-S may never happen at this rate. If AMD releases Matisse in a timely fashion, Intel is going to get wrecked. Comet Lake-S will not be suitable competition.

Never thought I'd say it, but Intel should have listened to Piednol!

Comet Lake-S would be faster than a theoretical Icelake-S in games because of the much higher clock speeds. We'll have to see what Matisse can clock to but it's pretty much the same deal.

 

[DrMrLordX]

Remind me.

Francois Piednol lobbied for Intel backporting IceLake to 14nm. He failed. Now he's gone (from Intel).

Comet Lake-S would be faster than a theoretical Icelake-S in games because of the much higher clock speeds. We'll have to see what Matisse can clock to but it's pretty much the same deal.

Remember that the expected clockspeed deficits between 14nm CoffeeLake/CoffeeLake Refresh/etc. and 10nm+ IceLake were due to the performance characteristics of 10nm/10nm+. On 14nm, I have every reason to believe that Intel could have achieved similar clocks with IceLake.

 

[crashtech]

Seems like the easiest route would be to adapt a mesh 10-core to dual channel, unless there is something specific about LGA 1151 that would prevent it.

 

[NTMBK]

Comet Lake-S would be faster than a theoretical Icelake-S in games because of the much higher clock speeds. We'll have to see what Matisse can clock to but it's pretty much the same deal.

Oh yeah sure, the 10 core part in a 95W socket is going to be a real clockspeed monster.

 

[NTMBK]

Seems like the easiest route would be to adapt a mesh 10-core to dual channel, unless there is something specific about LGA 1151 that would prevent it.

Well it would be missing an IGP and other stuff you'd expect to get integrated on LGA 1151, but they could probably make it work. After all they did it the other way around (Kabylake X).

 

[jpiniero]

Oh yeah sure, the 10 core part in a 95W socket is going to be a real clockspeed monster.

They could change the socket you know... Besides you don't necessairly have to have all 10 cores running at 5.2+ Ghz.

Remember that the expected clockspeed deficits between 14nm CoffeeLake/CoffeeLake Refresh/etc. and 10nm+ IceLake were due to the performance characteristics of 10nm/10nm+. On 14nm, I have every reason to believe that Intel could have achieved similar clocks with IceLake.

Possible, but the original Icelake CPU core was a marginal improvement at best besides AVX-512 support. And the new one (which is essientally Tigerlake's CPU core) is probally quite a bit bigger relatively speaking.

 

[Arachnotronic]

Possible, but the original Icelake CPU core was a marginal improvement at best besides AVX-512 support. And the new one (which is essientally Tigerlake's CPU core) is probally quite a bit bigger relatively speaking.

How do you know that "the original Icelake CPU core was a marginal improvement at best besides AVX-512 support"?

 

[jpiniero]

How do you know that "the original Icelake CPU core was a marginal improvement at best besides AVX-512 support"?

You can kind of tell. Hell, the fact that there weren't any rumors about Icelake Desktop coming (on 10 nm) should be enough - they figured the same as what I was saying, that the clock speed hit wasn't going to be worth the small IPC gain.

 

[Arachnotronic]

You can kind of tell. Hell, the fact that there weren't any rumors about Icelake Desktop coming (on 10 nm) should be enough - they figured the same as what I was saying, that the clock speed hit wasn't going to be worth the small IPC gain.

So just your assumption then.

 

[EXCellR8]

Comet Lake = 300 series + flash or new chipset?

 

[jpiniero]

Comet Lake = 300 series + flash or new chipset?

If they do end up using a new socket, they could use the 400 series chipset. Main new feature is Thunderbolt 3 on the PCH.

 

[BigDaveX]

Seems like the easiest route would be to adapt a mesh 10-core to dual channel, unless there is something specific about LGA 1151 that would prevent it.

It might be technically possible to stick an LCC Skylake-X die on an LGA 1151 package unless it were physically too big, but the biggest problems are that Skylake-X tends to use a fair bit more power than its consumer counterparts, and the mesh can really do a number on single-threaded performance in certain situations. Unless Intel really tweaked the design, a 10C Skylake-X derivative would almost certainly end up performing worse in a lot of situations compared to the 8C Coffee Lake.

 

[jpiniero]

They don't have to go to the mesh at 10 cores if they don't want to and that would be a lot more work. Broadwell-E XCC was 12 cores per ring.

 

[DrMrLordX]

You can kind of tell. Hell, the fact that there weren't any rumors about Icelake Desktop coming (on 10 nm) should be enough - they figured the same as what I was saying, that the clock speed hit wasn't going to be worth the small IPC gain.

Actually if you look at the misbegotten Cannonlake lappie chips that have already come out, you can see a few IPC improvements on those. Maybe 3-5%, depending on the workload. I think Ice Lake was rumoured to be in the ballpark of +5% over the Skylake family? It might have been closer to 10-15% assuming it was/is an iterative improvement over Cannonlake.

 

[Charlie22911]

It would make sense if intel planned to shift the pricing of its product stack down to put this where the 9900k is now. But they won’t, they’ll put it on top and price it outrageously, and AMD will continue to erode their market share with their superior price-performance offerings.

 

[JTaylor2005]

Doesn't seem to me as though Forbes' theory would hold up. Surely one stupidly expensive CPU doesn't class as a "Zen 2 Killer" The annoying thing is it's websites like these that investors believe the hype from!

https://www.forbes.com/sites/antony...zen-zen-2-killer-in-development/#3193929c3c43

 

[Vattila]

If this is true, you can look at it as Intel optimising their product range as best as they can.

Remember that 10-core is not a new offering. We already have 7900X Skylake-X for the HEDT platform. Intel's problem is that they were never prepared for the HEDT segment to approach the mainstream segment in pricing. Also, the workloads targeted by their respective platforms differ starkly. Their mainstream processors play games better, while their HEDT processors do better on parallel workloads.

This is how PC enthusiasts have been thinking about HEDT and mainstream — as two very different platforms.

But this is now changing, as I've speculated (here). AMD's HEDT platform is becoming affordable, with pricing reaching down to the upper part of mainstream. Also, due to advances in boost technology, their 16-core HEDT processor has higher single-thread performance than their 8-core mainstream counterpart (in addition to more cache and memory bandwidth, 2950X has boost at 4.4 GHz, vs 2700X at 4.3 GHz), which means there is no longer a stark difference in gaming performance. Consequently, AMD now has a pretty continuous range of desktop offerings from 4-core budget mainstream to 32-core workstation-grade HEDT. Zen 2 will likely extend core counts and performance even further, driving down pricing for the last generation, and making the AMD HEDT platform even more tempting for PC enthusiasts.

This is what Intel is facing.

PS. Hence, I see no need for AMD to offer more than 8 cores for AM4. If, as expected, Zen 2 is competitive in single-thread performance, then max 8 cores plus iGPU is what makes sense in the mainstream (i.e. single CPU chiplet plus IO chiplet with iGPU). For anything beyond 8 cores, their HEDT platform is well positioned to compete.

 

[chrisjames61]

One of Otellini's HUGE mistakes was turning down Apple for making their iPhone SOC. Forcing Apple to do their own made Apple into a giant brain drain on the CPU design engineers (Have a look at their linkedin sometime), plus a massive infusion of cash into the pay for play foundries. In any management at Intel still has a lick of sense they can steal some CPU talent from Global Foundries since they faceplanted with IBM's golden gift. But they better be quick, they aren't the only one who wants/needs talent.

Yup, Apple has sunk untold billions into TSMC's coffers. That could have been Intel's. This has allowed TSMC to take away Intel's process advantage. It has also allowed ARM to make massive inroads towards someday supplanting X86 cpu's.

 

[scannall]

Yup, Apple has sunk untold billions into TSMC's coffers. That could have been Intel's. This has allowed TSMC to take away Intel's process advantage. It has also allowed ARM to make massive inroads towards someday supplanting X86 cpu's.

It goes further than that even. Apple doesn't give flying monkeys if AMD rides their coattails on the process either. It gives them a second source, and an arm lock to use on Intel if they need. Intel's failure to meet road maps has screwed Apple quite a few times. The same reason they dumped their PowerPC line. They are less infatuated with Intel than most enthusiasts.

 

[PotatoWithEarsOnSide]

It's hard to meet a roadmap if you have to change it every 3 months.
It'd be better if they set realistic timescales, instead of ones essentially dictated by the stock market.
Unfortunately, short term thinking is rewarded too greatly.

 

[dark zero]

Why this remember me to the infamous AMD loop?

Poor Intel.. They did not deserved this.

 

[epsilon84]

10nm must be really underperforming for Intel to consider this route. That or AMD is going with 12 core Ryzen 7 for Zen 2.

I understand the concerns about a 10 core 14nm and the potential for furnace like thermals, but that will only apply if Intel clocks it like 9900K with unlimited TDP constraints.

The 95W 9900K still clocks at 4.2GHz all core so a 10 core chip will be able to fit under a 120W TDP under the same conditions.