Intel Skylake / Kaby Lake

[John Carmack]

It's a bit sad lookin at today's Intel press release. The answer to competition appears to be segment harder.

 

[DrMrLordX]

How is it "trying to bury Power9" by having barely half the threads per chip? Honest question. 52 vs 96...

Only 6 channels of DDR4 vs P9's 8, too.

I don't see that the mem channels will make a big difference but we'll see.

To address the former question, I don't think Intel has come that close to POWER-levels of threads per socket since . . . ever? POWER8 could get you 96 threads per socket as it was, while Haswell-EX capped out at . . . 36 I think?

The traditional x86 solution to challenging old-school "Big Iron" was just to sell more boards with more sockets in a fashion that was cost-effective compared to the big boys. Now we have Intel trying to cram 26 cores (with HT) into a single socket. They want to fight IBM on their own turf, and win doing so.

 

[raghu78]

So Intel has found a new metric to market . Last few generations it was Minimum Metal Pitch x Contacted Gate Pitch. Now its transistors / sq mm using an average density of a small and a large logic cell.

At 10nm, Intel will pack 100.8 million transistors per square millimeter. It estimated 10nm foundry processes now in production from TSMC and Samsung have about half that density.

Intel’s metric averages density of a small and a large logic cell. Specifically, it uses a two-input NAND cell with two active gates and a scan flip-flop cell with as many as 25 active gates.

“I think it’s a comprehensive, quantitative and honest metric,” said Mark Bohr, a senior fellow and director of process architecture and integration. “TSMC and, I think, Samsung used to quote it, but my guess is they weren’t looking very good with this metric anymore,” he said.

Anyway going by the past metric of minimum metal pitch x contacted gate pitch Intel 10nm achieves a 46.6% area reduction . 36 x 54 / 52 x 70 = 53.4% . I think TSMC has confirmed a 40nm minimum metal pitch at 7nm though they are yet to confirm contacted gate pitch. TSMC 16FF+ density was 90 x 64 = 5760 sq mm. Still we saw what Apple achieved with A9 and A10 Fusion in terms of transistor density in actual shipping product. With TSMC 7nm in 2018 and 7nm+ in 2019 I think its going to be even better. The problem for Intel is all of their technological lead in transistor density never meant a damn when they tried to make mobile Atom chips much better than the ARM competition. Anyway I am keen to see what the foundries come up with in 2018. TSMC will be first out of the gate with 7nm volume production ramping in Q1 or Q2 2018 (risk production is starting in early Q2 2017). GF 7nm is scheduled for a ramp by late 2018. Samsung has also said they will have 7nm ramping in 2018 but the exact timeline is not known. Intel 10nm will have products out roughly 2 quarters before TSMC 7nm. I think TSMC will lead Samsung and GF by anywhere from 6-12 months wrt 7nm time to market.

I am waiting to see how Cannonlake Core M vs Apple A12 compare in the <5w range. I think we are going to see 3+ Ghz smartphone CPUs in 2018. Just slightly over a decade after Core 2 . Amazing.

 

[Azuma Hazuki]

Those Kaby Xeons are impressively low TDP. I'd still rather have a Ryzen for the price, but Intel's server gear is solid in its own right.

 

[Ajay]

KBL-R and CFL are 14nm++, SKX is 14nm+. I expect a "SKX refresh" on 14nm++, Intel said they're committing to annual product refresh for all its product lines.

So, if there is a SKX refresh, then Intel will refresh it's server line on a buffed up node?

 

[Ajay]

More PDF from Intel's manufacturing Day can be found here-> https://www.semiwiki.com/forum/f293/intels-manufacturing-day-materials-9145.html

Still trying to figure out what exactly what "Hyper scaling" is for 10nm. Intel continues using FinFET with ever taller fins. Intel is going with SAQP as well as SADP for 10nm. Looks like frequency will take a hit at first on 10nm and won't match 14++ till 10nm+. Looks like Intel really needs 450mm wafers to bring down $/xtor costs.

 

[MarkizSchnitzel]

The problem for Intel is all of their technological lead in transistor density never meant a damn when they tried to make mobile Atom chips much better than the ARM competition.

Wasn't the opinion here that segmentation was the cause for this, they did not want to eat up in big core sales?

 

[coercitiv]

Wasn't the opinion here that segmentation was the cause for this, they did not want to eat up in big core sales?

Segmentation hurt Atom in some areas, but not in their mobile efforts. Intel simply did too little too slow, and neither their superior process or their very deep pockets were enough to compensate.

 

[nvgpu]

https://newsroom.intel.com/newsroom...3/Murthy-Renduchintala-2017-Manufacturing.pdf

Page 9 confirms future Intel 300 series chipset is built on 14nm process, but no idea whether it's 14+ or 14++.

Page 12 shows the future of Intel CPUs, built from various processes and all connected through EMIB.

 

[mikk]

Good find. Isn't Intels current chipset still made on 32nm?

 

[witeken]

Good find. Isn't Intels current chipset still made on 32nm?

I thought Brfoadwell-Y was 22nm, but upon further thinking and checking at AnandTech, it's actually 32nm indeed. Don't know about SKL.

 

[Bouowmx]

You can look in ARK database: https://ark.intel.com/products/98089/Intel-Z270-Chipset

22 nm

 

[raghu78]

Another important bit of information. Intel first gen 10nm is only slightly better than 14nm+ for transistor performance and is lower than Intel 14nm++ . In fact even Intel 10nm+ is not better than 14nm++. It will take 10nm++ to beat 14nm++ for transistor performance. (slide 29)

https://newsroom.intel.com/newsroom.../2017/03/Kaizad-Mistry-2017-Manufacturing.pdf

That does not bode well for high performance desktop CPUs from Intel based on 10nm. I think Coffeelake will probably be the fastest CPU core till 2020 in terms of clocks and single thread performance unless Icelake on 10nm+ can bring a large IPC gain to offset the lower clocks from a significantly less mature process with lower transistor performance than 14nm++.

Another interesting bit of information is Intel did not have yield information even for SRAM which they usually do while talking about a new process. Could this mean the yield problem rumours are probably true ?

 

[Sweepr]

10nm+ is only slightly below 14nm++ in the transistor performance graph, actually close. Ice Lake should be able to compensate this with higher IPC. Bodes well for 2018 Coffee Lake though, 14nm++ is in better shape than 10nm for desktops. Everybody wants the new 6C mainstream parts to be agressivelly clocked, but I'm also curious how far they can push the 4C - considering 7700K already Turbo's to 4.4GHz (all cores).

 

[thepaleobiker]

Another important bit of information. Intel first gen 10nm is only slightly better than 14nm+ for transistor performance and is lower than Intel 14nm++ . In fact even Intel 10nm+ is not better than 14nm++. It will take 10nm++ to beat 14nm++ for transistor performance. (slide 29)

https://newsroom.intel.com/newsroom.../2017/03/Kaizad-Mistry-2017-Manufacturing.pdf

That does not bode well for high performance desktop CPUs from Intel based on 10nm. I think Coffeelake will probably be the fastest CPU core till 2020 in terms of clocks and single thread performance unless Icelake on 10nm+ can bring a large IPC gain to offset the lower clocks from a significantly less mature process with lower transistor performance than 14nm++.

Another interesting bit of information is Intel did not have yield information even for SRAM which they usually do while talking about a new process. Could this mean the yield problem rumours are probably true ?

Brings me back to my question from a month ago : Would Coffeelake 14nm++ chips fit into the 200 series LGA 1151 socket? :/ ?

 

[nvgpu]

If it's LGA 1151 probably yes, but you'll need a UEFI BIOS update and that can only be done with a Kaby Lake or Skylake CPU before swapping out to a Coffee Lake CPU.

200 series chipset also does not have the new features of the 300 series chipset like native USB 3.1 Gen2 10Gbps support.

 

[Bouowmx]

Socket change for Coffee Lake: no, but still not confirmed whether 100 and 200 series chipsets will be forwards-compatible with Coffee Lake.

 

[Arachnotronic]

Another important bit of information. Intel first gen 10nm is only slightly better than 14nm+ for transistor performance and is lower than Intel 14nm++ . In fact even Intel 10nm+ is not better than 14nm++. It will take 10nm++ to beat 14nm++ for transistor performance. (slide 29)

https://newsroom.intel.com/newsroom.../2017/03/Kaizad-Mistry-2017-Manufacturing.pdf

Yep.

That does not bode well for high performance desktop CPUs from Intel based on 10nm. I think Coffeelake will probably be the fastest CPU core till 2020 in terms of clocks and single thread performance unless Icelake on 10nm+ can bring a large IPC gain to offset the lower clocks from a significantly less mature process with lower transistor performance than 14nm++.

Do you really think that Intel will release an Ice Lake that is slower than Coffee Lake? Come on, man.

Another interesting bit of information is Intel did not have yield information even for SRAM which they usually do while talking about a new process. Could this mean the yield problem rumours are probably true ?

Of course yields are bad, if they were good Intel would be talking about them.

 

[scannall]

Segmentation hurt Atom in some areas, but not in their mobile efforts. Intel simply did too little too slow, and neither their superior process or their very deep pockets were enough to compensate.

Not to mention, trying to drag the mobile market to x86, locking them in to a single vendor didn't help either. I'm sure Intel has more than enough talent to make a terrific ARM mobile SOC. But, then they have to compete.

 

[Arachnotronic]

So, if there is a SKX refresh, then Intel will refresh it's server line on a buffed up node?

Well, yeah. Skylake-X is just a LCC die of Skylake Xeon.

 

[raghu78]

Arachnotronic Kabylake is already clocking 5 Ghz . Coffeelake should do better. Do you think Coffeelake on 14nm++ can be beaten by Icelake on 10nm+ for max single thread performance. First of all we have to wait and see how close to 5 Ghz can Icelake get to. If they can't do 4.7-4.8 Ghz consistently on Icelake then I think their chances against a 5.2 Ghz Coffeelake (I am assuming only 200 Mhz higher OC for Coffeelake which could be higher) are not so great. 14nm++ has a 26% higher transistor performance over 14nm and roughly 10-12% higher performance than 14nm+. The first gen Skylake maxed out at 4.7-4.8 Ghz.

http://www.anandtech.com/show/9533/intel-i7-6700k-overclocking-4-8-ghz
https://www.hardocp.com/article/2015/08/05/intel_skylake_core_i76700k_ipc_overclocking_review/7

I think adding 8-10% IPC over Coffeelake is not going to be easy. These large cores are already damn good so every extra % of IPC increase comes at a much higher and non linear power increase. Coffeelake could turn out extremely difficult to beat for single thread performance even by Icelake.

 

[Suspicious-Teach8788]

Anyone familiar with Gigabyte Motherboards (Z270) and how to set the Vcore appropriately? Obviously there's manual settings but how would I go about setting an "Adaptive" (that's an ASUS term I believe) voltage for normal use with cstates enabled? Would I use the Normal + DVID feature? If any overclocker here has a Gigabyte Board that would be very helpful.

 

[mikk]

https://newsroom.intel.com/newsroom.../2017/03/Kaizad-Mistry-2017-Manufacturing.pdf

That does not bode well for high performance desktop CPUs from Intel based on 10nm. I think Coffeelake will probably be the fastest CPU core till 2020 in terms of clocks and single thread performance unless Icelake on 10nm+ can bring a large IPC gain to offset the lower clocks from a significantly less mature process with lower transistor performance than 14nm++.

14nm+ clocks very well based on Kabylake and 10nm+ does even better, not sure what the problem is. First version of 22nm didn't clock as good as 32nm, first 14nm was worse than matured 22nm and so on. Also you don't know clock speeds of Coffee Lake. Even if there is another 200-300 Mhz speedbump from 14nm++ (which I doubt), it would only add a 5% faster clock speed compared to 14nm+ and 1-2% to 10nm+ based on the performance graph and the fastest Kabylake.

Coffe Lake comes with 6 cores, not sure if it can beat 4.5 Ghz stock clock from Kabylake. It's not that easy.

 

[Sweepr]

Intel just leaked something interesting...

Skylake-SP Die Shot

Confirms 28C for the HCC die, monolithic monster.

 

[DrMrLordX]

Do you really think that Intel will release an Ice Lake that is slower than Coffee Lake? Come on, man.

The last time Intel did such a thing was Willamette. Well sort of, due to the clockspeed restrictions on the Pentium IIIs of the day (pre-Tualatin), a 1.8 GHz Willamette was still a winner. But against a 1.4 GHz Tualatin . . .