Discussion [SA] Intel should not launch Ice Lake-SP

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
OK lets get this out in the open right away, Intel’s Ice Lake-SP CPU is a dog, it loses at just about everything but a few benchmarks that heavily use it’s already deprecated AI instructions. For anything real world it is badly outclassed by both AMD’s and Intel’s own offerings, at times by multiples not percentages. Intel should simply not launch it, the damage it will do to their reputation, customer relations, and investor confidence far outweigh the meager marketing halo they will attempt to grab.


Not a subscriber, so I can't see the numbers... But that's a pretty bold headline. I dread to think how much Ice Lake has been neutered to get it out the door.
 

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
Well, as they plan to launch this Q4 this year, we are bound to find out sooner rather than later.

Charlie is known for some hyperbole regarding Intel, but if he is so vocal about it, things really must be quite bad.

If this is true, I'm beginning to believe that the rumored Intel's TSMC 6nm product is actually Sapphire Rapid. The've broken promises to OEMs so many times already since Skylake-X, that nobody would take 10nm Sapphire Rapid seriously with this track record (vs what was promised)

BTW, I already linked to this in the Ice-Lake thread.
 
  • Like
Reactions: lightmanek

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
If this is true, I'm beginning to believe that the rumored Intel's TSMC 6nm product is actually Sapphire Rapid.

It's possible, and that's one product that has to deliver for Intel. They've order a lot of wafers of 6nm too. Another possibility is Alder Lake. DG2 seems highly likely.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
It's possible, and that's one product that has to deliver for Intel. They've order a lot of wafers of 6nm too. Another possibility is Alder Lake. DG2 seems highly likely.

Intel confirmed for both Sapphire Rapids and Alder Lake being based on 10nm. Ponte Vecchio is likely based on a non Intel process.
 
  • Like
Reactions: lightmanek

rbk123

Senior member
Aug 22, 2006
743
345
136
Charlie has great sources, but its hard to know how bad it actually is. Its not as if this is the first time he's predicted Armageddon for Intel (or Nvidia).
Right. Intel isn't as dumb as they're made out here, and they saw what happened with Bulldozer. It most likely isn't as bad, or it'll be a targeted release of some kind.

On the other hand, Bob did say we need to stop using benchmarks..... (I kid!)
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
10 nm yields are still so bad, it's probably more profitable to push Tiger Lake. The client margins will take a deep hit but it is still profitable.

Especially when you factor in product viability; a 2 core 32 EU Tiger Lake is something you can sell... an 8, 10 core Icelake-SP is something that would be tough.
 

mikk

Diamond Member
May 15, 2012
4,112
2,108
136
How recently? Everything's up in the air now.

Two weeks ago they did confirm this. It must be concrete because both are scheduled for 2021 and we have seen the first ADL leaks in the wild. Ponte Vecchio is a safe bet for TSMC.

I wonder if ICL-SP is using the first 10nm version like ICL-U. Because it's much worse than 10+ apparently.
 

Kuiva maa

Member
May 1, 2014
181
232
116
It's possible, and that's one product that has to deliver for Intel. They've order a lot of wafers of 6nm too. Another possibility is Alder Lake. DG2 seems highly likely.

Do we know any rumors of the size of wafer order they placed on TSMC? Is it enough for just Aurora or will they try to actually bring it to the server market?
 

DrMrLordX

Lifer
Apr 27, 2000
21,582
10,785
136
Two weeks ago they did confirm this.

I'm still skeptical. We haven't seen TigerLake yet so we don't have anything to go on 10nm-wise except last year's IceLake-U, which in-and-of itself made their yields last year look kinda bad.

Intel has signaled that the formerly-7nm chiplets for Ponte Vecchio are now TSMC 5nm and Intel 7nm (mixed) so they may do the same with Sapphire Rapids (some 10nm, some 6nm). Alder Lake I could see them trying to push on their own 10nm exclusively, especially if TigerLake works out better than IceLake-U.

Rumor was 180,000 wafers I believe. Way more than they'd need for just Aurora.

That's the number I heard, and only part of Aurora could be covered by that wafer allocation anyway (the 10nm part of Ponte Vecchio could be moved to TSMC 6nm, but the 7nm part would have to be moved to TSMC 5nm).
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
Right. Intel isn't as dumb as they're made out here, and they saw what happened with Bulldozer. It most likely isn't as bad, or it'll be a targeted release of some kind.

On the other hand, Bob did say we need to stop using benchmarks..... (I kid!)

I wonder if it will be a "targeted release" so that they don't get in trouble for misleading shareholders about the state of 10nm...
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
Charlie is known for some hyperbole regarding Intel
Aside from selling his premium access (after all nothing sells like clickbait excerpts that make you want to read the rest...) I do feel his constant doomsaying was to counterbalance the unwarranted high rating Intel got in the market for way too long. It's telling that Intel's stock only significantly dropped when Swan admitted the issues with 7nm may warrant a change of course. I mean duh, everybody looking at the technical developments since 2015 should see that there are real serious issues, but apparently Intel's constant sugarcoating (and nice financial stats) was always sufficient for everybody to look the other way. Only Charlie repeatedly put it in (exaggerated) words.
 

ScopedAndDropped

Junior Member
Feb 15, 2020
7
3
41
Would SPR be the first TSMC product then? If ICX doesn't work I can't imagine a bigger SPR die would work on 10nm. Semiaccurate do say that they are making a "core" but which one could it be? With Aurora depending on SPR, would they let a dog of a node bring it down?

If they are buying 6nm wafers would it be something that was designed on 10nm as they are the closest dimensionally?
 
Last edited:

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
Two weeks ago they did confirm this. It must be concrete because both are scheduled for 2021 and we have seen the first ADL leaks in the wild. Ponte Vecchio is a safe bet for TSMC.

IMO dual sourcing Sapphire Rapids makes a lot of sense. The most from the entire Intel lineup of 2021.

It's not even so much whether they can deliver a good product on 10nm (Based on Tiger-Lake they probably can) but whether they can produce enough without sacrificing any other 10nm products. Intel has undelivered to server OEMs for multiple years now and trust is beginning to erode. If Ice-Lake SP is half as bad as Charlie claims (remains to be seen) then delivering Sapphire Rapid on time and in-mass becomes beyond critical.

Sapphire Rapid should have competitive core count to Rome/Milan (according to Charlie yet again), it also has DDR5, PCIe 5, newer Matrix/AI instructions and more. If Intel could deliver it in quantity in 2H2021 then they would finally have a product that has a clear edge on the competition at that point in time

Things wouldn't be nearly as rosy if:
a) It's postponed 6 months (as is almost a tradition @ Intel now) putting it against Genoa in 2022
b) They can only produce limited amounts of it (say 20% of the entire server demand). Be it due to yields, the need to produce all other kinds of (client) 10nm products or even unexpected demand.

Dual-sourcing would allow them to ramp up much quicker (and in larger quantities) and sell more Sapphire Rapids vs Ice Lake SP or god-forbid 14nm stuff and give them a safety net just in case 10nm "strikes again".

If I were In charge of delivering Sapphire Rapid on time and in mass, I'd surely have dual sourced it from both TSMC 6nm and Intel 10nm even if everybody underneath me in the organization would swear to god that: "In 2021 10nm will ramp up perfectly with godly parametric yields". Just on precedent of the previous years alone.

If the miracle indeed happens and 10nm really turns out so good that Intel can satisfy the demand of the entire server market with only 10nm Sapphire Rapids (without losing out on any desktop/laptop sales) they could just keep the amount of wafers they order from TSMC to the minimum (rumored TSMC lead time of 6 months is plenty to call it). It would just be a good "proof of concept" for engineering teams. And sure enough engineers would learn a ton about making dual-sourceable designs, should the need resurface again a later date, say for 7nm CPUs.

Because of the forementioned stuff dual sourcing Sapphire Rapid looks like a total no-brainer. Particularity if this is now communicated on the top level ("we'll outsource where/when it makes sense").

Then again I'm a software-engineer not an MBA :p
 
  • Like
Reactions: coercitiv

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
Because of the forementioned stuff dual sourcing Sapphire Rapid looks like a total no-brainer. Particularity if this is now communicated on the top level ("we'll outsource where/when it makes sense").
TSMC is not an attorney one keeps on retainer, having them as backup only makes financial sense with downsized internal production plans. There's no way to spin this internally, they will spot budget cuts and job losses from miles away.

Having said that, it's still the better choice. Backporting only works if you screw up one node every now and then, not two consecutive ones.
 

jur

Junior Member
Nov 23, 2016
15
1
81
There is Granite Rapids tag at the bottom of the article, so that may be the product Intel is out-sourcing. There are also GAA, 3nm and 5nm tags. So the article is probably also talking about Intel going gaa at 5nm since, if I remember correctly, TSMC announced that they are staying with finfet at 3nm.
 

ScopedAndDropped

Junior Member
Feb 15, 2020
7
3
41
There is Granite Rapids tag at the bottom of the article, so that may be the product Intel is out-sourcing. There are also GAA, 3nm and 5nm tags. So the article is probably also talking about Intel going gaa at 5nm since, if I remember correctly, TSMC announced that they are staying with finfet at 3nm.

He puts tags that may be completely unrelated in order to mess with people who don't have subsriptions but want info. I wouldn't take them seriously.
 
  • Like
Reactions: mikk

eek2121

Platinum Member
Aug 2, 2005
2,904
3,906
136
He puts tags that may be completely unrelated in order to mess with people who don't have subsriptions but want info. I wouldn't take them seriously.

I personally avoid his stuff. It is far too clickbait-y and he charges far too much for me to take him seriously. I have no problems plunking down cash for a subscription to a website if it has great content, but not that much cash. I would rather buy top end hardware every year.
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
There is Granite Rapids tag at the bottom of the article, so that may be the product Intel is out-sourcing. There are also GAA, 3nm and 5nm tags. So the article is probably also talking about Intel going gaa at 5nm since, if I remember correctly, TSMC announced that they are staying with finfet at 3nm.

GAA is interesting. There was the recent blog post from an Intel engineer which claimed that Intel tried to get GAAFETs working for 7nm, but had to abandon it and be less ambitious with 7nm.