[Anandtech] Intel's Architecture Day 2018

PeterScott

Platinum Member
Jul 7, 2017
2,605
227
96
#1
Intel's Architecture Day 2018: The Future of Core, Intel GPUs, 10nm, and Hybrid x86

I know some bits of this were getting some mention in other threads, but it seems like it was big enough to warrant it's own discussion.

Highlights for me were a new architecture (finally) Sunny cove, that looks to have potential (4-5 wide allocation) to deliver Intels first real IPC improvement in MANY years, and also what should be a nice improvement in the the IGP, and a look at Intels multichip solutions(Foveros) and Intel Big-little.

The Q&A was also somewhat interesting, like where they admit being node-locked on their designs really messed them up, and they will be more node agnostic in the future.

Overall, more actual news than we have seen from Intel in some time.
 

epsilon84

Senior member
Aug 29, 2010
943
121
136
#2
This is the most exciting news we've had from Intel for a long time and it's quite refreshing to see them being more open and also admitting to design mistakes, like node locked features which meant a lot of 10nm improvements couldn't be backported to 14nm.

I do hope that they will deliver on their 2H 2019 timeframe for Sunny Cove, but as I'm sure you will see in the posts following mine, we have a sizeable amount of skeptics here who believe Intel will be milking 14nm well into 2020...
 

Yotsugi

Senior member
Oct 16, 2017
794
196
96
#3
Nothing interesting besides that gimmicky Foveros SoC.
At least they admitted their mistakes, which is remarkable given that's Intel we're talking about, but still.

Oh and they're still dead-dead (even deader than dead) in servers.
 
Apr 27, 2000
11,827
1,042
126
#4
The only real question about Sunny Cove is, where will they launch it in 2019? Maybe if they hadn't announced Comet Lake 10c for desktop so recently, I'd have some hope of IceLake on the desktop after all.

To me, for the desktop/client group, it looks like they just slapped some new names on products that have been in the works for some time and called it a day. Seems like all their most-interesting products are showing up outside of desktop and server.

Oh and they're still dead-dead (even deader than dead) in servers.
Cooper Lake doesn't inspire much confidence. What happened to Sapphire Rapids? Is that going to be a 2021/2022 chip now?
 
Mar 13, 2006
10,126
121
126
#5
Oh and they're still dead-dead (even deader than dead) in servers.
Intel is doomed. Nevermind that Intel's Data Center Group has more revenue and profit than the entirety of AMD. Or that Intel has has raised its EOY projections by an amount greater than AMD's total revenue.

Yep, Intel is doomed.
 

maddie

Platinum Member
Jul 18, 2010
2,586
489
136
#6
Intel is doomed. Nevermind that Intel's Data Center Group has more revenue and profit than the entirety of AMD. Or that Intel has has raised its EOY projections by an amount greater than AMD's total revenue.

Yep, Intel is doomed.
Finally you get it. I'd really lost all hope. The meteorite has just struck and the kings are still roaring. :p
 

Mopetar

Diamond Member
Jan 31, 2011
4,480
418
126
#7
I do hope that they will deliver on their 2H 2019 timeframe for Sunny Cove, but as I'm sure you will see in the posts following mine, we have a sizeable amount of skeptics here who believe Intel will be milking 14nm well into 2020...
It probably depends on how 10nm performs and what their capacity and yields are like. Previously they indicated that until they got to 10nm+ they actually expected the mature 14nm process to offer higher total performance. If 10nm doesn't have all of the issues worked out, it may be necessary for Intel to keep using 14nm for certain market segments.

Hopefully they don't run into any more delays, because the architectural changes look nice and I'd like to see a return to substantial double digit performance gains like we see in the early days with SB.
 

Carfax83

Diamond Member
Nov 1, 2010
5,707
22
126
#8
So for those in the know, my question is, if Intel is bringing AVX 512 to the mainstream parts, how does it intend to increase the available bandwidth to make it more useful and powerful? From what I understand, with the current HEDT parts, even quad channel memory isn't enough to really take full advantage of AVX 512 let alone the dual channel mainstream CPUs. I know they are increasing the amount of cache and probably , but that can only get you so much right?
 

jpiniero

Diamond Member
Oct 1, 2010
6,407
260
126
#9
So for those in the know, my question is, if Intel is bringing AVX 512 to the mainstream parts, how does it intend to increase the available bandwidth to make it more useful and powerful? From what I understand, with the current HEDT parts, even quad channel memory isn't enough to really take full advantage of AVX 512 let alone the dual channel mainstream CPUs. I know they are increasing the amount of cache and probably , but that can only get you so much right?
Sunny Cove Client is probably still 2x256 units, and not 2x512.
 

NostaSeronx

Platinum Member
Sep 18, 2011
2,367
156
126
#10
The only real question about Sunny Cove is, where will they launch it in 2019? What happened to Sapphire Rapids? Is that going to be a 2021/2022 chip now?
Just guesses:
Icelake-Y => Q1 2019 paper launch
Icelake-U (& S) => Q2 2019 paper launch
Icelake-X (AP & SP) => H1 2019 paper launch, with actual products H2 2019.

Sapphire Rapids/Granite Rapids should be 2020(H2)/2021(H2).

Knight's Cove/ICLX(Sunnycove) -> Sapphire Rapids(Willowcove) -> Granite Rapids(Goldencove)
Sunny Cove Client is probably still 2x256 units, and not 2x512.
2x256 + 1x512 is probably going to be the standard. Bunch of legacy AVX/AVX2 256-bit optimziation + room for 2x512 when needed.
 

PeterScott

Platinum Member
Jul 7, 2017
2,605
227
96
#11
It probably depends on how 10nm performs and what their capacity and yields are like. Previously they indicated that until they got to 10nm+ they actually expected the mature 14nm process to offer higher total performance. If 10nm doesn't have all of the issues worked out, it may be necessary for Intel to keep using 14nm for certain market segments.

Hopefully they don't run into any more delays, because the architectural changes look nice and I'd like to see a return to substantial double digit performance gains like we see in the early days with SB.
Yeah, I suspect we will get Laptop/low power desktop parts first in 10nm. 14nm might have to carry the High Performance desktop load for a bit longer.

2H19 should be quite interesting with Sandy Cove on 10nm and Zen 2 on 7nm, though Zen 2 will almost certainly arrive sooner, and will definitely be on desktop.
 
Nov 6, 2018
94
99
66
#12
There's a really interesting interview with David Kanter (from realworldtech.com) on Gamer's Nexus Youtube channel related to this thread:

I'm adding it here since someone in here might not have yet seen it. I'm currently watching it myself.

Addition: David Kanter talks in the interview mostly about advanced packaging technologies including EMIB, silicon interposers and true 3D-packing and what are the problems with heat and power transform through stacked dies while on the other hand, you can have huge space savings and mix different kind of dies. It's not just about what Intel is doing but about where the whole semiconductor industry is heading. Advanced packaging technologies (like used in HBM memory stacks) are still expensive but there's a strong need to bring those to lower price points and to (mainstream) consumer market. There are other older interesting videos on Youtube if you just search for "David Kanter". It's always nice to get opinions from someone who knows the semiconductor industry so well.
 
Last edited:

jpiniero

Diamond Member
Oct 1, 2010
6,407
260
126
#13
Yeah, I suspect we will get Laptop/low power desktop parts first in 10nm. 14nm might have to carry the High Performance desktop load for a bit longer.
I am sort of wondering if Intel would port Willow Cove to 14nm; and that would be the CPU core used in the 2020 desktop and high power laptop parts. Use EMIB or Forvous to connect to the optional GPU.
 

epsilon84

Senior member
Aug 29, 2010
943
121
136
#15
I am sort of wondering if Intel would port Willow Cove to 14nm; and that would be the CPU core used in the 2020 desktop and high power laptop parts. Use EMIB or Forvous to connect to the optional GPU.
If Sunny Cove is 10nm what are the chances of the *successor* Willow Cove being backported to 14nm? I'll say zero to none. Think for a second how ridiculous that sounds.
 

jpiniero

Diamond Member
Oct 1, 2010
6,407
260
126
#16
If Sunny Cove is 10nm what are the chances of the *successor* Willow Cove being backported to 14nm? I'll say zero to none. Think for a second how ridiculous that sounds.
Very rediculous, but the alternative might be Skylake yet again.
 

tamz_msc

Platinum Member
Jan 5, 2017
2,230
219
106
#18
They say that allocation-width has increased from 4 to 5 but the block diagram shows the main execution units(excl. AGUs) still attached to four ports.
 

csbin

Senior member
Feb 4, 2013
816
8
136
#19
Just guesses:
Icelake-Y => Q1 2019 paper launch
Icelake-U (& S) => Q2 2019 paper launch
Icelake-X (AP & SP) => H1 2019 paper launch, with actual products H2 2019.

Sapphire Rapids/Granite Rapids should be 2020(H2)/2021(H2).

Knight's Cove/ICLX(Sunnycove) -> Sapphire Rapids(Willowcove) -> Granite Rapids(Goldencove)
2x256 + 1x512 is probably going to be the standard. Bunch of legacy AVX/AVX2 256-bit optimziation + room for 2x512 when needed.

:p

Cascade Lake-SP => Q1 2019

Cooper Lake-SP=> Q1 2020?

Ice Lake-SP=> Q1 2021?


 

Nothingness

Golden Member
Jul 3, 2013
1,895
32
106
#20
I would not call Sunny Cove a new micro-architecture. What was presented just looks like an evolution of the existing micro-arch. Increasing retirement width, cache size and uop cache, and adding a store unit does not make it a new micro-arch. No radical change.

But that's OK, because the current micro-arch already was very good.
 

Yotsugi

Senior member
Oct 16, 2017
794
196
96
#21
Cooper Lake doesn't inspire much confidence. What happened to Sapphire Rapids? Is that going to be a 2021/2022 chip now?
SPR is mid-late 2021, I guess (that's assuming ICL-SP is acktually Q3 2020).
Intel is doomed. Nevermind that Intel's Data Center Group has more revenue and profit than the entirety of AMD. Or that Intel has has raised its EOY projections by an amount greater than AMD's total revenue.

Yep, Intel is doomed.
Yes, they're pretty much doomed in servers and their only real chance for comeback is 7nm 2022-2023 products.
It's that simple.
 

NostaSeronx

Platinum Member
Sep 18, 2011
2,367
156
126
#22
I would not call Sunny Cove a new micro-architecture. What was presented just looks like an evolution of the existing micro-arch. Increasing retirement width, cache size and uop cache, and adding a store unit does not make it a new micro-arch. No radical change.

But that's OK, because the current micro-arch already was very good.
I think they are missing the nitty gritty details in the showing.
-> FO4 delay
-> Pipeline stage depth or samething Pipeline length
-> OoO Front-end/Retirement or InO Front-end/Retirement.

Skylake-SP: 2 LD/STA, 1 STA, 1 STD; Effective 2 64B Load/1 64B Store
Sunnycove: 2 LD, 2 STA, 2 STD; Effective 2 64B Load/2 64B Store
Skylake-SP: 1 execution scheduler and 1 memory scheduler
Sunnycove: 1 execution scheduler and 3 memory schedulers(2 for LD&STA units and 1 for STD units)

imho, based on the profile which I got Sunnycove/Willowcove(WLcove). It implied that Willowcove will be going for two execution schedulers.

Sunnycove's EX side is similar for speed of implementation, while its MEM side is completely different. With Willowcove bringing in the new execution design which will be ultra-wide. Compared to Sunnycove's wider EX side and Skylake's (& Cannonlake's) wide EX side.

Do to the abstraction of cores being built for set nodes, all the cove cores can technically be tocks/new architectures.
 
Last edited:
Feb 2, 2009
12,961
228
126
#23
Yes, they're pretty much doomed in servers and their only real chance for comeback is 7nm 2022-2023 products.
It's that simple.
They will sell way more 14nm server products in 2019 than AMD 14nm+7nm , hell im betting they will sell more 14nm server products in 2020 than AMD 7nm EPYC 2.

This Intel is dead thing is ridiculous, Intel can stay another 2-3 more year at 14nm and still outsell AMD easily both in Servers and Laptops.

On topic, why i remember that Ice Lake was end of 2019 with server products early 2020 ?? They changed the roadmap ??
 

TheELF

Platinum Member
Dec 22, 2012
2,848
122
126
#24
So for those in the know, my question is, if Intel is bringing AVX 512 to the mainstream parts, how does it intend to increase the available bandwidth to make it more useful and powerful? From what I understand, with the current HEDT parts, even quad channel memory isn't enough to really take full advantage of AVX 512 let alone the dual channel mainstream CPUs. I know they are increasing the amount of cache and probably , but that can only get you so much right?
Kapow
https://www.tomshardware.com/news/intel-optane-persistent-memory-dimms,37150.html
Also they increased store/load capabilities within the chip
 


ASK THE COMMUNITY

TRENDING THREADS