Discussion Intel current and future Lakes & Rapids thread

[IntelUser2000]

Like I said there's probably some overlap with different segments. The actual details, well we will have to see.

Intel is incredibly slow. Even AMD has an on-die PCH. Now Intel's PCH has more functionality, but AMD's way seems to be right to me. In mobile you don't need all the features so the one in Renoir will suffice.

Now they want to separate the GPU and CPU too? For mobile devices? Why they need a complex packaging scheme to get it working is beyond me. I'm hoping Alderlake at least has some of the PCH functionality on an active interposer with two compute dies on top.

The root cause is still process, which seems like will continue until 7nm. So the question is, is 7nm bad as initial 14nm was in, or is it worse?

TSMC's way of going in smaller steps seems like a better approach if you just look at end results. 28-20-16-10-7-5-3 are not full node jumps in a historical sense. They are more than a half node, something like a 0.6 node jump.

Intel on the other hand are doing full jumps, with what is increasingly looking like a fix in-between rather than improvements as they call it.

 

[mikk]

The part about 7nm is very disappointing. And with Jim Keller pushing for TSMC. Some said with Icelake-SP we'll see lower-core count variants first with the 38 core count part coming later.

All of this comes from Charlie, so better take it with a grain of salt. He also said that the GPU in ADL will be on 14nm which is nonsense given that TGL-U Xe LP is 10nm based.

 

[IntelUser2000]

All of this comes from Charlie, so better take it with a grain of salt. He also said that the GPU in ADL will be on 14nm which is nonsense given that TGL-U Xe LP is 10nm based.

Yea, that's true. I thought Rocketlake was supposed to be the one with two dies.

No matter what people say, 10nm is still preferrable to 14nm, unless you are talking about PCH chips.

A 14nm GPU will result in significant reduction in performance or raise cost. It's not like it needs to run at 4GHz either.

 

[mikk]

It could be doable for a GT1 variant just like RKL-S and for sure using a 14nm based GT2 Xe LP in ADL-P would result in efficieny and performance decreases, this is not realistic. Not to mention that ADL should use a further improved 10nm version over Tigerlake, there is no reason for a process downgrade in ADL-P.

 

[jpiniero]

Not to mention that ADL should use a further improved 10nm version over Tigerlake, there is no reason for a process downgrade in ADL-P.

Charlie did say it was about margins. It's cheaper to give up on the IGP than to add capacity to a broken node. Going to need the 10 nm wafers for servers since 7 nm is slipping.

 

[mikk]

Charlie did say it was about margins. It's cheaper to give up on the IGP than to add capacity to a broken node. Going to need the 10 nm wafers for servers since 7 nm is slipping.

He may said this but this is pure nonsense and just proves he isn't reliable.

 

[coercitiv]

Can Alderlake 8+8 be competitive?

If you simply add the two, you get ~9400, which is on par with the Ryzen 9 3950X.

One could think of the hybrid state as "Hyperthreading 2.0". Where potential gains in perf/watt is there but with further complications in coding and support.

I'm quite optimistic about P variants, especially the 2+8 one, but the S variant is in trouble even as we speak. It will be more cost effective than a 12 + 0 while trading blows in performance depending on thread count and small cores fmax, but that's about it. Competition will likely match it with a 12 core, not a 16 core flagship.

The hybrid approach needs a higher ratio than 1:1 between big and small cores to become effective in desktops. Mobile 2+8 and the (surprisingly) missing 4+8 configs would benefit a lot from small cores in 10-25W configs, while 6+8 sounds quite solid for 35W+.

 

[liahos2]

Charlie did say it was about margins. It's cheaper to give up on the IGP than to add capacity to a broken node. Going to need the 10 nm wafers for servers since 7 nm is slipping.

Charlie also said there were MASSIVE layoffs in DCG in Jan Feb - which was completely bogus
Charlie is not an engineer. Charlie is not a financial analyst - has no idea what intel's margins are per sku.
Charlie said that intel would have 40-50kwpm capacity at 10nnm by year end. They exited Q2'20 with 50kwpm and are targeting 80-85kwpm by year end.

Charlie is a purveyor of intel downfall porn and should be viewed as such.

 

[jpiniero]

He may said this but this is pure nonsense and just proves he isn't reliable.

It makes sense if they are doing it to save money. It'll get to the point where it'd be impossible to sell 14 nm CPUs. The IGP is something they can get away with. If fabbing the IGP even partially on 14 nm means they don't have to buy more 10 nm equipment, that's a ton of money being saved. That's kind of what I was getting at.

Charlie also said there were MASSIVE layoffs in DCG in Jan Feb - which was completely bogus

Sometimes it's not about less workers, it's about cheaper.

 

[Dave2150]

It makes sense if they are doing it to save money. It'll get to the point where it'd be impossible to sell 14 nm CPUs. The IGP is something they can get away with. If fabbing the IGP even partially on 14 nm means they don't have to buy more 10 nm equipment, that's a ton of money being saved. That's kind of what I was getting at.



Sometimes it's not about less workers, it's about cheaper.

Intel's 14nm still produces the fastest gaming CPU, so it will continue to sell very well. Intel's 10nm and 7nm will be huge steps forward, when they eventually release. Exciting times ahead!

 

[IntelUser2000]

The hybrid approach needs a higher ratio than 1:1 between big and small cores to become effective in desktops. Mobile 2+8 and the (surprisingly) missing 4+8 configs would benefit a lot from small cores in 10-25W configs, while 6+8 sounds quite solid for 35W+.

I like the 1:4 ratio. So a 4+16 and 8+32 seems cool to me.

According to Golem's(german site) latest review,

Tremont ST/Sunny Cove ST/Multi-Threading

Cinebench R11.5-2.9	1,08 ptk	1,35 ptk (+25 %)	3,09 ptk (+286 % bzw +128 %)
Cinebench R15-0.3.8	68 ptk	107 ptk (+57 %)	251 ptk (+369 % bzw +135 %)
Cinebench R20-0.6.0	211 ptk	272 ptk (+29 %)	577 ptk (+273 % bzw +112 %)
Geekbench 5.1.1	529 ptk	884 ptk (+67 %)	1758 ptk (+332 % bzw +99 %)

Interesting in that Tremont does way better than expected. That's with SNC having a ~5% clock advantage! In R11.5 and R15, Tremont performs at Haswell+ levels.

They should use the hybrid concept to maximize ST performance. A 4+32 part that uses big cores that would be larger and power hungry than normal(Rather than Golden Cove in the Golden Cove timeframe, it'd have Ocean Cove instead).

As for Alderlake-S, if they position and price it right it can be attractive for majority of the desktop market.

Yes, it probably won't take the lead. I feel that's only possible sometime in the 7nm successor generation.

 

[Markfw]

Intel's 14nm still produces the fastest gaming CPU, so it will continue to sell very well. Intel's 10nm and 7nm will be huge steps forward, when they eventually release. Exciting times ahead!

Well, thats barely true today, and when Ryzen 3 comes out, most likely will not be true.

 

[DrMrLordX]

Exciting times ahead!

Not really. They're taking too long. The market is already starting to stagnate (see: AMD's XT-series release, Zen3 delays, Comet Lake-S). For an enthusiast, this isn't very exciting.

 

[liahos2]

Well, thats barely true today, and when Ryzen 3 comes out, most likely will not be true.

i thought you didnt like to speculate on the future?

 

[Markfw]

i thought you didnt like to speculate on the future?

At the moment, thats not much of a speculation, its almost a fact.

 

[lobz]

Charlie also said there were MASSIVE layoffs in DCG in Jan Feb - which was completely bogus
Charlie is not an engineer. Charlie is not a financial analyst - has no idea what intel's margins are per sku.
Charlie said that intel would have 40-50kwpm capacity at 10nnm by year end. They exited Q2'20 with 50kwpm and are targeting 80-85kwpm by year end.

Charlie is a purveyor of intel downfall porn and should be viewed as such.

Really? Where's ICL Xeon then? OnTrack™, right? So much that even NVIDIA had to use Epyc with their A100. And you know they would never do that if it wasn't absolutely necessary.

 

[Asterox]

Intel's 14nm still produces the fastest gaming CPU, so it will continue to sell very well. Intel's 10nm and 7nm will be huge steps forward, when they eventually release. Exciting times ahead!

Keep this in mind, Apple has much beeter data about Intel and his manufacturing future success.Apple wants only the best, but Intel and its nm trenches are far from the best.

Behind the doors, Apple tried to arrange an order from Intel for Apple future A SoC-s or CPU-s.Today we all now the situation, for Apple the Intel nm success or five year plan was not a option.Intel is to expensive+unstable, and Samsung also was not a good option.If TSMC manufacturing success was not that good or superior(now and in the future), Apple or AMD they would have big problems.

 

[ondma]

Keep this in mind, Apple has much beeter data about Intel and his manufacturing future success.Apple wants only the best, but Intel and its nm trenches are far from the best.

Behind the doors, Apple tried to arrange an order from Intel for Apple future A SoC-s or CPU-s.Today we all now the situation, for Apple the Intel nm success or five year plan was not a option.Intel is to expensive+unstable, and Samsung also was not a good option.If TSMC manufacturing success was not that good or superior(now and in the future), Apple or AMD they would have big problems.

Intel is obviously having big problems, but supposedly AMD can do no wrong, yet Apple didnt switch to AMD either. So the reason for the switch is obviously more than just being unsatisfied with Intel chips.

 

[ajc9988]

Not really. They're taking too long. The market is already starting to stagnate (see: AMD's XT-series release, Zen3 delays, Comet Lake-S). For an enthusiast, this isn't very exciting.

The market is not stagnating. Literally, AMD is on a 14-16 month cadence since Zen. Zen2 was July. So September to November is clearly possible. Allegedly they already have B0 stepping, meaning they are ready to ramp for the inventory buildup. And now, a couple months before release, is when that should happen.

So there is no delay. In fact, the digitimes story seems to have been fully false, coming out around the time Intel's road map of rocket not coming out this year appeared.

If comet is stagnation, need i remind you it is skylake part 5? Skylake, kaby, coffee, coffee refresh, comet. All skylake.

At least with zen 3, you go to 8 core ccx/ccd. Latency on L3 goes from 40ns to 47ns, but you no longer need the rounds trip to the I/O die to access the other 4 cores on the same die, thereby being a net reduction in latency. This is combined with each core having full access to the 32MB rather than just 16MB victim cache. Join that with IPC and other changes and you get a much faster speed up.

The IPC reports have been 15-20%. If you revisit GNs video on the 3100x and 3300x, you will see the effect on latency sensitive gaming for going to a single CCX of 8 cores. And you can pick cores further away from each other, thereby less heat concentration, which gives higher boost clocks for light workloads of like 2 cores.

But, yes, let's talk if stagnation.

Intel is obviously having big problems, but supposedly AMD can do no wrong, yet Apple didnt switch to AMD either. So the reason for the switch is obviously more than just being unsatisfied with Intel chips.

If you make your own chips in house, you get more profit. Not hard to figure out.

Edited to remove my phone and:

To be clear, I believe the IPC reports are from server parts and are in non- latency sensitive tasks. That means the IPC will be multiplied by the percentage benefit found in latency sensitive tasks found in the GN video. If I am correct, that means there will be significant game uplift. So the AMD 20% versus the rocket lake 20% will not keep things where they currently are.

Also, on zen 3 rumors, I think the 15% on 64 cores versus the 20% on 32 cores either reflects lack of memory bandwidth or multiplier, as the rumor said performance and not IPC specifically on that part. Time will tell on that. But wanted to clarify.

 

[DrMrLordX]

The market is not stagnating.

Sure it is. We may get two (count em: two!) Vermeer chips in October; meanwhile, AMD rehashed Matisse in what was esentially a dud launch. Which they got away with because Comet Lake-S was also a dud launch.

Alder Lake-S may not show up until Q4 2021. AMD knows it, too.

 

[jpiniero]

Alder Lake-S may not show up until Q4 2021. AMD knows it, too.

Most likely it'll be a year from Rocket Lake-S's launch, whenever that ends up being.

 

[coercitiv]

Most likely it'll be a year from Rocket Lake-S's launch, whenever that ends up being.

Did you just say Q4 2020?!

 

[DrMrLordX]

Most likely it'll be a year from Rocket Lake-S's launch, whenever that ends up being.

Let's hope not. Rocket Lake-S is Q1 2021 if I recall correctly.

 

[jpiniero]

Let's hope not. Rocket Lake-S is Q1 2021 if I recall correctly.

My feeling is that will launch at CES.

 

[ajc9988]

Sure it is. We may get two (count em: two!) Vermeer chips in October; meanwhile, AMD rehashed Matisse in what was esentially a dud launch. Which they got away with because Comet Lake-S was also a dud launch.

Alder Lake-S may not show up until Q4 2021. AMD knows it, too.

You misunderstand the purpose of the XT lineup, which more likely was to iron out process used for the upcoming Zen 3 chips while technically having something competing against comet at each price point. You ignore Intel's done that multiple times before. Look how people reacted to R0 stepping. They lost their minds when it did what exactly?

My point is clear, you are mis- attributing purpose to that launch to contribute to your view.

Further, alder lake looks like a streaming pile on a 1700 pin socket, which will have to go against a 5nm Zen4 chip. We still will have to see how it turns out, but the limitation of the small cores and evidence of Microsoft not being able to implement advanced scheduler programming well, I just see very little on that side.

But, believe what you want. That is the beauty of life, right?