Discussion Intel current and future Lakes & Rapids thread

Page 116 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

TheGiant

Senior member
Jun 12, 2017
748
353
106
I have altered the roadmap, pray I don't alter it any further.

Intel is ramping for 7nm aggressively, don't expect anything to be on solid schedule till 7nm.
ok we pray, we pray strong
I can even calculate my prayer faster with brand new 3900X

so what is the new roadmap? care to share?
 

birdie

Member
Jan 12, 2019
98
71
51
I have altered the roadmap, pray I don't alter it any further.

Intel is ramping for 7nm aggressively, don't expect anything to be on solid schedule till 7nm.

So, looks like 10nm is considered almost a complete failure and it won't be used for desktop/server parts. Damn. Over 5+++ years in development and such a finale.
 
Last edited:

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
So, looks like 10nm is considered almost a complete failure and it won't be used for desktop/server parts. Damn. Over 5+++ years in development and such a finale.
10nm is still happening its just depreciated for x86. Low-volume, prefer 14nm on x86, with primary focus going to 7nm.

Future was delayed to 7nm, and aimed to be secure and exploit-less(No meltdown/spectre/etc). => Harming 10nm in conjunction with HVM difficulties from 10nmv2(2016-2018).

Future;
ISAs, Cores, Component IP that are created via ML(machine learning) techniques. Supposedly, it is better than current gen in every way. Including, Sunnycove and Willowcove which are from older projects ported to 10nmv2 plus and plus plus.
 
Last edited:

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
Server Icelake is coming, there is no news that they canceled this product. For desktop they might really go directly from 14nm to 7nm. I wonder what process Intels discrete GPU next year is based on.
 

Dayman1225

Golden Member
Aug 14, 2017
1,152
973
146
Server Icelake is coming, there is no news that they canceled this product. For desktop they might really go directly from 14nm to 7nm. I wonder what process Intels discrete GPU next year is based on.
Well if Tigerlake is 10nm and that has an Xe graphics engine... one would imagine the scaled up IP would also be on 10nm. And yes, Icelake server is still going to be released. As for Desktop. I don't know if Alder Lake still exists on the roadmap, that was meant to be the last 10nm desktop generation, then came Meteor Lake on 7nm...
 

ksec

Senior member
Mar 5, 2010
420
117
116
Absolutely not. I just cannot wrap my head around the fact that no one among Intel management had the guts to stand up against the company being actively destroyed/mismanaged for so long. Intel perfectly knew about Zen/Ryzen. Intel knew about TSMC/Samsung 7nm nodes. They still kept dragging their 14nm node when it was obvious they didn't even had the capacity to mass-produce many-cores CPUs not to mention that 14nm doesn't quite work for such a scenario 'cause you end up with CPUs which have an insane TDP.

1. Most of the decent people in Intel left for other jobs, that is to Google, Tesla, Apple etc.

2. It wasn't destroyed or mismanaged, you are viewing the situation with the present times. At that time they had no competition. Sales were increasing, they were making YoY increase in Revenue and Profits, breaking records after records. Everything is Great!

3. Intel didn't knew about Zen or Ryzen. Or they did but they didn't care. Intel's bread and butter, that is the Laptop CPU and Datacenter business even as of today is still perfectly fine. ( For now )

4. The only way to tell a perfect lie is to actually lie to yourself and you somehow believe in the lie. Which is 10nm is going to come. And everything will be fine.

5. Who cares if 14nm is dragged along for another generation if you had no competition.

6. Intel's 10nm is roughly TSMC 7nm with slightly better density, but not as good TSMC 7nm+ ( EUV ). Which is to say even TSMC 7nm roadmap were right at the time Intel would still have the lead if they had ship it in 2018.

7. The Capacity problem is still the biggest mystery to me. I still don't believe in Ian's take about Hyper Scale Server Customers taking more orders.

Basically, Intel got to be where it is at today mostly because of Andy Grove. Only Paranoid Survive mentality. And you can tell from the post they aren't really paranoid at all. The only one who is paranoid was Andy Grove disciple Pat Gelsinger. And he was driven out of Intel by the current Chairman Andy Bryant. Andy is a finance guy, he only looks at sales data, profits and revenue.

It is actually very common for big companies in monopoly or their own bubble to have little to no productivity. And sit there doing nothing. So I don't see how this so hard to believe. If anyone who has been in working for over 10 years and worked in several Fortune 500 companies will realise you most company are very much like this. Lots of Discussions and little action.

And one reason why people go to Startup.
 
  • Like
Reactions: OTG

ondma

Platinum Member
Mar 18, 2018
2,718
1,278
136
1. Most of the decent people in Intel left for other jobs, that is to Google, Tesla, Apple etc.

2. It wasn't destroyed or mismanaged, you are viewing the situation with the present times. At that time they had no competition. Sales were increasing, they were making YoY increase in Revenue and Profits, breaking records after records. Everything is Great!

3. Intel didn't knew about Zen or Ryzen. Or they did but they didn't care. Intel's bread and butter, that is the Laptop CPU and Datacenter business even as of today is still perfectly fine. ( For now )

4. The only way to tell a perfect lie is to actually lie to yourself and you somehow believe in the lie. Which is 10nm is going to come. And everything will be fine.

5. Who cares if 14nm is dragged along for another generation if you had no competition.

6. Intel's 10nm is roughly TSMC 7nm with slightly better density, but not as good TSMC 7nm+ ( EUV ). Which is to say even TSMC 7nm roadmap were right at the time Intel would still have the lead if they had ship it in 2018.

7. The Capacity problem is still the biggest mystery to me. I still don't believe in Ian's take about Hyper Scale Server Customers taking more orders.

Basically, Intel got to be where it is at today mostly because of Andy Grove. Only Paranoid Survive mentality. And you can tell from the post they aren't really paranoid at all. The only one who is paranoid was Andy Grove disciple Pat Gelsinger. And he was driven out of Intel by the current Chairman Andy Bryant. Andy is a finance guy, he only looks at sales data, profits and revenue.

It is actually very common for big companies in monopoly or their own bubble to have little to no productivity. And sit there doing nothing. So I don't see how this so hard to believe. If anyone who has been in working for over 10 years and worked in several Fortune 500 companies will realise you most company are very much like this. Lots of Discussions and little action.

And one reason why people go to Startup.

I agree with you about the complacency. It is really "understandable", and happens to a lot of companies, but of course a dangerous attitude. What I dont understand is how they could so badly botch the 10nm design, and worse yet, never really be able to correct whatever is wrong with it.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
The Cascade Lake-AP line is bit of a disappointment now, but with HBM2 stacks enabling much increased memory bandwidth, and possibly using EMIB could allow it to carve a niche.

Or maybe allow the option of Optane DIMMs?

(8 x 24GB HBM stacks + up to 6TB of Optane DIMMs per socket vs. 8 x 24GB HBM stacks + DRAM DIMMs)
 
Last edited:

TheGiant

Senior member
Jun 12, 2017
748
353
106
1. Most of the decent people in Intel left for other jobs, that is to Google, Tesla, Apple etc.

2. It wasn't destroyed or mismanaged, you are viewing the situation with the present times. At that time they had no competition. Sales were increasing, they were making YoY increase in Revenue and Profits, breaking records after records. Everything is Great!

3. Intel didn't knew about Zen or Ryzen. Or they did but they didn't care. Intel's bread and butter, that is the Laptop CPU and Datacenter business even as of today is still perfectly fine. ( For now )

4. The only way to tell a perfect lie is to actually lie to yourself and you somehow believe in the lie. Which is 10nm is going to come. And everything will be fine.

5. Who cares if 14nm is dragged along for another generation if you had no competition.

6. Intel's 10nm is roughly TSMC 7nm with slightly better density, but not as good TSMC 7nm+ ( EUV ). Which is to say even TSMC 7nm roadmap were right at the time Intel would still have the lead if they had ship it in 2018.

7. The Capacity problem is still the biggest mystery to me. I still don't believe in Ian's take about Hyper Scale Server Customers taking more orders.

Basically, Intel got to be where it is at today mostly because of Andy Grove. Only Paranoid Survive mentality. And you can tell from the post they aren't really paranoid at all. The only one who is paranoid was Andy Grove disciple Pat Gelsinger. And he was driven out of Intel by the current Chairman Andy Bryant. Andy is a finance guy, he only looks at sales data, profits and revenue.

It is actually very common for big companies in monopoly or their own bubble to have little to no productivity. And sit there doing nothing. So I don't see how this so hard to believe. If anyone who has been in working for over 10 years and worked in several Fortune 500 companies will realise you most company are very much like this. Lots of Discussions and little action.

And one reason why people go to Startup.
one american professor made a research why so many companies go bad after becoming big
http://www.jaygalbraith.com/images/pdfs/StarModel.pdf
he says that company above 3K employees without significant effort aligning those 5 elements can make so many tasks inside that they dont need any customers
they just work for themselves
it is a lack of leadership
those finance masters are not leaders, they just harvest
what we see at AMD (Lisa Su) is a clear vision, she presents AMD and can talk about their products without problems
Just sa Steve Jobs did, now a little less Jim Cook, Bill Gates has it, as Satya Nadella too in MS
Intel lacks this- at the highest places they dont execute
from my external PointOfView Intel top managers are cannibalising the company
 
  • Like
Reactions: OTG and ksec

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Just sa Steve Jobs did, now a little less Jim Cook, Bill Gates has it, as Satya Nadella too in MS

Microsoft isn't special either if you look at the bipolar mess that's called Windows 10.

If anyone who has been in working for over 10 years and worked in several Fortune 500 companies will realise you most company are very much like this. Lots of Discussions and little action.

It's kind of natural if you think about it. It's almost laws of nature/physics. Smaller companies are faster because hierarchy isn't needed. You have 10 people, then you have 1 CEO and 9 employees. If you want something done, then it takes minutes to gather them in one room.

If you have 100, you start needing managers. If you have 50K, you need a lot of layers, adding latency when sending out instructions. Larger groups also start becoming sort of political and have internal conflicts and competitions.

That's even forgetting things like the tendency to promote those that provide the most benefit to the company(at least in the short term) like financial people. It makes sense if you think about it. A company is there to make money right?

A story I read about Will Wright and Maxis says they were for the most part quite strapped for resources. But his team made good games. When they started sacrificing that, their revenue increased a lot.

How will you convince people that running a project on extreme budget is better than being very profitable and be able to provide more benefits to employees and such?

(8 x 24GB HBM stacks + up to 6TB of Optane DIMMs per socket vs. 8 x 24GB HBM stacks + DRAM DIMMs)
Pretty sure memory capacity isn't a priority on the -AP parts. They go from 2 DIMM/channel support to 1 DIMM/channel support for the same reason. Greater memory capacity degrades signal integrity at high speeds so the maximum speed the modules can support goes down.
 
Last edited:
  • Like
Reactions: lightmanek

repoman27

Senior member
Dec 17, 2018
342
488
136
16 CPU + 30 chipset = 46 platform

If CPU had 30 PCIE lanes, I would understand the new socket, but bleh.
Mind you, that's 30 chipset HSIO lanes, up to 24 of which can be PCIe, so zero change from CNP / 300 series.
And there is essentially zero chance of CML using mesh instead of ring, because there will also be 2, 4, 6, and 8 core variants.
Comet Lake = Coffee Lake Refresh Refresh. You get 2 more cores, higher TDP, new socket, rebranded PCH.
 

ksec

Senior member
Mar 5, 2010
420
117
116
I agree with you about the complacency. It is really "understandable", and happens to a lot of companies, but of course a dangerous attitude. What I dont understand is how they could so badly botch the 10nm design, and worse yet, never really be able to correct whatever is wrong with it.

To be fair, it is kind of human nature. You are blindsided when you are very successful, like I said despite what happen in the past few years in terms of innovation and management, they were breaking records profits and revenue YoY. Whats the rush for 10nm if everything is working so well. ( On Paper )
 

repoman27

Senior member
Dec 17, 2018
342
488
136
While Intel has been more or less up front about Ice Lake being a limited volume product on the client side, I hoped that might mean something akin to Broadwell levels of bad (still a disaster, but not the unmitigated disaster that was Cannon Lake). But with Lenovo shipping the X1 Yoga G4 and Carbon G7 with Whiskey Lake, and now Apple using off-roadmap Coffee Lake-U 4+3e 15W in the 2019 13-inch MacBook Pro and the same garbage Amber Lake as last year in the MacBook Air... Either Dell got an exclusive on Ice Lake or we're just not going to see any flagship products with Intel 10nm CPUs in 2019. That's not good. The only way Intel can ease their 14nm capacity pressure is to deliver some measurable volume of 10nm parts that their customers can actually use.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
Microsoft isn't special either if you look at the bipolar mess that's called Windows 10.
well I dont know about you but win10 is by far the best windows I ever used
It is not super excellent for tablets (my surfacepro 4 as tablet sucks) but otherwise its excellent
MS has a vision and its improving
by this situation I wont be very surprised if in the next 2 years Intel splits to many companies
there must be fight of the competent and incompetent
usually (sad) the incompetent "win" because the competent leave, that is what IMO happened with engineers
 
  • Like
Reactions: Drazick

cbn

Lifer
Mar 27, 2009
12,968
221
106
Pretty sure memory capacity isn't a priority on the -AP parts. They go from 2 DIMM/channel support to 1 DIMM/channel support for the same reason.

Yes, It isn't.

Looking at the sleds below the main focus appears to be compute density:

Intel-Server-System-9200WK-Compute-Overview.jpg


P.S. Interestingly enough RDIMM price per GB doesn't increase until the 128GB capacity (128GB RDIMM is ~4x more expensive than a 64GB RDIMM). 24 x 128GB RDIMMs is about $14,500 more expensive than 24 x 128GB Optane DIMMs or 48 x 64GB RDIMMs. $14,500 extra for two processor's memory seems like a lot of money but perhaps in the grand scheme in things it isn't significant.
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
The only way Intel can ease their 14nm capacity pressure is to deliver some measurable volume of 10nm parts that their customers can actually use.

Intel is in the process of moving almost all of the capacity for 10 nm back to 14, if they haven't already completed that.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
$14,500 extra for two processor's memory seems like a lot of money but perhaps in the grand scheme in things it isn't significant.

It's not for the markets that require it.

It's important enough that they bothered making another memory type for it. FB-DIMMs were a result of that. FB-DIMMs increased the price of DIMMs in general so with Nehalem-EX they put the buffer on a separate chip instead. The buffer chips cost extra, adds latency, and uses more power, but still it was used. That's how much the particular market wanted it. IBM's Power chips also use buffers to extend memory capacity.

But with Lenovo shipping the X1 Yoga G4 and Carbon G7 with Whiskey Lake, and now Apple using off-roadmap Coffee Lake-U 4+3e 15W in the 2019 13-inch MacBook Pro and the same garbage Amber Lake as last year in the MacBook Air...

Apple is typically late with product introductions. If they wanted a January introduction, then Icelake would make sense.

Also, despite Ivy Bridge U chips being introduced June at Computex, I had to wait until November to get my XPS 12. After waiting a month for shipping, I had it on my hands in January.

well I dont know about you but win10 is by far the best windows I ever used
It is not super excellent for tablets (my surfacepro 4 as tablet sucks) but otherwise its excellent

The best was Windows 7. They should have used the basic UI and added features and call it Windows 10. Windows 8 took two step backwards, and Windows 10 went 1 step forwards, so we still end up with the worse one.
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
It's important enough that they bothered making another memory type for it. FB-DIMMs were a result of that. FB-DIMMs increased the price of DIMMs in general so with Nehalem-EX they put the buffer on a separate chip instead. The buffer chips cost extra, adds latency, and uses more power, but still it was used. That's how much the particular market wanted it. IBM's Power chips also use buffers to extend memory capacity.

So instead of four channels with eight DIMM slots (supporting FB-DIMMs) they went four channels with sixteen DIMM slots (not supporting FB-DIMM). Lowered price per GB at the cost of density.

With these Xeon-AP Servers I do wonder how much memory they have (on average)? I wouldn't be suprised if most of the high end ones use 128GB RDIMMs, but I wonder what is better a lower amount of HBM2 + 128GB RDIMMs or a higher amount of HBM2 + 128GB Optane DIMMs?
 
Last edited:

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
So instead of four channels with eight DIMM slots (supporting FB-DIMMs) they went four channels with sixteen DIMM slots (not supporting FB-DIMM). Lowered price per GB at the cost of density.

It's not about lower price per GB, but doubling capacities.

Considering what the -EX customers went through to get higher capacities, I think Optane PMMs will fit for a lot of them, in addition to workloads only PMMs can enable.

With these Xeon-AP Servers I do wonder how much memory they have (on average)? I wouldn't be suprised if most of the high end ones use 128GB RDIMMs, but I wonder what is better a lower amount of HBM2 + 128GB RDIMMs or a higher amount of HBM2 + 128GB Optane DIMMs?

I don't think many(if any) -AP customers will use Optane DIMMs even if it was possible. The bandwidth when the workload is greater than the size of HBM2 will plummet, so it'll result in a performance loss. The read bandwidth is about half, and write bandwidth is about 1/6th.

There's likely a technological limitations as well. Current Optane DIMMs require 1:1 mix of DIMMs to PMMs. Being able to use the on-package HBM2 as a cache will need significant changes.
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
Intel is in the process of moving almost all of the capacity for 10 nm back to 14, if they haven't already completed that.
its a good move atm
but..can someone explain how can you not fix a broken process for 3 years?- settings? tools? materials? staff?
 

cbn

Lifer
Mar 27, 2009
12,968
221
106
The bandwidth when the workload is greater than the size of HBM2 will plummet, so it'll result in a performance loss. The read bandwidth is about half, and write bandwidth is about 1/6th.

8 x 24GB HBM2 + 12 x 128GB Optane DIMMs for workloads that are more predictable vs. 12 x 128GB RDIMMs for workloads that are more random?

vs.

8 x 16GB (or 24GB) HBM2 + 12 x 128GB RDIMMs for workloads that vary from predictable to random?
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Intel mentions 1 part DRAM to eight parts Optane as the recommended ratio for Memory Mode:

No what I mean is you need Optane to be paired with at least equal number of DIMMs. So you can have 6 DIMMs and 3 DC PMMs but not 3 DIMMs and 6 DC PMMs.