Discussion Intel current and future Lakes & Rapids thread

[jpiniero]

Kaby Lake R, Coffee Lake U and Coffee Lake H.

The Macbook uses the Y parts, ie: fanless.

 

[Arachnotronic]

Yeah I don't know what Apple is gonna do if they intend to refresh the Macbook this year.

Intel could give them a +100-200MHz Kaby Lake-Y sold under the 8th gen Core brand if Apple intends to refresh this year.

 

[Eug]

Intel could give them a +100-200MHz Kaby Lake-Y sold under the 8th gen Core brand if Apple intends to refresh this year.

It seems like a 200 MHz boost would be a pretty significant boost, even if just 14 nm Kaby Lake again. Is that feasible? Also remember Apple is already using TDP up for the fanless MacBooks. How far would they go? I'd be inclined to believe that Apple wouldn't pay extra for custom chips in this context and would prefer to wait ~18 months instead of forcing an update at 1 year... meaning they could wait as long as until 2019 Q1 if absolutely necessary.

BTW, just for everyone's information, here are the previous Apple MacBook releases:

2015:
1.1 GHz M-5Y31 (Normal 4.5 W is 0.9 GHz. TDP up at 6 W is 1.1 GHz)
1.2 GHz M-5Y51 (Normal 4.5 W is 1.1 GHz. TDP up at 6 W is 1.3 GHz)
1.3 GHz M-5Y71 (Normal 4.5 W is 1.2 GHz. TDP up at 6 W is 1.4 GHz)

2016:
1.1 GHz m3-6Y30 (Normal 4.5 W is 0.9 GHz. TDP up at 7 W is ? GHz.)
1.2 GHz m5-6Y54 (Normal 4.5 W is 1.1 GHz. TDP up at 7 W is ? GHz.)
1.3 GHz m7-6Y75 (Normal 4.5 W is 1.2 GHz. TDP up at 7 W is ? GHz.)

2017:
1.2 GHz m3-7Y32 (Normal 4.5 W is 1.1 GHz. TDP up at 7 W is 1.6 GHz.)
1.3 GHz i5-7Y54 (Normal 4.5 W is 1.2 GHz. TDP up at 7 W is 1.6 GHz.)
1.4 GHz i7-7Y75 (Normal 4.5 W is 1.3 GHz. TDP up at 7 W is 1.6 GHz.)

I find it curious that for 7th generation Kaby Lake, all three of those CPU tiers have a TDP up of 1.6 GHz for the base frequency.

 

[Dayman1225]

The Macbook uses the Y parts, ie: fanless.

Oh? I thought they also made use of U parts and I believe the MBP uses H parts.

 

[jpiniero]

Oh? I thought they also made use of U parts and I believe the MBP uses H parts.

They do but those are other products.

 

[Dayman1225]

They do but those are other products.

Ah alright, my mistake.

 

[mikk]

https://www.youtube.com/watch?v=XppRV6tkkik

A new CFL 4+3e SKU appeared:

i7-8559U 2.7 Ghz

And some generic Icelake ES:

ICL-U 4+2 2.4 Ghz

2.4 Ghz base is quite good for such an early ES without CPUID, much better than the current 15W KBL-R SKUs which are in the range of 1.6-1.9 Ghz. Looks promising for 10nm+, although there is no Turbo enabled on this ES yet.

 

[Dayman1225]

https://www.youtube.com/watch?v=XppRV6tkkik

A new CFL 4+3e SKU appeared:

i7-8559U 2.7 Ghz

And some generic Icelake ES:

ICL-U 4+2 2.4 Ghz

2.4 Ghz base is quite good for such an early ES without CPUID, much better than the current 15W KBL-R SKUs which are in the range of 1.6-1.9 Ghz. Looks promising for 10nm+, although there is no Turbo enabled on this ES yet.

That ICL U ES is also monumentally better than what we've been seeing with Cannonlake, double the cores, more clocks and a recognised Gen 11 LP IGPU

 

[IntelUser2000]

That ICL U ES is also monumentally better than what we've been seeing with Cannonlake, double the cores, more clocks and a recognised Gen 11 LP IGPU

If Icelake U is 10% faster per clock and has 400MHz higher clocks, I'd call that a huge disappointment.

Their "small-core" based Geminilake chip may have performance in the range of Core M3-6Y30 in single thread! In multi-thread it can beat the Core i7 7Y75 by 20-25%. All at similar TDP, and inherent platform differences mean Gemini Lake devices will have noticeably better battery life. If they had the same GT2 GPU we could replace Core m3 with Goldmont Plus cores.

The chip going in the next generation Gemini Lake+ platform may have to be artificially handicapped not to kill the Core lines.

The hope is then Icelake does more than the mediocre, 10% jump. The alternative is that the small core team may be doing better so replace all Intel cores with the Goldmont Plus derivatives. Certainly, they planned for the process stagnation better than the Core team.

 

[mikk]

If Icelake U is 10% faster per clock and has 400MHz higher clocks, I'd call that a huge disappointment.

400 Mhz in this range gives 20% better performance without any IPC increase under full load.

Their "small-core" based Geminilake chip may have performance in the range of Core M3-6Y30 in single thread! In multi-thread it can beat the Core i7 7Y75 by 20-25%.

Core i7 7Y75 is a dualcore design and still based on the first mediocre 14nm generation from 2016. The big improvement comes with ICL-Y because it will be a Quadcore design.

The hope is then Icelake does more than the mediocre, 10% jump. The alternative is that the small core team may be doing better so replace all Intel cores with the Goldmont Plus derivatives. Certainly, they planned for the process stagnation better than the Core team.

We have seen mediocre IPC increases from the small cores as well in the past with Airmont, it's not certain that they can bring big IPC boosts every generation, it would be amazing if they could repeat this.

 

[jpiniero]

Their "small-core" based Geminilake chip may have performance in the range of Core M3-6Y30 in single thread! In multi-thread it can beat the Core i7 7Y75 by 20-25%.

Seems like the die size for the CPU cores is getting close though. It's almost like why bother.

 

[IntelUser2000]

Seems like the die size for the CPU cores is getting close though. It's almost like why bother.

On Gemini Lake and predecessors they are true SoCs with PCH integrated onto the die. That skews it. Increase in die size over Apollo Lake is likely due to Gen 10 display block and more I/O features like WiFi MAC integration.

You can fit 4 Goldmont cores into an area smaller than single Skylake core with 256KB L2 cache.

The big improvement comes with ICL-Y because it will be a Quadcore design.

This needs a link. Kabylake-Y can't even take full advantage of the dual core. Unless otherwise shown, I'd assume a process shrink and a new architecture would allow a 4.5W 2 core chip, but with high clocks like Kabylake U.

 

[jpiniero]

This needs a link. Kabylake-Y can't even take full advantage of the dual core. Unless otherwise shown, I'd assume a process shrink and a new architecture would allow a 4.5W 2 core chip, but with high clocks like Kabylake U.

It makes sense when the Icelake core size is going to be pretty tiny. Not much point in doing a separate die unless it also cuts the EU count back down as well, given that's what is taking up most of the space.

Base clock will obviously be very low.

 

[IntelUser2000]

It makes sense when the Icelake core size is going to be pretty tiny. Not much point in doing a separate die unless it also cuts the EU count back down as well, given that's what is taking up most of the space.

It's much easier to create a smaller die out of a base one, than creating another one that's bigger.

If the cores are made to be modular, the designers can cut cores to make them into smaller dies, like having a 2 core die out of the 4 core base design. If you want to get 6 cores out of it though, you'd have to add from the 4 core base and requires extra work.

going to be pretty tiny.

There's also nothing known about Icelake, so we can't say its much smaller. It would be 7 years from the last architecture change with genuine new features(Sandy Bridge), and an architecture from scratch takes 5-7 years so time is ripe for one.

With Tick/Tock I think Intel moved from 5 year new uarch changes with no core changes in between to ~2 years for an expansion uarch, but retaining 5-6 years for a really new architecture. That's because process gains have slowed, and frequent smaller changes address competitive landscape better.

P6 - 1996
Netburst - 2001
Merom - 2006
Sandy Bridge - 2011

With 2-year process cadence, it would have put the first 10nm chips in early 2016. Assuming it was little over 2 years, but with little 10nm troubles, we generally expected Cannonlake by late 2016. That would put Icelake in 2017.

Again, its ripe for an architecture with new ideas, not mere expansion of existing units. Of course what they really do is another story.

 

[jpiniero]

The architecture with new ideas is Sapphire Rapids.

One thing I was thinking could happen would be to increase the AVX units to dual 512-bit from dual 256-bit. That would take up some space. No extra L2 I imagine though.

 

[IntelUser2000]

The architecture with new ideas is Sapphire Rapids.

One thing I was thinking could happen would be to increase the AVX units to dual 512-bit from dual 256-bit. That would take up some space. No extra L2 I imagine though.

Nah. Ashraf says Icelake - Tiger Lake - Alder Lake
For servers its Icelake - Sapphire Rapids - Granite Rapids

AVX-512 units for the relevant parts are done by 2x256 + 1x512. Unless, you mean by consumer line. I don't think they'll need full AVX-512 for a while. Just supporting the ISA is fine with it having features to vectorize previously un-vectorizable code.

 

[mikk]

This needs a link. Kabylake-Y can't even take full advantage of the dual core. Unless otherwise shown, I'd assume a process shrink and a new architecture would allow a 4.5W 2 core chip, but with high clocks like Kabylake U.

You thought ICL-Y is a dualcore design? Looks like you are a little too pessimistic about ICL.


ICL-Y BGA1377 5.2W 4+2
https://www.ptt.cc/bbs/PC_Shopping/M.1503467362.A.A6A.html

 

[IntelUser2000]

ICL-Y BGA1377 5.2W 4+2
https://www.ptt.cc/bbs/PC_Shopping/M.1503467362.A.A6A.html

Hmm. Impressive. It suggests once they get 10nm+ out it will be really good.

Architecture-wise I'm not impressed with where Intel Core chips are going. Are we really going to end up having bunch of CPUs having the same perf/clock only differentiated by auxiliary features, TDP, and clocks?

 

[Glo.]

Hmm. Impressive. It suggests once they get 10nm+ out it will be really good.

Architecture-wise I'm not impressed with where Intel Core chips are going. Are we really going to end up having bunch of CPUs having the same perf/clock only differentiated by auxiliary features, TDP, and clocks?

For me, if the mainstream CPU TDP target shifts from 65W to 35-45W is best news, and biggest innovation in quite a lot of time in CPU department.

 

[jpiniero]

Architecture-wise I'm not impressed with where Intel Core chips are going. Are we really going to end up having bunch of CPUs having the same perf/clock only differentiated by auxiliary features, TDP, and clocks?

Isn't that how it is now?

 

[frozentundra123456]

Isn't that how it is now?

Yes, but is that a good thing? Skylake launched in August of 2015, so we have had 2.5 years with no ipc improvements. And in fact, it could be 6 months to a year or two *additional* to that before we see a new architecture with an ipc increase. Now granted, we have seen significant performance increases from process improvements and increased clockspeed, but we are sorely lacking architectural improvements.

 

[Eug]

You thought ICL-Y is a dualcore design? Looks like you are a little too pessimistic about ICL.


ICL-Y BGA1377 5.2W 4+2
https://www.ptt.cc/bbs/PC_Shopping/M.1503467362.A.A6A.html

It appears to get Thunderbolt too? Finally!

What is that site anyway?

 

[Dayman1225]

It appears to get Thunderbolt too? Finally!

What is that site anyway?

It's a Taiwan PTT Forum, I believe Chrisdar is one of the more reputable leakers on that site, iirc.

 

[Eug]

So, if the leaks are to be believed (and I know many people don't believe them), it's starting to look like:

2017: Kaby Lake Y 4.5 Watt dual-core (1.0 GHz base clock for 7Y30). No Thunderbolt.
2018: Cannon Lake Y 5.2 Watt dual-core (1.1 GHz base clock for entry level). No Thunderbolt.
2019: Ice Lake Y 5.2 Watt quad-core (no clockspeeds known yet). Includes Thunderbolt 3.

It seems the 2018 listing above may be the most controversial for some, even though the 2019 listing is just from a lone bulletin board message.

 

[LTC8K6]

So, if the leaks are to be believed (and I know many people don't believe them), it's starting to look like:

2017: Kaby Lake Y 4.5 Watt dual-core (1.0 GHz base clock for 7Y30). No Thunderbolt.
2018: Cannon Lake Y 5.2 Watt dual-core (1.1 GHz base clock for entry level). No Thunderbolt.
2019: Ice Lake Y 5.2 Watt quad-core (no clockspeeds known yet). Includes Thunderbolt 3.

It seems the 2018 listing above may be the most controversial for some, even though the 2019 listing is just from a lone bulletin board message.

Where does the chrisdar info say 2019?
It appears to have been posted 6 months ago, and nothing else since?

What am I missing?